- Email: [email protected]
Fails-to-deliver, short selling, and market quality
Journal of Financial Economics 114 (2014) 493–516
Contents lists available at ScienceDirect
Journal of Financial Economics journal homepage: www.elsevier.com/locate/jfec
Fails-to-deliver, short selling, and market quality$ Veljko Fotak a, Vikas Raman b, Pradeep K. Yadav c,n a
School of Management, University at Buffalo, The State University of New York, USA Warwick Business School, UK c Price College of Business, Division of Finance, University of Oklahoma, 307W. Brooks, Room 205A, Norman, Oklahoma 73019, USA b
a r t i c l e i n f o
abstract
Article history: Received 5 September 2011 Received in revised form 16 December 2013 Accepted 14 January 2014 Available online 17 August 2014
We investigate the aggregate market quality impact of equity shares that fail to deliver (hereafter “FTDs”). For a sample of 1,492 NYSE stocks over a 42-month period from 2005 to 2008, greater FTDs lead to higher liquidity and pricing efficiency, and their impact is similar to our estimate of delivered short sales. Furthermore, during the operative period of a Security and Exchange Commission (SEC) order mandating stock borrowing prior to short sales, the securities affected display relatively lower liquidity and higher pricing errors. Finally, we do not find any evidence that FTDs caused price distortions or the failure of financial firms during the 2008 financial crisis. & 2014 Elsevier B.V. All rights reserved.
JEL classification: G14 G18 Keywords: Naked short selling Short selling Failure to deliver
1. Introduction Trades in US stock markets are settled on a three-day cycle: for trades on day t, if the net delivery obligations of a clearing member are not fulfilled on day tþ3, any undelivered position becomes a “failure-to-deliver” (or FTD).
☆ The authors thank Leslie Boni, Tarun Chordia, John Griffin, Stewart Mayhew, Bill Megginson, David Musto, Narayan Naik, Adam Reed, Paul Schultz, Chester Spatt, “Vish” Viswanathan, Andriy Shkilko, and participants at the Western Finance Association Meetings, INQUIRE UK conference, Notre Dame Conference on Market Regulation, American Finance Association Meetings and seminars at Case-Western, Indian School of Business, the Fondazione Eni Enrico Mattei and the University of Oklahoma, for helpful comments and discussions. The authors gratefully acknowledge the support of the Center for Financial Research at the University of Koln, Germany, the Fondazione Eni Enrico Mattei, and the Institute for Quantitative Investment Research (INQUIRE), UK. The authors remain responsible for all errors. n Corresponding author at: Price College of Business, Division of Finance, University of Oklahoma, 307 W. Brooks, Room 205A, Norman, Oklahoma 73019, USA. Tel.: þ 1 405 325 5591; fax: þ1 405 325 5491. E-mail address: [email protected] (P.K. Yadav).
http://dx.doi.org/10.1016/j.jfineco.2014.07.012 0304-405X/& 2014 Elsevier B.V. All rights reserved.
Regulators and financial journalists have widely perceived FTDs, and specifically “naked” short sales, as having a very negative impact on markets.1 For example, a report in Time Magazine (“Watch out, they bite,” November 9, 2005), quoting estimates by former Under Secretary of Commerce Robert J. Shapiro, alleges that naked short selling has “cost investors $100 billion and driven 1,000 companies into the ground”; and an article in Euromoney (“Short selling: the naked truth,” December 2008) claims that, “Fails to deliver in the US equity market have exacerbated the sharp declines in share prices of financials. Although the SEC is clearing up the mess caused by naked short selling, more drastic measures might be needed to restore confidence.” In this context, this paper investigates the overall net impact (on day tþ1) of the change in the open interest of undelivered positions (i.e., the change in FTDs) generated
1 The term “naked short sale” is used (in this paper) to describe a short sale that, irrespective of any intent considerations, fails to deliver because of the short seller not making timely stock borrowing arrangements.
494
V. Fotak et al. / Journal of Financial Economics 114 (2014) 493–516
collectively by trades on day t (and observed on day t þ3). First, we analyze the effect of FTD changes on pricing efficiency and liquidity. Second, we examine whether FTD changes played a causal role in the major price declines or the demise of financial institutions during the 2008 financial crisis. Settlement on US stock exchanges is done electronically through the Depositary Trust and Clearing Corporation, its subsidiaries, and associated agencies (hereafter collectively referred to as the DTCC). Ownership records are largely held, tracked, and transferred electronically through the DTCC, with more than 99.9% of all trades involving only electronically held securities (Morris and Goldstein, 2009). The DTCC becomes the central counterparty of all dulymatched error-reconciled trades, and electronically checks and updates relevant stock ownership accounts. If the stock ownership accounts associated with a clearing member with a net delivery obligation do not actually include the stock needed for delivery on settlement day, the undelivered position becomes a FTD.2 FTDs can potentially arise from both short and “long” sales (i.e., sales of duly-owned stock). In a short sale, such a delivery shortfall arises when the stock is not borrowed in time to credit the stock ownership account by settlement day.3 In a long sale, a delivery shortfall arises, for example, when (duly-owned) sold stock has been lent out and not returned to the stock ownership account in time for settlement; or because the broker has not ensured electronic recording of a paper certificate prior to the trade; or in conjunction with oversubscribed securities at security issuance. There has been a strong regulatory focus on reducing FTDs. In January 2005, Regulation SHO introduced requirements to “locate” stock prior to every short sale to reduce FTDs due to stock-borrowing delays; it also forced firm stock-borrowing arrangements prior to short sales in “threshold-list” stocks (that is, securities with high levels of FTDs) unless existing FTDs were “closed-out.” In July/ August 2008, a temporary SEC order mandated stock borrowing arrangements prior to short selling in select financial stocks to “eliminate any possibility that naked short selling may contribute to the disruption of markets” (SEC Release 58166, 2008). After Lehman's collapse in September 2008, through Rule 204T (later made permanent), FTDs arising from short sales were virtually eliminated by requiring borrowing or purchasing by the broker by the morning after the failure day; FTDs arising from long 2 A delivery failure by an individual client account will not become a FTD at DTCC level if another client account held with the same clearing member lends the stock explicitly, or implicitly because of being due to receive delivery. Due to such “netting,” a FTD at DTCC level cannot generally be uniquely attributed to a specific trade or trader; nor is it possible to identify whether it originated from a short or long sale. 3 The Regulation SHO “locate” requirement introduced in January 2005 requires broker-dealers to ensure (prior to a short sale) that there are “reasonable grounds” to believe that the security could be borrowed for timely delivery, rather than identify a separate block of shares for each specific short sale. As a consequence, a lender may indicate to multiple potential borrowers that shares are available, but be unable to meet all demands. The requirement could also often be fulfilled by using published lists of easy-to-borrow securities (Welborn, 2008).
sales or market-maker FTDs were given three additional days. Similarly, in November 2012, new EU rules instituted a pre-trade borrowing requirement for short sales. Regulatory concerns have been further exacerbated by the widespread allegations and the extensive litigation about FTDs being used manipulatively by naked short sellers. Such “abusive” naked short selling is widely alleged to have contributed to the financial crisis by precipitating sharp price declines of financial firms.4 Notwithstanding extensive concerns, the SEC (in Report 450, March 2009) says that “there is hardly unanimity in the investment community or the financial media” on the dangers involved, and “despite its assertions regarding the potential of danger… the [SEC] Report can cite to no [relevant] bona fide studies” on the associated market impact. In the context of the strong regulatory focus, we begin by examining whether changes in FTDs are followed by changes in prices, pricing errors, intraday volatility, bid-ask spreads, and order imbalances. Our main sample consists of all the 1,492 NYSE ordinary common-share issues for which all relevant data are available over the period January 2005 to June 2008. For robustness, we replicate our core analysis on a similarly constructed sample of 2,381 Nasdaq ordinary common-share issues over the same period. First, we analyze liquidity and pricing efficiency for securities in different portfolios based on the number of FTDs. Second, we utilize ordinary least-squares (OLS) regressions and Granger-causality tests to investigate the link between FTD changes and market quality. Third, given the complex and endogenous interrelationships between market quality metrics and other market variables, we fit vector autoregressive models and test these relationships using impulse response functions. Fourth, we examine whether FTDs by market-makers affect securities differently from FTDs by other “public traders,” by using a proxy based on the reduction in FTDs subsequent to SEC Rule 204T in September 2008, which selectively precluded public-trader FTDs but not market-maker FTDs. We find that, during our sample period, FTDs affect about 95% of NYSE securities. For this NYSE sample, we conclude that an increase in FTDs equivalent to 0.1% of the number of outstanding shares leads to a 3% reduction in the magnitude of positive and absolute pricing errors; a 0.2% decline in intraday volatility; a 1.7% reduction in bid-ask spreads; and a 0.70% reduction in order imbalances. Each of these changes is statistically significant and not subject to significant subsequent corrections. Similar results are obtained in subsamples of securities affected by high levels of FTDs, or by persistent FTDs. Further, the beneficial 4 Over a two-year period leading up to the 2008 financial crisis, the SEC received more than 5,000 complaints in this regard (Wall Street Journal Asia, March 20, 2009), and a Factiva search shows over 4,600 printed English-language articles on the subject. Several investor associations and high-profile chief executive officers (CEOs) lobbied aggressively. The huge volume of litigation alleging associated stock price manipulation led to “naked” short selling being called the “Holy Grail….bigger than tobacco” for plaintiffs' lawyers (Stokes, 2009). Examples of lawsuits include the Biovail lawsuit against Stephen Cohen, Gradient, and others; the Overstock lawsuit against Rocker Partners, Gradient, and others; and the NFI lawsuit against Bank of America (the Specialist) and the Prime Brokers.
V. Fotak et al. / Journal of Financial Economics 114 (2014) 493–516
impact on pricing efficiency and liquidity is driven by FTDs of both market-makers and public traders. Overall, our results are consistent with failing-to-deliver traders acting as value arbitrageurs and liquidity suppliers, enhancing pricing efficiency and providing liquidity as needed. We observe that the regulatory focus on reducing FTDs has been particularly concerned with FTDs arising from short sales, despite extant literature documenting that short sales have a beneficial overall impact on pricing efficiency and liquidity by enabling value-traders to more effectively bring prices of overpriced securities in line with their true value, and facilitating financial intermediaries in providing liquidity more expeditiously.5 This suggests an underlying regulatory belief that short sales that fail to deliver have an impact on liquidity and pricing efficiency that is different from the corresponding impact of short sales that result in timely delivery. However, from an economic perspective, that should arguably not be the case. First, short sales that fail to deliver might not differ from timely delivered short sales at the time of the trade since a short seller need not definitively know on day t of the trade whether or not she will deliver or fail on settlement day tþ3. This is because that decision may rationally be taken only on day t þ3 on the basis of the rebate rates in the overthe-counter (OTC) stock-borrowing market on day tþ3.6 The short seller may want to borrow and deliver if these rebate rates are positive and fail if they are negative. Second, as explained by Culp and Heaton (2008), a FTD results either in automatic stock-borrowing at zero rebate rates from a pool of voluntary lenders under the DTCC Stock Borrow Program, or in forced stock-borrowing at zero rebate rates from a randomly assigned broker with a long stock position—and little incentive to enforce delivery by forcing a “buy-in” from the market, since the random assignment potentially changes every day (Putniņš, 2010). As far as the market is concerned, since the stock is being temporarily borrowed anyway (voluntarily or otherwise), there should not be any functional consequences arising from whether a short sale ultimately results in a FTD or not. Hence, any FTDs that arise from short sales should contribute to price discovery and liquidity on the day of trade in the same way as timely delivered short sales. In the context of the above, we extend our empirical analysis (on the impact of FTD changes on pricing efficiency and liquidity) to also examine the corresponding impact of delivered short sales, and test whether that impact is different. We find strong evidence that the impact on our market-quality metrics of daily changes in FTDs, and of
5 While there are some dissenting opinions, there is broad agreement behind an extensive literature finding that short selling leads to improvements in pricing efficiency and liquidity (Diamond and Verrecchia, 1987; Abreu and Brunnermeier, 2002, 2003; Miller, 1977; Bris, Goetzmann, and Zhu, 2007; Diether, Lee, and Werner, 2009; Boehmer, Jones, and Zhang, 2008). 6 Since stock-borrowing arrangements are ordinarily accompanied by same-day delivery, it is economically not rational for short sellers to borrow prior to the delivery date and pay extra borrowing fees (Geczy, Musto, and Reed, 2002). Evans, Geczy, Musto, and Reed (2009) show that short sellers fail to deliver when rebate rates are negative. Boni (2006) offers evidence of correlation between fails and borrowing costs, which she interprets as indicative of “strategic” fails.
495
daily changes in short interest net of FTDs (our proxy for delivered short sales), is similar in both magnitude and significance. To focus on FTDs more likely to originate from short sales, and to control for any omitted variables, we investigate the securities affected by an exogenous event: the imposition of a temporary SEC Emergency Order (Release 58166 dated July 15, 2008) that remained in force between July 21 and August 12, 2008. This order required pre-borrowing arrangements prior to all short sales in 19 selected financial stocks and hence did not directly affect any FTDs arising from “long” sales. We find that during the order validity period, reported FTDs for the affected securities reduce to about 12% of their pre-order average within one week, to about 2% of their pre-order average within three weeks, and finally, on the last day of the order period, reported FTDs are below the minimum reportable threshold for each and every affected security. We also find that securities affected by the order display a significant increase in intraday volatility, absolute pricing errors, and bid-offer spreads while the order was in force: i.e., the stock-borrowing prior to short sales mandate designed to reduce FTDs significantly worsened pricing efficiency and liquidity. While this natural experiment may not be generalizable because very few securities in the same industry and during a particularly turbulent time are affected, our results indicate that FTDs for these firms appear to arise largely from short sales, and contribute beneficially to market liquidity and pricing efficiency. In the context of the extensive concerns articulated in the press, and by investor groups and company CEOs, the final section of our empirical analysis investigates whether FTDs caused price crashes and distortions for specific securities during the 2008 financial crisis period.7 Since FTDs have been blamed for the price crashes of Bear Stearns Companies Inc. (“Bear Stearns”), Lehman Brothers Holdings Inc. (“Lehman”), Merrill Lynch & Co. Inc. (“Merrill”), and American Insurance Group (“AIG”), we test for high levels of FTD changes prior to the large price declines in the stock prices of those companies. We find that FTDs were too few for any significant stock price distortions on most days, and when FTDs did become abnormally high, it was after price declines, not before. In each case, FTDs were responding to information about the firms, rather than being responsible for triggering price declines. Overall, we find no evidence that FTDs played a causal role in the demise of financial institutions or in generating price distortions during the 2008 financial crisis. This research contributes to the literature on settlement systems and delivery failures. Fails are not actual “failures” of delivery, but delays with a median age of three days as per Boni (2006). Evans, Geczy, Musto, and Reed (2009) show that “the alternative to fail is valuable and important 7 Press sentiment is exemplified by statements such as “when Bear and Lehman made their final leap off the cliff of history, both undeniably got a push—especially in the form of a flat-out counterfeiting scheme called naked short selling” (Rolling Stone, October 2009). Concerned investor associations include The Movement for Market Reform, National Coalition against Naked Short Selling, and Coalition for the Reform of Regulation SHO. An example of a crusading high-profile CEO is Patrick Byrne of Overstock.com, which filed several lawsuits against alleged naked-short sellers and financial institutions they accused of facilitating “naked-shorting”.
496
V. Fotak et al. / Journal of Financial Economics 114 (2014) 493–516
to the pricing and trading of options.” Our results additionally and importantly show that the ability to fail has a beneficial impact on liquidity and pricing efficiency of equity markets as well. Earlier, Merrick, Naik, and Yadav (2005) have shown that the ability to fail is an important release valve for settlement-related pressures and manipulative distortions. Similarly, the difficulty of borrowing shares when short selling demand is high leads to costly frictions which negatively impact market quality; and the ability to fail is a release valve that helps to protect traders from the worst effects of these frictions: it is particularly valuable when stock-borrowing is so costly that short selling rebate rates become negative, which is exactly when liquidity is most needed in the stock-borrowing market (Evans, Geczy, Musto, and Reed, 2009).8 Hence, regulatory removal of the ability to fail for all public traders is debatable when progressive fines for settlement delays can arguably be effective without being an extreme solution. It is more important for regulators to focus on removing the economic incentives for delivery failures by improving the liquidity, transparency, and regulation of the stock-borrowing market. This research also contributes significantly to the short selling literature. First, extending the extant evidence on the market-quality benefits of short selling, our results suggest that short sales that fail to deliver are as beneficial for market quality as short sales that deliver on time. Second, our findings have important implications for the extensive press coverage of naked short selling, and the strong regulatory response to it. Finally, we extend the literature on the impact of short sales restrictions: the ban on FTDs in some stocks in July 2008 was effectively an additional constraint on short selling and our evidence shows that it had predictably negative consequences on market quality.9 The remainder of the paper is structured as follows. Section 2 outlines the data, the variables, and the measures utilized. Section 3 presents empirical results on the impact of FTD changes on pricing efficiency and liquidity. Section 4 investigates the role of FTD changes in creating price distortions during the 2008 financial crisis. Finally, Section 5 offers concluding remarks. 2. Data, variables, and measures 2.1. Data and sample The main sample consists of all NYSE listed common stocks in the Center for Research in Security Prices (CRSP) share codes 10 and 11 for which complete data are 8 These frictions are well documented by Kolasinski, Reed, and Ringgenberg (2013): They find that high levels of stock-borrowing demand lead to dramatic increases in loan fees and search costs related to supply constraints. 9 Several recent papers examine short selling restrictions in 2008: Boulton and Braga-Alves (2010) and Kolasinski, Reed, and Thornock (2013) examine aspects of the July/August 2008 Emergency Order that we investigate in this paper. Other papers—Battalio and Schultz (2011), Boehmer, Jones, and Zhang (2013), Autore, Billingsley, and Kovacs (2011), and Beber and Pagano (2013)—focus on the subsequent September 2008 short selling ban, which we do not examine.
available. By restricting the analysis to share codes 10 and 11, we exclude exchange-traded funds (ETFs) (in case FTDs are caused by delays in the creation of units) and other securities that are not common stocks. We thereby exclude securities classified by CRSP as certificates, American trust components, American Depositary Receipts (ADRs), shares of beneficial interest (SBIs), and units such as “depositary units” and “units of limited partnership interest,” as well as securities issued by other types of entities such as foreign firms, closed-end funds, and real estate investment trusts (REITs). We also restrict the sample to securities that are listed for at least six months during the sample period (in order to have adequate data to fit vector autoregressive models) and that have been trading for at least one year prior to the sample period. Having a year of trading data allows for the estimation of pricing errors as discussed later in this section. We further require that the number of shares outstanding does not vary by more than 10% over any single day, to control for the unusual volumes of FTDs around primary market transactions documented by Edwards and Hanley (2010); though our results are robust to resampling without this constraint. In view of the potentially confounding influence of the July 2008 Emergency Order and the restrictions on short selling and FTDs subsequently imposed in September and October 2008, we confine our analysis to the period up to June 30, 2008. The main sample thus consists of 1,492 NYSE securities over January 2005 to June 2008. For robustness tests, we construct a comparable sample of 2,381 Nasdaq securities for the same period. Short interest data are from www. shortsqueeze.com and short sales data (available from January 2005 onwards) are from the NYSE. The number of shares outstanding is from CRSP. Market quality measures are based on NYSE Trade and Quote (TAQ) data. 2.2. Outstanding fails-to-deliver ratio (OFR) and outstanding delivered shorts ratio (ODR) Our metric for FTDs is the Outstanding fails-to-deliver ratio (OFR), defined for each day t as the number of outstanding failed positions reported on day tþ3 scaled by the total number of shares outstanding of the firm. We obtain data on the number of shares outstanding from CRSP, while data on outstanding FTDs have been made available by the SEC under the Freedom of Information Act (FOIA) since March 22, 2004. Reported raw FTD data from the SEC represent the open interest of failed-to-deliver trades that occurred up to three days prior. For each day t, we look at the number of outstanding FTDs on day tþ3, as we are interested in investigating the impact of FTD changes on the market on the day on which the transaction actually occurs (day t), rather than on the day on which the resulting failures are disclosed (day t þ3). Hence, OFR on day t is equal to the number of FTDs for day t þ3, scaled by shares outstanding on day t. Our approach of scaling the number of FTDs by the number of shares outstanding is similar in spirit to previous literature that relies on measures of “relative short interest” (DeChow, Hutton, Meulbroek, and Sloan, 2001; Desai, Thiagarajan, Ramesh, and Balachandran, 2002; Chen and Singal, 2003), and mirrors recent studies using FTD-based data (for example,
V. Fotak et al. / Journal of Financial Economics 114 (2014) 493–516
Boulton and Braga-Alves, 2010). FTD changes are measured as the daily change in OFR. Over our sample period, the SEC reports the number of outstanding FTDs only for a trading day on which it exceeds 10,000 shares of a particular security. Accordingly, on a day on which no figure is reported for a particular security, we are unable to determine the true number of outstanding FTDs—we simply know that it is less than 10,000 shares. In the absence of reported data, we record the number of FTDs as zero. Our approach is consistent with extant literature using FTD data (for example, Boulton and Braga-Alves, 2010). Accordingly, reported figures somewhat understate the true extent of fails. To put this into perspective, we note that the average number of shares outstanding for the securities in our NYSE sample is about 212 million. Further, over the days on which nonzero FTDs are reported, the average number of reported FTDs is 97,068 shares, which is well in excess of the minimum reporting threshold. In terms of our tests regarding the market impact of FTDs, this data bias makes our findings more conservative, rather than less. First, by inducing a level of noise, it is lowering the statistical power of our tests. Second, it is arguably unlikely that securities with low levels of FTDs would suffer from distortions in market quality due to failed trades, given that we find that securities with high levels of FTDs do not display such problems. In order to examine whether claims regarding the market impact of “naked” short selling are valid, we construct a control variable, the Outstanding delivered shorts ratio, or ODR, which is the estimated short ratio net of FTDs. The Outstanding delivered shorts ratio is intended to be the open interest of timely delivered short sales positions scaled by the number of shares outstanding. We estimate the open interest of timely delivered short sales by the difference between the estimated short interest and the number of outstanding failed positions. The implicit assumption in this estimate of delivered short sales is that all FTDs arise from short selling. Additionally, while we have daily data on short sales from TAQ, short interest data are biweekly; accordingly, we calculate the short interest on a day within a biweekly period as the sum of the short interest reported at the start of the period and the cumulative daily changes in short interest since the start of that period. The daily change in short interest is calculated as the difference between the daily total short sales and an estimate of the daily short positions covered. Following Diether, Lee, and Werner (2009), we assume that the number of shares covered on each day of the biweekly interval is equal to the total short sales over the period minus the change in short interest over that period, divided by the number of trading days in that period. We use the daily change in ODR as our measure of timely delivered short sales. 2.3. Pricing error We define the “pricing error” on any day as the difference between the observed price on that day and the estimated information-efficient price for the same day. The unobservable daily estimate of the informationefficient random-walk or “fundamental” price of the security, a latent stochastic variable, is estimated for each sample
497
security using a Kalman-filter methodology as in Hamilton (1985). The procedure involves establishing two equations. The first equation dictates the evolution of the latent variable, and in our case we assume, in the spirit of Hasbrouck (1993), that the logarithm of the stock's underlying or information-efficient value, F(t), follows a random walk with a drift, μ, and a white noise innovation, ε(t), with mean zero and variance σ 2ε : FðtÞ ¼ μ þFðt 1Þ þ εðtÞ;
ε Nð0; σ 2ε Þ:
ð1Þ
The second equation relates the observed and latent variables, i.e., specifies the pricing error process. We assume that the pricing error Y(t) follows a meanreverting process around zero, with α, the rate of meanreversion, ranging between zero and one. Pricing errors correct fully in one period when α is equal to one, and do not correct at all when α is equal to zero:
ΔYðtÞ ¼ αYðt 1Þ þ ϕðtÞ;
ϕ Nð0; σ 2ϕ Þ:
ð2Þ
The observed log of stock price S(t) is the sum of the fundamental price and pricing error: SðtÞ ¼ FðtÞ þYðtÞ:
ð3Þ
Hence; SðtÞ ¼ μ þ ð1 αÞSðt 1Þ þ αFðt 1Þ þ θðtÞ;
θðtÞ ¼
ϕðtÞ þ εðtÞ: The Expectation Maximization (EM) algorithm (Dempster, Laird, and Rubin, 1977) is employed to compute the Maximum Likelihood (ML) estimate of the unobservable variable, F(t), based on data relating to the observed variable, S(t). Hamilton (1985) employs such an approach to estimate expected quarterly inflation, the latent variable, based on observed actual inflation. In exactly the same way, we utilize the observed daily stock prices to infer the daily unobserved fundamental price, and hence the daily pricing error, using daily closing price data from CRSP. The statespace representation of the system is as follows: Measurement equation: 2 3 SðtÞ 6 7 ½SðtÞ ¼ 1 0 0 4 FðtÞ 5 ð5Þ εðtÞ Transition equation: 2 3 2 SðtÞ ð1 αÞ α 6 FðtÞ 7 6 1 4 5¼4 0
εðtÞ
0
0
3 3 2 Sðt 1Þ θðtÞ 7 76 7 6 0 54 Fðt 1Þ 5 þ 4 εðtÞ 5 0 1 0
μ
32
ð6Þ
To test the bias and efficacy of the pricing error estimation process, we run 500 simulations of both fundamental price and pricing error for a hypothetical stock over 252 trading days assuming a range of volatility parameters and mean-reversion parameters. In each case, we add the fundamental price and the pricing error to arrive at the equivalent of a simulated “observed” price. Then, we run our Kalman-filter estimation procedure on this observed price series to determine our Kalman-filter estimate of the originally simulated fundamental price. Finally, we run a regression of changes in the originally simulated fundamental price on changes in our Kalman-filter estimate of that fundamental price. In each and every case, the regression intercept is not significantly different from zero, and
498
V. Fotak et al. / Journal of Financial Economics 114 (2014) 493–516
the regression slope is not significantly different from one; and the root mean square error in the estimated fundamental price is economically small in magnitude. We employ three measures of pricing efficiency related to pricing errors. First, we use duly-signed pricing errors as estimated. Second, we use the absolute values of pricing errors, which given the zero-mean distribution of pricing errors, are also a direct measure of the volatility of pricing errors. Finally, given that the impact of FTDs on pricing errors could be asymmetric, we use a dummy for positive pricing errors to distinguish between positive and negative pricing errors. 2.4. Liquidity measures and other variables The liquidity measures we use are relative order imbalance, bid-ask spread, and trading volume defined as in Table 1. Our measure of volatility is the intraday volatility computed as the daily standard deviation of five-minute stock returns. The definitions of all of our variables are based on extant empirical literature analyzing NYSE TAQ data, and are summarized in Table 1. 3. Empirical results: FTDs, pricing efficiency, and liquidity 3.1. Preliminary descriptive analysis of overall sample We sort securities into ten deciles based on mean OFR, computed over the entire sample period (January 1, 2005 to June 30, 2008). In Table 2, we report means, medians, and standard deviations of the main variables, for the overall sample and for the top and bottom deciles. We find that overall mean OFR is about 0.04%. By construction, OFR varies by decile: the mean is less than 0.01% in decile 1 and about 0.29% in decile 10. ODR is similarly higher in
decile 10 (13.35%) than in decile 1 (4.08%), with an almost monotonic positive relationship between mean OFR and mean ODR across (OFR-based) deciles of different securities. However, even though securities with higher mean OFR tend to also have a higher mean ODR, the average timeseries correlation between ODR and OFR of a particular security is, in contrast, extremely small—only 0.08. Securities in deciles 1 and 10 do not differ significantly in terms of mean pricing errors. Securities with higher OFR display significantly higher positive pricing errors and absolute pricing errors, which is consistent with FTDs intensifying when securities are overpriced, although we should not infer causality from this correlation. Similarly, higher OFR is associated with higher and positive order imbalances and higher intraday volatility, but lower spreads. Our results also indicate that FTDs are significantly more common for smaller firms, as average market capitalization for firms in decile 1 (USD 9.1 billion) is almost six times larger than for firms in decile 10 (USD 1.6 billion). As discussed in Section 2.2, FTD reporting is subject to a minimum threshold of 10,000 shares. It is reasonable to conjecture that this reporting threshold is relatively more onerous for securities with smaller market capitalization and, likely, a lower number of shares outstanding. That is not actually so. The proportion of security-days with no reported FTDs is 97.6% in OFR-decile 1, and drops monotonically across deciles to 48.0% for securities in decile 10; the daily average number of outstanding FTDs for the days on which FTDs exceed the 10,000 share-reporting threshold ranges from 29,031 for securities in decile 1 to 357,238 for securities in decile 10. Since securities with smaller mean market capitalization generally have higher mean OFR, they also have a relatively larger average number of reported FTDs (in number of shares) on the days on which
Table 1 Variable definitions. This table lists the variables used in the analysis and provides definitions. All variables are daily, unless otherwise specified. FTDs and short selling Outstanding fails-to-deliver ratio (OFR) Outstanding delivered shorts ratio (ODR) Pricing error Pricing error (PE) Negative pricing error (Negative PE) Positive pricing error (Positive PE) Absolute pricing error (Abs PE) Positive PE dum Liquidity Order imbalance (OIB)
For each day t, ratio of the number of outstanding fails-to-deliver reported on day t þ3, over total shares outstanding on day t. Ratio of the estimated number of outstanding delivered shorted shares over total shares outstanding.
The non-random walk component of a daily return series estimated using a Kalman-filter methodology. A variable set equal to the pricing error when pricing error is negative, to zero otherwise. A variable set equal to the pricing error when pricing error is positive, to zero otherwise. The absolute value of the pricing error. A binary variable set equal to one if pricing error is positive, to zero otherwise.
Positive OIB dum Spread Volume
The daily sum of the five-minute difference between the market value of shares traded in buyer-initiated trades and the market value of shares traded in seller-initiated trades, divided by total daily dollar trading volume. A binary variable set equal to one if OIB is positive and to zero otherwise. The daily average of the ratio of the difference between bid and ask and the mid-quote. Daily number of shares traded.
Other Price Return Intraday volatility Market value
The The The The
natural log of the daily average of the five-minute midquote. daily stock price return. daily average of standard deviation of the five-minute stock price return. number of shares outstanding multiplied by the closing price for the day.
V. Fotak et al. / Journal of Financial Economics 114 (2014) 493–516
499
Table 2 Summary statistics. All variables are as defined in Table 1. The Outstanding fails-to-deliver ratio (OFR) is computed for all NYSE common stock of US-based firms (CRSP share codes 10 and 11) listed for at least 18 months over the interval January 1, 2005 to June 30, 2008, and with no large daily change ( 410%) in the number of shares outstanding and for which all required data are available (n¼ 1,492). Securities are ranked by mean OFR and allocated to decile 1 (lowest) through 10 (highest). Daily statistics are computed by security, security averages are further averaged over deciles and for the entire sample. This table reports mean, median, and standard deviation for the sample, means for deciles 1 and 10, and the difference between those, along with results of a t-test for differences in means across deciles 1 and 10. “n”, “nn”, and “nnn” indicate significance at the 10%, 5%, and 1% level, respectively. Sample
Outstanding fails-to-deliver ratio Outstanding delivered shorts ratio Pricing error Positive pricing error Abs pricing error Price (log) Intraday volatility Return Spread Order imbalance Market value (US$ M) Obs
OFR decile means
Difference
Mean
Median
Std
OFR D1
OFR D10
OFR D10–D1
p-Value
0.04% 5.46% 0.15% 0.51% 1.18% 3.3001 0.21% 0.07% 0.18% 6.72% $8,223.33 1,492
o0.01 4.34% o 0.01% 0.10% 0.21% 3.4171 0.20% 0.06% 0.11% 6.92% $2,087.97 1,492
0.18% 5.32% 4.33% 1.47% 4.96% 0.7631 0.08% 0.11% 0.27% 5.71% $24,212.97 1,492
o 0.01 4.08% o 0.01% 0.37% 0.74% 3.1484 0.21% 0.05% 0.39% 4.00% $9,128.81 149
0.29% 13.35% 0.55% 1.06% 2.66% 2.9121 0.28% 0.03% 0.26% 8.26% $1,598.89 150
0.29% 9.27% 0.55% 0.68% 1.92% 0.2363 0.08% 0.02% 0.13% 4.26% $7,529.91 149
o 0.01nnn o 0.01nnn 0.35 o 0.01nnn o 0.01nnn 0.03nn o 0.01nnn 0.26 o 0.01nnn o 0.01nnn o 0.01nnn
FTDs are reported, and a lower proportion of days with no reported FTDs.
3.2. Portfolio approach As a first test of the relationship between FTD changes and subsequent market quality, on each day t in our sample spanning January 2005 to June 2008, we group the 1,492 securities into nine portfolios based on changes in OFR and ODR. We start by estimating the daily time-series standard deviation in OFR by security. On each day, we include securities with a one-standard deviation or greater decrease in OFR into a “FTD decrease” portfolio and, similarly, securities with a one-standard deviation or greater increase in OFR into a “FTD increase” portfolio. We include all remaining securities into a “FTD constant” portfolio. In order to control for the extent of timely delivered short sales, we replicate the same procedure on the basis of timely delivered short sales, proxied by changes in ODR, thus forming the portfolios “Delivered short sales decrease,” “Delivered short sales constant,” and “Delivered short sales increase”. Finally, we intersect those groups of securities, forming nine final portfolios. For each portfolio, we compute average changes in our metrics of price levels, intraday volatility, market liquidity, pricing errors, absolute pricing errors, and order imbalances for the following day (t þ1). All variables are standardized and winsorized (at three standard deviations) by security. We compute nextday average changes for the portfolios “FTD decrease, Delivered short sales constant” and “FTD increase, Delivered short sales constant” and we further test for differences between these averages. Our interest lies in changes in OFR. We form portfolios on changes in ODR to control for the impact of estimated delivered short sales on liquidity and pricing efficiency. We present results for the two key portfolios (the ones
including low and high levels of FTD changes while experiencing normal levels of delivered shorts), rather than for the full set of nine portfolios, primarily for the sake of brevity. Our focus on the portfolios with high or low levels of FTD changes and “normal” levels of delivered shorts, allows us to draw inferences about FTD changes under “normal” short selling conditions. Also, the samples with abnormal levels of both FTDs and estimated delivered shorts vary greatly in size, leading to comparisons of data samples of dramatically different sizes. The results of this analysis are presented in Table 3. For the portfolio “FTD increase, Delivered short sales constant,” we find that, on the following day, there is a 4.2% reduction in absolute pricing error, a 0.1% decrease in spreads, a 0.4% decrease in order imbalances, and a 0.67% decrease in intraday volatility—all results being statistically significant at the 1% level, except for the result on order imbalances, which is significant at 5%. We also find a decrease in price and pricing error, but estimates are not statistically significant. For the portfolio “FTD decrease, Delivered short sales constant,” we similarly observe a decrease in absolute pricing error, but the magnitude of the estimated effect is smaller (0.18%). We further investigate the differences in next-day metrics between the two portfolios. The “FTD increase” portfolio, compared to the “FTD decrease” portfolio, displays lower absolute pricing errors, spreads, order imbalances, and intraday volatility, with all results being statistically significant at the 1% level. This initial analysis offers evidence that large FTD changes are followed by next-day improvements in market liquidity and pricing efficiency. Given the limitations of this univariate framework, in the following section we employ OLS regressions to investigate the relationship between FTD changes and subsequent market quality while controlling for timely delivered short sales and relevant systematic factors.
500
V. Fotak et al. / Journal of Financial Economics 114 (2014) 493–516
Table 3 FTDs and market quality, portfolio approach. All variables are as defined in Table 1. All variables are standardized and winsorized by security. The sample includes all NYSE common stock of US-based firms (CRSP share codes 10 and 11), listed in the CRSP and TAQ databases for at least 18 months over the interval January 1, 2005 to June 30, 2008, and with no large changes ( 410%) in the number of shares outstanding. The resulting sample includes 1,492 securities. On each day t between January 2005 and June 2008, securities with a one-standard deviation or greater decrease in OFR are allocated to a “FTD decrease” portfolio and, similarly, securities with a one-standard deviation or greater increase in OFR to a “FTD Increase” portfolio. All remaining securities are allocated to a “FTD constant” portfolio. The same procedure is repeated on the basis of the intensity of delivered short selling, proxied by changes in ODR, thus forming the portfolios “Delivered short sales increase,” “Delivered short sales decrease,” and “Delivered short sales constant.” Those portfolios are intersected, forming nine final portfolios. Averages of next-day (tþ 1) changes in returns, return volatility, market liquidity, pricing errors, and order imbalances are computed. The table presents results for the “FTD increase, Delivered short sales constant” and the “FTD decrease, Delivered short sales constant” portfolios, along with results of a test for differences between them. p-Values are from two-sided t-tests. “n”, “nn”, and “nnn” indicate significance at the 10%, 5%, and 1% level, respectively. FTD increase, delivered short sales constant
p-Value
FTD decrease, delivered short sales constant
p-Value
Difference
p-Value
2.31% 0.28% 4.20% 0.10% 0.40% 0.67%
0.31 0.41 o 0.01nnn o 0.01nnn 0.01nn o 0.01nnn
0.15% 0.04% 0.18% o 0.01% 0.09% 0.01%
0.36 0.36 o0.01nnn 0.22 0.12 0.14
2.46% 0.24% 4.02% 0.10% 0.31% 0.66%
0.75 0.77 o 0.01nnn o 0.01nnn o 0.01nnn o 0.01nnn
ΔPrice(tþ 1) ΔPE(tþ 1) ΔAbs PE(t þ1) ΔSpread(t þ1) ΔOIB(tþ 1) ΔIntraday volatility (tþ 1) N (days)
844
849
3.3. Panel OLS regressions We next estimate panel regressions to investigate the relationship between FTD changes and next-day marketquality metrics. Accordingly, we estimate six separate regressions, with changes in pricing errors, absolute pricing errors, prices, intraday volatility, spreads, and order imbalances as responses. Our main explanatory variable is the previous-day change in OFR. Given the regulatory focus on reducing FTDs arising from short sales, and the welldocumented impact of short sales on liquidity and pricing efficiency, our empirical tests aim to control for the estimated impact of delivered short sales on market quality. Accordingly, we also add the previous-day change in ODR to the model. As further controls, we include, in each regression, previous-day changes of the dependent variable, to account for possible autocorrelations. In addition, to control for systematic effects, we include market averages of the changes in the same quality metrics. Finally, when modeling changes in pricing errors, we add interaction variables between previous-day changes in OFR and ODR and a binary variable identifying positive pricing errors, to allow for an asymmetric impact of fails on pricing errors. All regressions contain security fixed-effects. The models estimated are:
ΔPEi;t ¼ λ0;i þ λ1 ΔOFRi;t 1 þ λ2 ΔODRi;t 1 þ λ3 ΔOFRi;t 1 Positive_PE_Dumi;t 1 þ λ4 ΔODRi;t 1 Positive_PE_Dumi;t 1 þ λ5 ΔPEM;t þ λ6 ΔPEi;t 1 þ δi;t ð7Þ ΔPE_Volatilityi;t ¼ γ 0;i þ γ 1 ΔOFRi;t 1 þ γ 2 ΔODRi;t 1 þ γ 3 ΔPE_VolatilityM;t þ γ 4 ΔPE_Volatilityi;t 1 þ νi;t
ð8Þ
ΔVolatilityi;t ¼ δ0;i þ δ1 ΔOFRi;t 1 þ δ2 ΔODRi;t 1 þ δ3 ΔVolatilityM;t þ δ4 ΔVolatilityi;t 1 þ θi;t
ð9Þ
844
ΔSpreadi;t ¼ β0;i þ β1 ΔOFRi;t 1 þ β2 ΔODRi;t 1 þ β3 ΔSpreadM;t þ β4 ΔSpreadi;t 1 þ εi;t
ð10Þ
ΔOIBi;t ¼ κ 0;i þ κ 1 ΔOFRi;t 1 þ κ 2 ΔODRi;t 1 þ κ 3 ΔOIBM;t þ κ 4 ΔOIBi;t 1 þ ρi;t ð11Þ ΔLogmidi;t ¼ μi þ μ1 ΔOFRi;t 1 þ μ2 ΔODRi;t 1 þ μ3 ΔLogmidM;t þ μ4 ΔLogmidi;t 1 þ σ i;t :
ð12Þ
We estimate the above parameters with OLS regressions using daily data and standard errors are time-clustered. Variable definitions are included in Table 1; the i and t subscripts index, respectively, securities and days. The M subscript indicates a “market” equally weighted average computed over our entire sample. Δ is used to identify daily changes (for example, ΔOFRt ¼OFRt OFRt 1). Results for the overall sample are presented in Table 4. As reported, an increase in OFR is associated with a nextday decrease in absolute pricing errors, spreads, order imbalances, and intraday volatility (all significant at 1%). We find no significant relationship between increases in FTDs and price levels, nor with positive and negative pricing errors, when those are separately modeled. This second set of results is highly consistent with our previous findings: an increase in FTDs is related to improvements in liquidity and pricing efficiency. We recognize that the correlations documented in this analysis are not proof of causation. Hence, we attempt to gain insights into the direction of causality in our system by first testing for Granger-causality. For each metric of interest, we test both directions of Grangercausality: that is, we test whether previous-day FTD changes “Granger-cause” next-period market-quality metrics. We first find that FTD changes Granger-cause absolute pricing errors, spreads, order imbalances, and intraday volatility. Yet, when testing the other side of causality, we find it equally meaningful, as each of those variables Granger-causes FTD changes as well. FTD changes and market-quality metrics
V. Fotak et al. / Journal of Financial Economics 114 (2014) 493–516
501
Table 4 FTDs and market quality, OLS regressions. Coefficient estimates and p-values (in italics) from panel regressions with firm fixed effects and time-clustered standard errors. Metrics of market quality (in column headings) are regressed on lagged values of the predictors (listed in the first column). All variables are standardized and winsorized by security, and defined as in Table 1. The sample includes all NYSE common stocks of US-based firms (CRSP share codes 10 and 11), listed in the CRSP and TAQ databases for at least 18 months over the interval January 1, 2005 to June 30, 2008, and with no large changes ( 410%) in the number of shares outstanding. The resulting sample includes 1,492 securities and 1,001,656 firm-days. The “Market” variable on day t is obtained by averaging the related variable across all securities on day t. Δ is used to identify changes (for example, ΔOFRt ¼OFRt OFRt 1). “n”, “nn”, and “nnn” indicate significance at 10%, 5%, and 1% level, respectively.
ΔOFR(t 1) ΔODR(t 1) ΔPricing error (t 1) ΔOFR(t 1) Positive PE(t 1) ΔODR(t 1) Positive PE(t 1)
ΔPricing error(t)
ΔAbs PE(t)
ΔPrice(t)
ΔIntraday volatility(t)
ΔSpread(t)
ΔOIB(t)
0.0019 0.4127 0.1932 o 0.01nnn 0.3489 o 0.01nnn 0.0048 0.1375 0.4240 o 0.01nnn
0.0069 o 0.01nnn 0.2733 o 0.01nnn
o 0.0001 0.9718 0.0042 0.03nn
0.0068 o0.01nnn 0.0721 o0.01nnn
0.0125 o0.0nnn 0.1008 o0.01nnn
0.0093 o 0.01nnn 0.2599 o 0.01nnn
ΔAbs PE(t 1)
0.3763 o 0.01nnn
ΔPrice(t 1)
0.0255 o0.01nnn
ΔIntraday volatility(t 1)
0.1937 o0.01nnn
ΔSpread(t 1)
0.3214 o0.01nnn
ΔOIB(t 1) ΔMarket pricing error(t)
0.4480 o 0.01nnn 0.8559 o 0.01nnn
ΔMarket abs PE(t)
0.8363 o 0.01nnn
ΔMarket price(t 1) ΔMarket intraday volatility(t) ΔMarket spread(t) ΔMarket OIB(t)
Granger-cause each other, and the whole system displays feedback effects over time. Hence, it is necessary for us to investigate the impact of FTD changes within a statistical model that allows us to control for intertemporal interdependency amongst all metrics. Accordingly, in the next section, we present results from vector autoregressive (VAR) models, duly accounting for endogenous interrelationships between marketquality metrics, FTD changes, and timely delivered short sales. 3.4. Vector autoregressive framework In a third set of tests, we control for endogenous interrelationships using three vector autoregressive (VAR) models, with additional exogenous variable(s) added as controls. The system of equations underlying each of these models is formally described in Table 5. In VAR Model 1, the variables included are changes in OFR, ODR, Pricing error, Intraday volatility, Spread, and Order imbalance. We add, as predictors in the pricing error equation, (1) an interaction between lagged changes in OFR and a lagged binary variable set equal to one when pricing error is positive, and (2) an interaction between lagged changes in ODR and
0.9971 o0.01nnn 0.9150 o0.01nnn 0.8751 o0.01nnn 0.8016 o 0.01nnn
a lagged binary variable set equal to one when pricing error is positive. Accordingly, we are able to separately estimate the impact of FTD changes and delivered short sales on pricing error when the pricing error is positive in contrast to when the pricing error is negative. We add two more predictors to each of the equations in which the changes in OFR or ODR are the response variables: (3) a lagged binary variable set equal to one when order imbalance is positive and zero otherwise, and (4) a lagged binary variable set equal to one when pricing error is positive and zero otherwise. Finally, in each equation we add, as an exogenous variable, the market-wide equally weighted average of the dependent variable, to control for possible systematic effects. In VAR Model 2, we replace Pricing error by Absolute pricing error; given that we do not model pricing error here, we do not include the interaction-variables described above in points (1) and (2). VAR Model 3 is similar to Model 2, but we replace Pricing error by Price. We estimate VAR Models 1, 2, and 3 separately for each security. The results we report in Table 5 are obtained by first standardizing all variables by subtracting the securityspecific mean and dividing by the security-specific standard
502
V. Fotak et al. / Journal of Financial Economics 114 (2014) 493–516
Table 5 FTDs and market quality, VAR analysis. Results for the estimation of three vector autoregressive models of order one used to investigate the impact of changes in OFR and ODR: VAR with Pricing error (Model 1), VAR with Pricing error volatility (Model 2), and VAR with Price (Model 3). All variables are defined in Table 1 and are standardized and winsorized by security. The “entire market” sample includes all common stocks of US-based firms (CRSP share codes 10 and 11) with NYSE as the primary exchange, listed in the CRSP and TAQ databases for at least 18 months over the interval January 1, 2005 to June 30, 2008, and with no large changes (410%) in the number of shares outstanding. For the “Most FTDs” sample, securities from the sample described above are ranked by mean OFR and allocated on that basis to decile 1 (lowest) through 10 (highest); the securities included in decile 10 constitute the “Most FTDs” sample. The VAR models are defined below; the subscripts M and t indicate, respectively, the market-wide average of the variable of interest and day t; Δ is used to identify changes (for example, ΔOFRt ¼ OFRt OFRt 1). Panels A and B are extracts of estimates of the parameters in Models 1, 2, and 3 above and report results pertaining to the impact of, respectively, OFR and ODR. Reported parameter estimates are averages of parameters estimated by security. The p-values from two-sided t-tests are in italics below the parameter estimate. “n”, nn “ ”, and “nnn” indicate significance at the 10%, 5%, and 1% level, respectively. Panels C and D present the effect on the market-quality metrics included in Models 1, 2, and 3 of an increase in Outstanding FTD ratio (OFR) and Outstanding delivered shorts ratio (ODR) equivalent to 10 basis points of the number of outstanding shares. The panels report the impact on the response variable as a proportion of the mean of the response variable. Percentages presented in bold are significant at 10% or lower. Panels E and F are extracts of estimates of the parameters in Models 1, 2, and 3 above and report results pertaining to the impact of OFR. Panel E presents results from a subset of the “entire market” data set discussed in Panels A–D; the sample includes all common stocks of US-based firms (CRSP share codes 10 and 11) with NYSE listed as the primary exchange, listed in the CRSP and TAQ databases for at least 18 months over the interval January 1, 2005 to June 30, 2008, with no large changes (410%) in the number of shares outstanding and that appear, for at least one day, on the NYSE-published list of Threshold Securities. Panel F presents results from a data set including Nasdaq securities: all common stocks of US-based firms (CRSP share codes 10 and 11) with Nasdaq listed as the primary exchange, available in the CRSP and TAQ databases, listed for at least 18 months over the interval January 1, 2005 to June 30, 2008, and with no large changes (410%) in the number of shares outstanding. Reported parameter estimates are averages of parameters estimated by security. Significance is tested employing a cross-sectional estimate of the standard error of the parameter estimate. The p-values from two-sided t-tests are in italics below the parameter estimate. “n”, “nn”, and “nnn” indicate significance at the 10%, 5%, and 1% level, respectively. Model1: ΔY1t ¼ c1 þ β1ΔY1t 1 þ M1t þ S1t þ S2t þ ε1t Model2: ΔY2t ¼ c2 þ β2ΔY2t 1 þ M2t þ S1t þ ε2t
ε1t i:i:d: Nð0; ΩÞ
ε2t i:i:d: Nð0; ΩÞ
Model3: ΔY3t ¼ c3 þ β3ΔY3t 1 þ M3t þ S1t þ ε3t ε3t i:i:d: Nð0; ΩÞ 1 1 1 0 0 0 OFRt OFRt OFRt C C C B B B ODRt C B ODRt C B B ODRt C C C C B B B C C C B B B PEt C B B PE Volatilityt C B Pricet C Y1t ¼ B C; Y2t ¼ B C; Y3t ¼ B C B Volatilityt C B V olatilityt C B Volatilityt C C C C B B B C C B Spreadt C B B Spread Spread t t A A A @ @ @ OIBt OIBt OIBt 3 3 3 2 2 2 ψ1 ΔOFRM;t ψ7 ΔOFRM;t ψ13 ΔOFRM;t 7 7 7 6 ψ2 ΔODR 6 6 ψ8 ΔODRM;t 7 7 6 6 6 ψ14 ΔODRM;t 7 M;t 7 7 7 6 6 6 7 7 7 6 6 6 ψ3 ΔPE ψ9 ΔPEVolatility ψ15 ΔPrice M;t M;t M;t 7 7 7 6 6 6 M1t ¼ 6 7; M2t ¼ 6 7; M3t ¼ 6 7 6 ψ4 ΔVolatilityM;t 7 6 ψ10 ΔVolatilityM;t 7 6 ψ16 ΔVolatilityM;t 7 7 7 7 6 6 6 6 ψ5 ΔSpreadM;t 7 6 ψ11 ΔSpreadM;t 7 6 ψ17 ΔSpreadM;t 7 5 5 5 4 4 4 ψ6 ΔOIBM;t ψ12 ΔOIBM;t ψ18 ΔOIBM;t 3 2 τ1 Positive OIB dumt 1 þ τ2 Positive PE dumt 1 7 6 τ3 Positive OIB dum t 1 þ τ4 Positive PE dumt 1 7 6 7 6 7 6 0 7 S1t ¼ 6 7 6 0 7 6 7 6 5 4 0 0
3 0 7 6 0 7 6 7 6 6 τ5 ðPositive PE dumt 1 ΔOFRt 1 Þ þ τ6 ðPositive PE dumt 1 ΔODRt 1 Þ 7 7 6 S2t ¼ 6 7 0 7 6 7 6 5 4 0 2
0 ΔX t ¼ X t X t1
Panel A: Impact of FTDs
A: Entire market 1,407 Securities B: Most FTDs 150 Securities A: Entire market 1,407 Securities B: Most FTDs 150 Securities A: Entire market 1,407 Securities B: Most FTDs 150 Securities
ΔPE
ΔPE (incremental effect when lag PE 40)
0.0045 0.20 0.0451 o 0.01nnn
0.0168 o 0.01nnn 0.0785 o 0.01nnn
ΔAbs PE
ΔPrice
ΔIntraday volatility
ΔSpread
ΔOIB
0.0002 0.70 o 0.0001 0.99
0.0054 o 0.01nnn 0.0121 0.11 0.0071 o 0.01nnn 0.0189 0.01nn 0.0049 0.02nn 0.0133 0.08n
0.0126 o0.01nnn 0.0204 o0.01nnn 0.0095 o0.01nnn 0.0168 o0.01nnn 0.0103 o0.01nnn 0.0192 0.05nn
0.0093 o 0.01nnn 0.0130 0.31 0.0071 0.04nn 0.0115 0.37 0.0068 0.05nn 0.0140 0.28
0.0077 o 0.01nnn 0.0361 o 0.01nnn
V. Fotak et al. / Journal of Financial Economics 114 (2014) 493–516
503
Table 5 (continued ) Panel B: Impact of delivered short selling
A: Entire market 1,407 Securities B: Most FTDs 150 Securities A: Entire market 1,407 Securities B: Most FTDs 150 Securities A: Entire market 1,407 Securities B: Most FTDs 150 Securities
ΔPE
ΔPE (incremental effect when lag PE40)
0.5414 o 0.01nnn 0.8416 o 0.01nnn
1.3759 o 0.01nnn 1.4300 o 0.01nnn
ΔAbs PE
ΔPrice
0.8516 o 0.01nnn 0.7913 o 0.01nnn 0.0221 o 0.01nnn 0.0196 0.04nn
ΔIntraday volatility
ΔSpread
ΔOIB
0.9128 o 0.01nnn 0.9759 o 0.01nnn 0.9594 o 0.01nnn 1.0141 o 0.01nnn 0.9126 o 0.01nnn 0.9872 o 0.01nnn
0.2267 o 0.01nnn 0.0690 0.06n 0.2104 0.09n 0.0615 0.09n 0.2106 o 0.01nnn 0.0762 0.04nn
0.7665 o 0.01nnn 0.6021 o 0.01nnn 0.7629 o 0.01nnn 0.6035 o 0.01nnn 0.7628 o 0.01nnn 0.6079 0.01nn
Panel C: Impact of a change in Outstanding FTD ratio (OFR) equal to 10 basis points of the number of outstanding shares Overall sample Response variable Positive PE Abs PE Price Intraday volatility Spread OIB
Model 1 (PE)
Model 2 (Abs PE)
Most FTDs sample Model 3 (Price)
Model 1 (PE)
3.24%
Model 2 (Abs PE)Model 3 (Price)
2.59% 3.00%
0.18% 1.68% 0.73%
3.84% o 0.01% 0.17% 1.77% 0.55%
0.24% 1.27% 0.53%
0.12% 0.70% 0.35%
0.18% 0.58% 0.31%
o 0.01% 0.13% 0.66% 0.38%
Panel D: Impact of a change in Outstanding delivered shorts ratio (ODR) equal to 10 basis points of the number of outstanding shares Overall sample Response variable Positive PE Abs PE Price Intraday volatility Spread OIB
Model 1(PE)
Model 2 (Abs PE)
Most FTDs sample Model 3 (Price)
Model 1 (PE)
4.52%
Model 2 (Abs PE)Model 3 (Price)
2.39% 6.78%
3.49%
0.66% 0.58% 1.22%
0.01% 0.63% 0.58% 1.22%
ΔPE
ΔPE (incremental effect when lag PE 40)
ΔAbs PE
0.0409 o 0.01nnn
0.0857 o 0.01nnn
0.63% 0.62% 1.23%
0.49% 0.09% 0.96%
0.51% 0.08% 0.96%
0.01% 0.50% 0.10% 0.97%
ΔIntraday volatility
ΔSpread
ΔOIB
0.0176 0.02nn 0.0223 o 0.01nnn 0.0171 0.02nn
0.0158 0.01nn 0.0137 0.02nn 0.0146 o 0.01nnn
0.0142 0.23 0.0150 0.25 0.0178 0.17
ΔIntraday volatility
ΔSpread
ΔOIB
0.0053 0.02nn 0.0059 0.09n 0.0025 0.48
0.0048 0.07n 0.0042 0.09n 0.0045 o 0.01nnn
0.0054 0.04nn 0.0029 0.46 0.0004 0.93
Panel E: Impact of FTDs, Threshold Securities
163 Securities
ΔPrice
0.0268 0.02nn
175 Securities
0.0024 0.24
175 Securities Panel F: Impact of FTDs, Nasdaq sample
2223 Securities 2983 Securities 2983 Securities
ΔPE
ΔPE (incremental effect when lag PE 40)
0.0069 0.10
0.0162 o 0.01nnn
ΔAbs PE
ΔPrice
0.0135 o 0.01nnn 0.0019 0.02nn
504
V. Fotak et al. / Journal of Financial Economics 114 (2014) 493–516
deviation, and then winsorizing at three standard deviations from the mean. Based on an analysis of the Schwarz information criteria, we determine that, for all models, it is most appropriate to use a VAR of order one. In order to draw inferences about the true population parameters, we average coefficient estimates obtained for each security as in Fama and MacBeth (1973). In significance testing, we use cross-sectional estimates of standard errors. For robustness, we also estimate all three models without standardizing variables, but we face problems with the convergence of our maximum likelihood estimation algorithm, likely driven by the extreme differences in magnitude across our variables. For those securities for which we manage to obtain results without standardizing variables, our results are qualitatively similar to those reported. We estimate VAR Models 1, 2, and 3 for both our “entire market” sample and for a subset of the sample containing only the 10% of securities with the highest level of FTDs based on mean OFR over the sample period, January 2005 to June 2008 (the same subsample described as “decile 10” in Table 2). We focus on this subset of securities with the highest FTDs to find whether the impact of FTD changes differs when they occur most frequently. We find that the sign and the significance of the various inter-relationships involved are economically reasonable for both samples. For compactness, ease of interpretation, and to preserve the focus on the specific equations of interest to us, we report in Table 5 only those relationships that are relevant to the issues addressed in the paper. In particular, we report, for all three models and for both samples, the coefficient estimates related to the effects of FTD changes and timely delivered short sales on each of the market-quality metrics. We present estimated coefficients and related statistical significance in Table 5, Panels A and B. First, and most importantly, irrespective of the measure used as proxy for pricing efficiency or liquidity, increases in FTDs lead to higher liquidity and improve pricing efficiency. In particular: 1. Increases in FTDs significantly reduce positive pricing errors, consistent with those traders acting as value arbitrageurs when stocks are overpriced (statistically significant at 1%). 2. Increases in FTDs significantly reduce absolute pricing errors, consistent with an increase in pricing efficiency (statistically significant at 1%). 3. Increases in FTDs significantly reduce intraday volatility, consistent with improved market stability (significant at 1% in two models, at 5% in two models, at 10% in one model, and not significant in the sixth). 4. Increases in FTDs significantly reduce order imbalances, consistent with an improvement in liquidity (significant at 1% for the overall sample, but not for the sample of securities with most fails). 5. Increases in FTDs reduce spreads and thus lower trading costs (significant at 1% in five models and at 5% in the sixth). 6. The impact of changes in FTDs on prices is not significant overall.
Second, we find that the direction and significance of all results relating to the impact of FTD changes is qualitatively
similar to the direction and significance of all results relating to the impact of delivered short sales. The main differences lie in the fact that delivered short sales appear to have a statistically significant impact on prices, while the results are not statistically significant for FTD changes. Our main analysis does not include any measure of trading volume, as trading volume is highly correlated with other market-quality metrics. Nevertheless, for robustness, we add to the VAR models a measure of trading volume based on the daily number of trades for the security of interest. The core results describing the relationship between FTD changes and metrics of market liquidity and pricing efficiency are unaffected, while we find that an increase in OFR is related to lower subsequent trading volume. We also obtain similar results by utilizing a different metric of trading volume based on the monetary value of shares traded daily. The results are not included in tables for brevity. 3.5. FTD changes vs. timely delivered short sales Overall, when comparing the impact of FTD changes with the impact of delivered short sales, the direction and significance of the results are virtually identical, with the exception of the impact on stock prices. That said, the coefficients measuring the impact on market quality in Table 5, Panels A and B are, in most cases, larger in magnitude for delivered short sales relative to FTD changes. This is arguably what we should expect given that the vector autoregressive models we employ are based on standardized variables, and the average standard deviation of delivered short sales is much greater than that of FTD changes, as reported in Table 2. Hence, a one standard deviation change in delivered short sales predictably results in a much greater volume of trades relative to a one-standard deviation change in FTDs and should arguably have a greater impact on market quality. In this context, we next examine the economic significance of the market-quality impact of an equal size (in terms of the number of shares outstanding) of FTD changes and delivered short sales. The results are reported in Table 5, Panels C and D for the impact of FTD changes and delivered short sales, respectively. Based on the Model 1 estimates for the “entire market” sample, we find that an increase in FTDs corresponding to 10 basis points of the total number of outstanding shares— approximately equal to a one-standard deviation increase in total short interest—leads to a 3.2% reduction in the magnitude of positive pricing errors; a 3.2% reduction in absolute pricing errors; a 0.2% point decline in intraday volatility; a 1.7% reduction in spreads; and a 0.7% reduction in order imbalance. The impact on prices, aside from not being statistically significant, is tiny, with a reduction of less than 0.01%. In comparison, again based on the “entire market” sample, we find that delivered short sales corresponding to 10 basis points of the total number of outstanding shares leads to about a 4.5% reduction in the magnitude of positive pricing error; a 6.8% reduction in absolute pricing error; a 0.6% decline in intraday volatility; a 0.6% reduction in spreads; a 1.2% reduction in order imbalance. The impact on prices is again small, with a reduction of approximately 0.01%. We test for the statistical significant of differences between the coefficients estimated for OFR (Panel C) and ODR (Panel D)
V. Fotak et al. / Journal of Financial Economics 114 (2014) 493–516
and find that those are, in all cases, not statistically significant.10 Overall, the economic significance of the estimated impact of FTD changes and delivered short sales is of roughly similar magnitude. Inferences based on the sample with the highest FTD changes are weaker (the same quantum of FTD changes leads to smaller market-wide reactions). This suggests that a further increase in FTDs has a weaker impact when existing levels of outstanding FTDs are higher. The bottom-line is that the market-quality impact of FTD changes is economically significant and very similar to that of timely delivered short sales. Both types of trades appear to have a positive effect on market quality, first by enhancing pricing efficiency through correction of security overvaluation and reduction of intraday volatility, and second by providing and improving liquidity through a reduction of order imbalances and spreads.
3.6. Impulse response functions To further investigate the impact of FTD changes on pricing efficiency and liquidity, we compute impulse response functions based on the vector autoregressive models of Table 5. We estimate accumulated impulse response function parameters for each security and then present cross-sectional means of the parameter estimates. We construct confidence intervals using cross-sectional estimates of standard errors. We present the results related to accumulated impulse response functions depicting the impact of a one-standard deviation increase in OFR on pricing error volatility, intraday volatility, spreads, and order imbalances in Fig. 1. We opt to utilize accumulated rather than orthogonalized impulse response functions because, consistent with econometric literature in this regard, the latter are extremely sensitive to the ordering of variables in the model, thus adding a level of arbitrariness to the estimation procedure. For comparison, we present a similar set of impulse response functions for the impact of delivered shorts in Fig. 2. As the impulse response functions depict, the impact of FTD changes on market-quality metrics documented in Table 5 occur over the following day. On subsequent days, accumulated responses are mostly flat, indicating no further correction or reversal. Given the very close similarities in the estimated VAR coefficients, we only report impulse response functions related to the VAR models with pricing error volatility, and omit the other models including pricing errors and prices. We observe very similar patterns for changes in ODR, thus providing additional evidence that delivered short sales have a similar impact on market quality.11
10 We do not separately report all the estimated t-statistics related to the tests for significance because all tests uniformly fail to reject the null hypothesis that the mean difference between estimated coefficients is zero. 11 Given the different magnitude of the estimated coefficients, which reflect the scaling issues described in Section 3.5, the scales used in the vertical axes in the plots in Figs. 1 and 2 differ. Such unavoidable differences in scaling render a visual comparison of the magnitude of the estimated responses and of the related confidence intervals difficult: while the confidence intervals associated with OFR appear wider than those associated with ODR, they are, in reality, narrower.
505
3.7. Robustness check: threshold-list securities The impact of FTD changes could be different when failures are persistent or long-lived. Accordingly, we separately focus on a sample of securities which displays persistent FTDs—securities that appear on the NYSE Threshold List. In accordance to the SEC Regulation SHO, “threshold securities” are securities for which the aggregate FTD position for five consecutive settlement days at a registered clearing house totals 10,000 shares or more and equals at least 0.5% of the total number of shares outstanding for the issuer. From our sample, we select all securities which appear on the NYSE Threshold List for at least one day during the period spanning January 1, 2005 to June 30, 2008: we identify 175 such securities. Accordingly, we re-estimate the VAR model for this subset of securities with the addition of a binary variable, equal to one for each day on which a security is included on the threshold list, as a predictor in each equation. Results, reported in Table 5, Panel E, are largely consistent with what was previously reported: coefficient estimates are of similar magnitude to those obtained with the data subset containing the securities with the highest level of FTDs. While we do not report the coefficients on the “threshold” binary variable in tables for brevity, we note that inclusion on the threshold list is associated with higher spreads, higher intraday volatility, and higher prices.12 Overall, these results are consistent with inclusion on the threshold list hampering liquidity (as, under regulation SHO, shorting without pre-borrowing of securities on the threshold list is curtailed), while leading to higher prices, perhaps due to the resulting buying pressure of traders forced to close their non-delivered positions.
3.8. Robustness check: Nasdaq securities As a further robustness test, we re-estimate the VAR model with a sample of securities with Nasdaq as a primary exchange. We apply the same filters in creating the sample as we did for our overall, NYSE-based sample: we focus on securities with share codes 10 and 11 in CRSP, for which we have at least 18 months of continuous trading data and with no significant changes in the number of shares outstanding; we obtain a full sample of 2,983 securities. Results are presented in Table 5, Panel F. The total number of securities for which we are able to estimate the models presented in Table 5, Panel F ranges from 2,983 to 2,223. Estimation is not possible for some securities because of inadequate number of positive pricing errors, or negative pricing errors, or order imbalances. In this data subset, similar to our NYSE results, we find that increases in FTDs lead to a next-day decline in positive pricing errors and absolute pricing errors, intraday volatility, and in spreads, but we also find an increase in prices. We also find a decline in order imbalances, but results are statistically significant in only one of three models. However, in this context, it is 12 This test is based on a time-series analysis and, thus, should be interpreted as indicative of changes to market quality affecting a specific security at threshold-list inclusion.
506
V. Fotak et al. / Journal of Financial Economics 114 (2014) 493–516
Fig. 1. Plots of accumulated impulse response functions depicting the impact of a one-standard deviation shock in the Outstanding fails-to-deliver ratio (OFR). The VAR model used for estimation is Model 2 and the sample is the “entire market,” as described in Table 5. The response variables in the four panels are, respectively, Absolute pricing error, Intraday volatility, Spread, and Order imbalance, as defined in Table 1. The horizontal axes are time periods (days), following the initial shock (day 0). All variables are standardized and winsorized by security. Impulse response coefficients are estimated by security; mean values are depicted. 5% confidence intervals are computed using cross-sectional standard error estimates. Panel A: Absolute PE, Panel B: Intraday volatility, Panel C: Spread and Panel D: Order imbalance.
relevant to recognize data limitations, particularly liquidity metrics, for Nasdaq securities.13 Most importantly in the context of this paper, while NYSE has played the dominant role in our sample period in the supply of liquidity for stocks with NYSE as their primary exchange, this has not been so for stocks with Nasdaq as their primary exchange. For example, Diether, Lee, and Werner (2009) document that, while NYSE accounted for almost 80% of shares sold short in NYSE stocks, Nasdaq accounted for less than 35% of the shares sold short in Nasdaq securities. Hence, our estimated order imbalances can potentially be a noisy proxy of the overall order imbalances of Nasdaq stocks.
13 The fragmentation of trading across several different platforms and systems on Nasdaq has led in the past to difficulty in observing and reacting to short-term changes in market variables. For example, quote adjustment has been documented to be slower (Jones and Lipson, 1999) and time stamps of trades and quotes on Nasdaq could not always be matched (Schultz, 2000). More recently, Pastor and Stambaugh (2003) and Watanabe and Watanabe (2008) do not use Nasdaq data in their pricing of liquidity studies apparently because of problems in using Nasdaq-based liquidity measures.
3.9. FTD changes: market-makers vs. public traders The SEC has regarded the use of FTDs by market-makers as bona fide and positive for market liquidity. SEC Report “Key Points about Regulation SHO,” April 11, 2005, updated 2008 says: “because it may take a market maker considerable time to purchase or arrange to borrow the security, a market maker engaged in bona fide market making, particularly in a fast-moving market, may need to sell the security short without having arranged to borrow shares.” In a similar vein and context, SEC Report No. 450, March 18, 2009, states that the SEC “has repeatedly stressed the fact that the practice can provide needed market liquidity in certain circumstances”. It is therefore conceivable that the beneficial effects on market quality of FTDs arise due to such market-making activities, and that there are no beneficial effects from FTDs of public traders. We define the term “public trader” in this context to mean all market participants other than registered market-makers. Notably, it is these public traders that have been the primary focus of negative media and regulatory attention. If there are no beneficial effects of public-trader FTDs, the estimated FTD-related improvement in market quality will be relatively stronger for securities that have a relatively greater proportion of market-maker FTDs. To gain insights into the impact of
V. Fotak et al. / Journal of Financial Economics 114 (2014) 493–516
507
Fig. 2. Plots of accumulated impulse response functions depicting the impact of a one-standard deviation shock in the Outstanding delivered shorts ratio (ODR). The VAR model used for estimation is Model 2 and the sample is the “entire market,” as described in Table 5. The response variables in the four panels are, respectively, Absolute Pricing error, Intraday volatility, Spread, and Order imbalance, as defined in Table 1. The horizontal axes are time periods (days), following the initial shock (day 0). All variables are standardized and winsorized by security. Impulse response coefficients are estimated by security; mean values are depicted. 5% confidence intervals are computed using cross-sectional standard error estimates. Panel A: Absolute PE, Panel B: Intraday volatility, Panel C: Spread and Panel D: Order imbalance.
public-trader FTDs relative to the impact of market-maker FTDs, we classify securities based on the proportion of FTDs originating from market-makers. In order to accomplish that, we employ an exogenous event—an SEC ban in September 2008 that affected only public-trader FTDs but not market-maker FTDs: SEC Temporary Rule 204T (later made permanent). SEC Rule 204T required all market participants other than registered market-makers to purchase or borrow securities to close out their FTD position by the beginning of day tþ4, effectively banning FTDs for anyone other than a market-maker. Hence, we compute average FTDs across our entire sample of NYSE securities for two periods: pre204T from January 1, 2005 to June 30, 2008; and post-204T from January 1, 2009 to December 31, 2010. We intentionally omit the period July to October 2008 because of multiple short selling rule changes, and omit November and December 2008 to allow outstanding FTD positions to fully clear and thereby enable accurately gauging the impact of post-204T FTDs. Consistent with a 2009 s report, we find a drop in average OFR of approximately 71% across all securities, suggesting that, on average, about 29% of all FTDs were due to market-makers prior to September 2008.
For each sample security, we use the proportional change in mean OFR between the two subperiods as our proxy for FTDs by public traders. Our underlying assumption here is that the number of market-maker FTDs is fairly constant over time.14 We accordingly rank securities on the basis of this proportional decline and allocate securities to terciles. Table 6 presents results for the two terciles corresponding to “high proportion of estimated marketmaker trades” (i.e., securities with the lowest decline in mean OFR) and “low proportion of estimated marketmaker trades” (i.e., securities with the highest decline in mean OFR). For the “low proportion of estimated market-maker trades” subsample, despite the sample size being much smaller than the overall sample, all previously documented results remain strong and statistically significant, with the exclusion of the coefficient estimates associated with order imbalance. For the “high proportion of estimated 14 A brief example can be useful in clarifying our methodology: if, for a given security, average OFR in the pre-204T period is 1% and OFR in the post-204T period is 0.25%, then we estimate that 75% of FTDs are due to public traders and 25% are due to market-makers.
508
V. Fotak et al. / Journal of Financial Economics 114 (2014) 493–516
Table 6 Impact of FTDs for securities with high/low proportion of FTDs by market-makers. This table provides results for the estimation of three vector autoregressive models of order one: VAR with Pricing error (Model 1), VAR with Pricing error volatility (Model 2), and VAR with Price (Model 3), as described in Table 5. All variables are standardized and winsorized by security, and are as defined in Table 1. The sample includes all NYSE common stocks of US-based firms (CRSP share codes 10 and 11), listed in the CRSP and TAQ databases for at least 18 months over the interval January 1, 2005 to June 30, 2008, and with no large changes ( 410%) in the number of shares outstanding. The resulting sample includes 1,492 securities. Mean OFR is computed for all securities included in the sample for the period preceding the introduction of SEC Rule 204T (January 1, 2005 to June 30, 2008) and for a period following (January 1, 2009 to December 31, 2010). Securities are ranked according to the proportional change in mean OFR. Securities are then allocated to three quantiles—those with the highest decline in mean OFR are allocated to the “Low proportion of estimated market-maker trades” sample, while those with the lowest decline in mean OFR are allocated to the “High proportion of estimated market marker trades” sample. Reported parameter estimates are averages of parameters estimated by security. Significance is tested employing a cross-sectional estimate of the standard error of the parameter estimate. The p-values from two-sided t-tests are in italics below the parameter estimates. “n”, “nn”, and “nnn” indicate significance at the 10%, 5%, and 1% level, respectively. Low/high proportion of estimated market-maker trades A: Low 340 Securities B: High 329 Securities A: Low 340 Securities B: High 340 Securities A: Low 340 Securities B: High 340 Securities
ΔPE
ΔPE (incremental effect when lag PE40)
0.0216 o 0.01nnn 0.0094 0.30
0.0287 o0.01nnn 0.0033 0.74
market-maker trades” subsample, we observe statistically significant parameter estimates for the impact on spreads, but not for the other variables. Our overall results indicate that, if anything, the impact of FTD changes is stronger for the subset of securities with a relatively low involvement of market-makers. Hence, at the very least, our results are inconsistent with the hypothesis that the beneficial effects of FTD changes arise entirely from market-makers. Our intuition is that, while market-maker trades contribute to an improvement in liquidity (and, hence, lead to lower spreads), they do not lead to an improvement in pricing efficiency—which is consistent with market-makers not engaging in price arbitrage. 3.10. The market-quality impact of restrictions on FTDs As discussed in the introduction, FTDs can plausibly originate from both short and long sales. In this subsection, we use a natural experiment based on an exogenous regulatory change to focus on FTDs most likely arising from short sales. In particular, we look at the imposition on July 15, 2008 by the SEC of an Emergency Order selectively mandating borrowing requirements prior to short sales and applicable only to the stocks of a select group of 19 publicly traded financial institutions from July 21 to August 12, 2008. The SEC order required that “no person may effect a short sale in these securities… unless such person or its agent has borrowed or arranged to borrow the security or otherwise has the security available to borrow in its inventory prior to effecting such short sale.” (SEC Release 58166, 2008). The only regulatory change affecting these stocks during this period was this order mandating firm
ΔAbs PE
ΔPrice
ΔIntraday volatility
ΔSpread
ΔOIB
0.0018 0.09n 0.0007 0.52
0.0085 0.03nn 0.0003 0.95 0.0114 o0.01nnn 0.0070 0.15 0.0090 0.02nn 0.0050 0.30
0.0152 o 0.01nnn 0.0105 o 0.01nnn 0.0142 o 0.01nnn 0.0119 o 0.01nnn 0.0147 o 0.01nnn 0.0127 o 0.01nnn
0.0082 0.15 0.0108 0.05nn 0.0079 0.16 0.0104 0.11 0.0074 0.21 0.0099 0.13
0.0169 o 0.01nnn 0.0036 0.47
stock-borrowing arrangements prior to executing short sales. Any FTDs arising from “long” sales should clearly not be affected by this order. The overall effect of the SEC order in relation to FTDs should therefore be driven by FTDs arising from short sales. We employ an event study methodology to investigate the variation in FTDs and in market quality around this mandate requiring stockborrowing arrangements prior to short sales. Data for 17 of the 19 securities (the “event sample”) affected by the order are available from CRSP. We construct a matched control sample: for each of our 17 affected securities, we identify the firm that is not affected by the order, but shares the same four-digit Standard Industrial Classification (SIC) code and has the closest market capitalization (least absolute deviation) as of January 1, 2008. Overall, the securities affected by the ban have a higher mean and median market capitalization, but the differences are not statistically significant. Then, we analyze the number of reported FTDs (duly scaled by the number of shares outstanding) and relevant market-quality variables for both event and control samples, over the following periods: (1) a Pre-order period from January 1 to July 14, 2008 ending on the day prior to the announcement of the Emergency Order; (2) the Order period—July 21 to August 12, 2008—for which the Emergency order was in force; and (3) a Post-order period from August 13 to September 9, 2008 following the order and ending before Lehman's bankruptcy. We do not include the four trading days on July 15, 16, 17, and 18, 2008, in the Pre-order period to avoid any potentially confounding effects arising from the announcement of the order on July 15, 2008. We report our results on FTDs in Table 7. Reported FTDs (as a proportion of shares outstanding) for the 17 securities
V. Fotak et al. / Journal of Financial Economics 114 (2014) 493–516
509
Table 7 FTDs and short sales. This table presents the average number of fails-to-deliver (FTDs) scaled by the contemporaneous number of shares outstanding for “Event sample” and “Control sample.” Securities in the “Event sample” are 17 stocks that were subject to the imposition of the SEC Order requiring stock-borrowing arrangements prior to short sales between July 21, 2008 and August 12, 2008, i.e., the “Order period.” While the order affected 19 securities, data are available for 17 of these: Bank of America Corporation, Barclays, Bear Stearns Companies Inc., Citigroup Inc., Credit Suisse Group, Deutsche Bank Group AG, Allianz SE, Goldman Sachs Group Inc., Royal Bank ADS, HSBC Holdings PLC ADS, J.P. Morgan Chase & Co., Merrill Lynch & Co., Inc., Mizuho Financial Group, Inc., Morgan Stanley, UBS AG, Freddie Mac, and Fannie Mae. “Control sample” securities are 17 market capitalization and industry-matched stocks that were not subject to any relevant regulatory changes during the “Order period.” The “Pre-order period” refers to the interval January 1, 2008 to July 14, 2008. The “Post-order period” refers to the interval August 13, 2008 to September 9, 2008. The number of FTDs scaled by shares outstanding is reported in basis points and computed on a daily basis over the interval January 1, 2008 to August 12, 2008. Reported p-values are for a test of differences in mean FTDs (scaled by shares outstanding) between the indicated order subperiods and the “Pre-order period” and between “Event sample” and “Control sample”. “n”, “nn”, and “nnn” indicate significance at the 10%, 5%, and 1% level, respectively. Event companies Period
Mean (bps)
Pre-order Order, 1st week (Order, 1st week) – Pre-order Order, 2nd week (Order, 2nd week) – Pre-order Order, 3rd week (Order, 3rd week) – Pre-order Order, Last day (Order, Last day) – Pre-order Post-order (Post-order) – (Order, 3rd week)
4.78 0.57 4.22 0.17 4.61 0.12 4.67 0.00 4.78 1.12 1.01
Control companies
p-Value
o 0.01nnn o 0.01nnn o 0.01nnn nnn
o 0.01
0.03nn
in our sample average 4.78 basis points (“bp”) during the pre-order period. After the order, FTDs reduce rapidly to about 0.57 bp, or 12% of their pre-order average within one week; to about 0.12 bp, or 2% of the pre-order average, within three weeks; and, on the last day of the order period, reported FTDs are below the reporting threshold of 10,000 shares for each and every affected security. In contrast, reported FTDs for control sample securities increase during the first week of the ban, slightly decrease over the second week of the ban, and show no significant change during the third week of the ban. As soon as the temporary order period expires, reported FTDs of affected securities again increase monotonically and significantly. The difference in reported FTDs between event and control firms is positive and significant prior to the mandate period, negative and significant during the mandate period, and again positive and significant after the mandate is lifted. Given that reported FTDs cannot fall immediately to zero because of the time required for fails outstanding from prior to the SEC order to clear (Boni, 2006), this natural experiment suggests that, at least for these particular securities and around this specific period, a very large portion of FTDs originated from short sales.15 We report our results on market quality in Table 8. To test the impact of the order on market quality, we analyze Absolute pricing error, Return, Intraday volatility, Spread, and Order imbalance over each of our three subperiods. We average each of these five variables across securities to obtain a daily mean for the affected and control samples. We compare differences in means 15 For our larger NYSE sample, we also find that new FTDs on day tþ 3 are significantly (p-value⪡0.01) and positively related only to the daily trading volume arising from short sales on day t, and not the daily trading volume arising from “long” sales on day t.
Mean (bps) 0.47 2.79 2.32 0.46 0.01 0.48 0.01 0.23 0.24 0.18 0.31
p-Value
o0.01nnn 0.72 0.21 o0.01nnn 0.51
Event-Control Mean (bps)
p-Value
4.31 2.23 6.54 0.29 4.60 0.36 4.68 0.23 4.55 0.95 1.31
o 0.01nnn o 0.01nnn o 0.01nnn 0.53 o 0.01nnn 0.43 o 0.01nnn 0.61 o 0.01nnn 0.04 o 0.01
between the order-period (July 21 to August 12, 2008) and the “pre- and post-order” (January 1 to July 14 and August 13 to September 9) periods for both sample and control securities. Our results are presented in Table 8. Comparing securities during the benchmark “pre- and post-order” period, we note that “event sample” securities display larger pricing errors, lower returns, higher intraday volatility, and lower spreads than “control sample” securities. While these results are indicative of the difficulty of obtaining perfect matches, we mitigate the impact of these differences by employing a difference-in-difference methodology: we compute the difference between the “order” and the “pre- and post-order” periods for both “event sample” and “control sample” securities and then compute the difference between those differences. We find that, while both samples display higher absolute pricing errors, intraday volatility, and spreads during the “order” period, the increase for the “event sample” securities is larger and the difference-in-difference tests are all statistically significant—at the 1% level for absolute pricing errors and for intraday volatility and at the 10% level for spreads. The impact of the SEC order on returns and order imbalances is not statistically significant. In robustness tests (that are not reported for brevity), we replicate a similar analysis excluding the period subsequent to the order from the benchmark, and obtain virtually identical results. Boulton and Braga-Alves (2010) also document a negative impact of the July 2008 Emergency Order on liquidity, i.e., higher bid-ask spreads and lower trading volume relative to a matched sample of financial firms. However, an SEC Office of Economic Analysis (OEA) memorandum dated January 14, 2009 titled “Analysis of the July Emergency Order Requiring a Pre-Borrow on Short Sales” examines selected market-quality variables around the July 2008 Emergency Order, but is unable to conclude statistically
510
V. Fotak et al. / Journal of Financial Economics 114 (2014) 493–516
Table 8 The impact of restrictions on FTDs imposed by the SEC between July 21, 2008 and August 12, 2008. The table presents means of Absolute pricing error, Return, Intraday volatility, Spread, and Order imbalance, as defined in Table 1. The SEC mandated preborrowing prior to short selling 19 securities during the period spanning July 21 to August 12, 2008 (the “Order” period). The Order affected 19 securities, but data are available for 17 of these (the “Event sample”): Bank of America Corporation, Barclays, Bear Stearns Companies Inc., Citigroup Inc., Credit Suisse Group, Deutsche Bank Group AG, Allianz SE, Goldman Sachs Group Inc., Royal Bank ADS, HSBC Holdings PLC ADS, J. P. Morgan Chase & Co., Merrill Lynch & Co., Inc., Mizuho Financial Group, Inc., Morgan Stanley, UBS AG, Freddie Mac, and Fannie Mae. Securities in the “Control sample” are market capitalization and industry-matched stocks that were not subject to any relevant regulatory changes during the “Order” period. Means during the “Order” period are compared to variable means computed over a period surrounding the “Order,” the “Pre- and Post-order” periods, as indicated in the table. The p-values from two-sided t-tests for the significance of mean differences are in italics below the parameter estimates. “n”, “nn”, and “nnn” indicate significance at the 10%, 5%, and 1% level, respectively.
Abs PE
Pre- and post-order (Jan 1–Jul 14 and Aug 13–Sep 9, 2008)
Order (Jul 21–Aug 12, 2008)
Difference (Order – Pre- and postorder)
20.47%
28.08%
1.34%
1.48%
7.61% o0.01nnn 0.14%
19.13% o 0.01nnn
26.60% o0.01nnn
0.24 7.47% o0.01nnn
0.37%
0.27%
Control sample
0.14%
0.46%
0.10% 0.93 0.59%
Difference (EventControl) Event sample
0.24% 0.06n
0.73% 0.07n
0.49 0.49% 0.26
0.33%
0.44%
0.11%
0.29%
0.37%
o0.01nnn 0.07%
0.03% o 0.01nnn
0.08% o0.01nnn
o0.01nnn 0.04% o0.01nnn
0.10%
0.15%
0.17%
0.21%
0.05% o0.01nnn 0.04%
0.07% o 0.01nnn
0.06% o0.01nnn
o0.01nnn 0.01% 0.07n
1.78%
0.16%
Control sample
0.18%
0.34%
1.94% 0.33 0.53%
Difference (EventControl)
1.59% o 0.01
0.18% 0.90
0.71 1.41% 0.37
Event sample Control sample
Return
Intraday volatility
Difference (EventControl) Event sample
Control sample
Spread
Difference (EventControl) Event sample Control sample
OIB
Difference (EventControl) Event sample
significant differences between their “event” and “control” samples for any of the market-quality variables they examine. There can be several reasons for the difference in our conclusions. First, except for returns (for which our conclusions are the same), the OEA do not examine the variables that we analyze all through this paper, or their variable is defined in a materially different way: (1) they do not examine absolute pricing errors and order imbalances; (2) as against our intraday volatility measure (computed daily over five-minute intervals), they examine daily closeto-close or open-to-close volatility (over several weeks); and (3) as against our relative spreads (i.e., dollar spreads
scaled by price), they examine dollar spread (which also means that their averages implicitly give higher weight to higher priced securities). Second, while we use a control sample of financial firms that are as closely matched to the affected securities as possible, the OEA uses two control samples—a sample of all financial firms and a sample of all nonfinancial firms—none of which are specifically matched with the affected securities. Third, the OEA uses a benchmark period prior to the Order that is constructed by randomly selecting only 20 trading days over the previous nine months. The small number of days and the randomness may arguably increase the standard errors of their
V. Fotak et al. / Journal of Financial Economics 114 (2014) 493–516
estimates, besides making it impossible for us to replicate their results or more fully explain any differences. The longer benchmark period we utilize, our closely matched construction of the control sample, and the higher frequency of some of our variables, are all reasons why our tests are likely to have greater power, and hence potentially accounts for the fact that we detect significance in tests in which the OEA may not. Overall, our results indicate that mandating stockborrowing arrangements prior to executing a short sale leads to a sharp drop in overall FTDs to almost zero (as it would if FTDs arose largely from short sales over the sample period), but at the same time, there is also a significant adverse impact on the stock's liquidity and its price discovery process. Banning FTDs should also arguably lead to an upward pressure on stock-lending rates as short sellers scramble to make stock-borrowing arrangements at short notice, and thereby also concomitantly increase costs for all short sellers. Kolasinski, Reed, and Thornock (2013) and the OEA memo cited above document that this is indeed the case with rebate rates for order-affected securities becoming negative on the first day of order-validity and remaining negative for several days thereafter. While our results clearly show that restricting FTDs reduces market quality, this could result directly from the absence of FTDs, or because informed short selling decreases due to consequential increase in stock-lending rates, or both. In
511
this context, we examine short interest of our event and control sample securities over our Pre-order period and Order-period, and find, consistent with similar findings of the OEA memo cited above, that the short interest of event sample securities over the Order-period is not significantly different from the short interest over the Pre-order period: the mean short interest for event firms reported on July 15 is 5.07% (of shares outstanding) and, by the end of the order period, on August 15, it is 5.45%; the slight increase is not statistically significant and neither is it if we measure it relative to the control sample. Given that FTDs arising from short sales decrease sharply during the ban while the short interest does not change significantly, we conjecture that the adverse market-quality impact is likely to be related at least partially to the absence of failing-to-deliver short sales. Irrespective, our results show that restricting FTDs reduces market quality. That said, our results in this subsection should be interpreted or generalized with caution. First, this Emergency Order affected only a few securities, all from one industry, and during a fairly unique period of time. Second, while Kolasinski, Reed, and Thornock (2013) note that it is sensible to analyze the impact of a short selling regulation during a crisis, as that is exactly when short selling regulation is most likely to be operational, the OEA memo cited above suggested a “Hawthorne effect,” whereby “firms could have been on their best behavior during the
Fig. 3. Plot of Outstanding fails-to-deliver ratio and Return index related to Bear Stearns Companies Inc., common stock (ticker: BSC) against calendar date. The Return index is set to 1 on January 1, 2008; Return indext ¼ Σ ij ¼ 1 ð1þ RBSC;j Þ RBSC;j is the observed total return for BSC on day j, from the CRSP database.
512
V. Fotak et al. / Journal of Financial Economics 114 (2014) 493–516
Table 9 Outstanding fails-to-deliver ratio (OFR) of Bear Stearns (ticker: BSC) and of Lehman Brothers Holdings Inc. (ticker: LEH) in 2008. The Outstanding fails-to-deliver ratio (OFR) is computed as the ratio of outstanding fails-to-deliver and shares outstanding. Index OFR is calculated as the equal-weighted average of OFR of common stocks of four firms that are matched on primary SIC and market capitalization as of the end of the fiscal year 2007 to BSC in Panel A and to LEH in Panel B. Tests for significance and related t-statistics are computed using the standard error of the OFR difference over the time interval January 1, 2008 to February 15, 2008 for BSC and over the time interval January 1, 2008 and ending 20 trading days prior to September 9, 2008 for LEH; p-values are two-sided. “n”, “nn”, and “nnn” indicate significance at the 10%, 5%, and 1% level, respectively. Panel A: Bear Stearns Date 03/03/08 03/04/08 03/05/08 03/06/08 03/07/08 03/10/08 03/11/08 03/12/08 03/13/08 03/14/08 03/17/08 03/18/08 03/19/08 03/20/08 03/24/08 03/25/08 03/26/08 03/27/08 03/28/08 03/31/08
Panel B: Lehman Brothers Holdings Inc.
BSC stock price (USD)
BSC OFR
Index OFR
Difference in OFR
p-Value
79.86 77.32 77.17 75.80 69.90 70.08 62.30 62.97 61.58 57.00 30.85 2.00 5.91 5.26 6.39 11.25 10.94 11.21 11.23 10.78
0.0580% 0.0069% 0.0154% 0.0634% 0.0574% 0.0082% 0.0854% 0.5284% 0.3175% 0.8979% 5.8383% 5.3300% 4.9694% 3.2489% 3.9550% 5.2488% 1.9806% 2.3933% 2.7779% 1.1950%
0.0003% 0.0000% 0.0062% 0.0008% 0.0017% 0.0001% 0.0010% 0.0249% 0.0041% 0.0155% 0.0174% 0.0004% 0.0031% 0.0180% 0.0030% 0.0240% 0.0076% 0.0110% 0.0000% 0.0000%
0.0577% 0.0069% 0.0092% 0.0626% 0.0558% 0.0081% 0.0845% 0.5035% 0.3134% 0.8824% 5.8209% 5.3296% 4.9663% 3.2310% 3.9520% 5.2248% 1.9730% 2.3823% 2.7779% 1.1950%
0.52 o 0.01nnn o 0.01nnn 0.66 0.47 o 0.01nnn 0.64 o 0.01nnn o 0.01nnn o 0.01nnn o 0.01nnn o 0.01nnn o 0.01nnn o 0.01nnn o 0.01nnn o 0.01nnn o 0.01nnn o 0.01nnn o 0.01nnn o 0.01nnn
time of the Emergency Order because of the likelihood of intense scrutiny by regulators of failed positions”. 4. Empirical results: FTDs and the 2008 financial crisis In the wake of the 2008 financial crisis, it has been widely alleged that short sellers have, by failing to deliver, caused or accelerated sharp declines in stock prices, particularly of financial firms, in order to profit from eventual collapse of the financial institutions involved. For example, in relation to Bear Stearns, former Under Secretary of Commerce Robert J. Shapiro has been quoted as saying that “… the pace of the collapse of the stock price was clearly accelerated by the enormous naked short sale activity” (Euromoney, December 2008). Similarly, Richard Fuld, former CEO of Lehman Brothers, in his October 2008 testimony before the US House of Representatives Committee on Oversight and Reform, stated that “naked short selling […] dealt a critical, if not fatal, blow to Bear Stearns” and “contributed significantly” to the collapse of Lehman Brothers. In this context, we examine whether significant increases in FTDs preceded (and hence potentially triggered) the price crashes associated with the four major casualties of the 2008 financial crisis, i.e., Bear Stearns, Lehman, AIG, and Merrill, or did significant increases in FTDs take place after these price crashes and in response to negative public news. The first large financial casualty of the 2008 financial crisis was Bear Stearns, the fifth largest investment banking firm in the nation at the time of its demise. We analyze OFR on the days preceding and immediately following the
Date 08/25/08 08/26/08 08/27/08 08/28/08 08/29/08 09/02/08 09/03/08 09/04/08 09/05/08 09/08/08 09/09/08 09/10/08 09/11/08 09/12/08 09/15/08 09/16/08 09/17/08
LEH stock price (USD)
LEH OFR
Index OFR
Difference in OFR
p-Value
13.45 14.03 14.78 15.87 16.09 16.13 16.94 15.17 16.20 14.15 7.79 7.25 4.22 3.65 0.21 0.30 0.13
0.0010% 0.0017% 0.0000% 0.0070% 0.0008% 0.0045% 0.0038% 0.0009% 0.0000% 0.0023% 0.0786% 0.4232% 1.6291% 2.3673% 2.0212% 2.1495% 3.6041%
0.0001% 0.0000% 0.0016% 0.0027% 0.0016% 0.0015% 0.0000% 0.0038% 0.0095% 0.0005% 0.0055% 0.0084% 0.0161% 0.0167% 0.0202% 0.0777% 0.0797%
0.0009% 0.0017% 0.0016% 0.0043% 0.0008% 0.0030% 0.0038% 0.0029% 0.0095% 0.0018% 0.0731% 0.4148% 1.6130% 2.3506% 2.0010% 2.0718% 3.5244%
0.89 0.79 0.79 0.50 0.90 0.63 0.55 0.64 0.13 0.78 o 0.01nnn o 0.01nnn o 0.01nnn o 0.01nnn o 0.01nnn o 0.01nnn o 0.01nnn
dramatic loss of market value that led to the demise of the firm. Fig. 3 provides a time-line about the crisis. Outstanding suspicions about liquidity problems at Bear Stearns were reported in the media from March 10 onwards, along with news that the company's management was repeatedly denying rumors about such problems. The first major price-crash took place on Friday, March 14, when the price per share dropped from $57 to $30 after a 9 a.m. announcement that Bear Stearns would receive an unprecedented loan from the Federal Reserve System; and two days later, on Sunday, March 16, JP Morgan Chase proposed buying Bear Stearns for $2 per share (Wall Street Journal, March 15, 2008; J.P. Morgan News Release, March 16, 2008). When markets opened on March 17, a second major price-crash materialized, and the price dropped to a close of $4.81. We compute OFR for Bear Stearns on each trading day from January 1 to March 28, 2008. We also compute, as a control variable, an equal-weighted average OFR for four other financial institutions—Raymond James Financial Corporation, Ameritrade Holdings Corporation, Ameriprise Financial Inc., and Charles Schwab Corporation —with the same primary SIC code as Bear Stearns, and with the closest market value as of the end of the fiscal year 2007. We test for the statistical significance each day of the difference, i.e., the “abnormal” OFR of Bear Stearns. To construct a t-statistic for the difference in means, we compute the standard error of this difference over the period January 1 to February 15, 2008. Results are presented in Table 9, Panel A. Even though negative media attention started on March 10, Table 9, Panel A and Fig. 3 show that, while OFR for Bear
V. Fotak et al. / Journal of Financial Economics 114 (2014) 493–516
513
Fig. 4. (A) Plot of Outstanding fails-to-deliver ratio and Return index related to Lehman Brothers Holdings Inc. (ticker: LEH) against calendar date. The Return index is set to 1 on January 1, 2008; Return indext ¼ Σ ij ¼ 1 ð1 þ RLEH;j Þ. RLEH,j is the observed total return for LEH on day j, from the CRSP database. (B) Detail of plot of Outstanding fails-to-deliver ratio and Return index related to Lehman Brothers Holdings Inc. (ticker: LEH) against calendar date. The Return index is set to 1 on January 1, 2008; Return indext ¼ Σ ij ¼ 1 ð1þ RLEH;j Þ . RLEH,j is the observed total return for LEH on day j, from the CRSP database.
Stearns was greater than for the benchmark securities, its small magnitude (below 0.1%) indicates that fails were unlikely to have a significant economic impact up to March 11. While OFR did increase significantly on March 12 and 13, the increase was still relatively tiny from an economic perspective relative to the total number of shares outstanding, to the typical overall short volume, and to what took place on or after March 14: OFR was only 0.32% of shares outstanding until market close on March 13. OFR increased to 0.90% on March 14, but increased massively
only on March 17, reaching 5.84%. Importantly, given that the Fed announcement was at the start of trading on March 14, even the (relatively small) increase in OFR on March 14 was clearly subsequent to the public release of tangibly negative news in the form of the announcement and the consequent immediate precipitous price-drop. By the time OFR spiked on March 17, the company was already in an open distress sale. The evidence clearly shows that the abnormal incidence of increases in FTDs did not precede the price decline but followed it; and the decline in stock
514
V. Fotak et al. / Journal of Financial Economics 114 (2014) 493–516
price was triggered by other well-identified negative economic news. Consistent with our previous results, short sellers appear to be facilitating price discovery, rather than increasing pricing errors. The second notable casualty of the financial crisis of 2008 was Lehman. In Fig. 4A we present the relationship between OFR and stock price for the period from January 2008 to Lehman's bankruptcy on September 15, 2008. We present a closer look at the period surrounding Lehman's bankruptcy in Fig. 4B. It is clear that OFR remained minuscule in magnitude all through till September 10; and till September 9, the firm's stock price had already lost approximately 87% of its value as of the beginning of the year. There were some spikes in OFR over the period, but these were too small to be economically relevant. For example, OFR increases on June 27 just prior to an 11% drop in share price on the next trading day of June 30, but this spike is just 0.06%, and ostensibly linked, according to an article in Bloomberg (March 19, 2009), to a rumor about Barclays PLC buying Lehman for a 25% discount to market value. The biggest single-day price drop of about 45% occurred on September 9, again following news that talks with the Korea Development Bank (previously rumored to be considering a 25% stake in Lehman) had failed. The OFR on the days immediately surrounding the dramatic price fall on September 9, 2008, is presented in Table 9, Panel B. The table indicates very low OFR, below 0.01%, in the days leading to September 9. While OFR increases on September 9, it still remains less than 0.09% of shares outstanding. Abnormal OFR increased more dramatically only after September 10, well after widespread coverage of negative news about Lehman and the associated price crash. On September 11, as shareholders rejected a management rescue plan, the stock price fell by an additional 42% and OFR further increased to 1.63%. Over the following days, talks of a possible acquisition by Bank of America and Barclays failed, triggering further declines in stock price and an increase in OFR to 2.37%. Lehman announced its
bankruptcy on September 15, and OFR increased beyond 3.6% on September 17. In sum, our analysis shows that the incidence of increases in FTDs occurred only after the large price drops on and after September 9, not before. We report the relationship between OFR and the stock price for Merrill and AIG in Figs. 5 and 6, respectively. In both cases, the stock price declines were fairly gradual through the year and not accompanied by any significant increase in FTDs. For Merrill, OFR reached its highest value of only 0.09% on October 10, 2008, well after Merrill had lost most of its value. AIG had also lost about 40% of its market value by the end of August 2008, and the largest single-day price drop was on September 15, 2008 when Standard & Poor's (S&P) cut AIG's credit rating. Following the announcement, the company's stock price dropped by about 60% and, on the following day, OFR climbed to its highest value up to that point, at 0.13%; OFR reached its highest value only a fortnight later on September 29, 2008, and even this highest value was only 0.14%. Given that OFR remained so low all through the financial crisis period for both firms, we do not engage in any further statistical testing. 5. Conclusions Recent regulatory action has focused on eliminating settlement day delivery failures in equity markets, particularly delivery failures arising from short sales, even though the alternative to fail has been shown to be valuable and important for pricing and liquidity (Evans, Geczy, Musto, and Reed, 2009; Merrick, Naik, and Yadav, 2005); and notwithstanding the extensive evidence on the beneficial effect of short sales on market quality (Diamond and Verrecchia, 1987; Abreu and Brunnermeier, 2002, 2003; Miller, 1977; Bris, Goetzmann, and Zhu, 2007; Diether, Lee, and Werner, 2009; Boehmer, Jones, and Zhang, 2008). In this context, we investigate the aggregate impact on pricing efficiency and liquidity of changes in FTDs, i. e., the market-quality impact of the set of trades over the days that collectively affect the open interest of undelivered positions.
Fig. 5. Plot of Outstanding fails-to-deliver ratio and Return index related to Merrill Lynch & Co., Inc. (ticker: MER) against calendar date. The Return index is set to 1 on January 1, 2008; Return indext ¼ Σ ij ¼ 1 ð1þ RMER;j Þ. RMER,j is the observed total return for MER on day j, from the CRSP database.
V. Fotak et al. / Journal of Financial Economics 114 (2014) 493–516
515
Fig. 6. Plot of Outstanding fails-to-deliver ratio and Return index related to American International Group (ticker: AIG) against calendar date. The Return index is set to 1 on January 1, 2008; Return indext ¼ Σ ij ¼ 1 ð1 þ RAIG;j Þ . RAIJ,j is the observed total return for AIG on day j, from the CRSP database.
For samples of both NYSE and (in robustness checks) Nasdaq securities, we find that increases in FTDs lead to a reduction in pricing errors, intraday volatility, spreads, and order imbalances. Further, consistent with our expectation that the effect on market quality of a short sale will not depend on whether or not the short sale finally results in timely delivery at settlement, we also find that the positive impact of FTD changes on pricing efficiency and liquidity is similar to the beneficial impact of our estimate of delivered short sales. We then focus specifically on FTDs most likely arising from short sales by examining the effect of a July 2008 temporary SEC order restricting FTDs by mandating stock-borrowing arrangements prior to short sales in select financial securities. We find that, while the SEC order banning FTDs arising from (naked) short sales did lead (as expected) to a drastic reduction in FTDs, it also led to a significant increase in absolute pricing errors, relative bid-ask spreads, and intraday volatility. These findings are consistent with other studies documenting a negative impact on market quality associated with regulatory restrictions on short sales in general, for example, Battalio and Schultz (2011), Boehmer, Jones, and Zhang (2013), Autore, Billingsley, and Kovacs (2011), and Beber and Pagano (2013). In the context of the widespread allegations that FTDs arising from (naked) short sales contributed to the collapse of major financial firms, we find that large abnormal increases in FTDs took place after and not before their major stock price declines and associated negative news; and hence, FTDs were responding to information about the firms, rather than triggering the observed price declines. Overall, our results indicate that, contrary to media and regulatory perceptions about naked short selling, FTDs did not precipitate the collapse of major financial firms in 2008. Our findings have important implications for regulatory policy. While we do not advocate tolerance for FTDs tied to
manipulative episodes, we find no evidence indicating that FTDs, in aggregate, systematically and manipulatively precipitated price declines, even in the extreme situation of the 2008 financial crisis. Instead, the gently regulated FTD regime that existed after Regulation SHO up to mid-2008 was net beneficial for pricing efficiency and market liquidity. In this context, the removal of the ability to fail for the vast majority of market participants is very questionable, since the ability to fail arguably reduces stock-borrowing costs at the time when such costs are the highest, and thereby protects traders from the lack of liquidity sometimes seen in the less regulated stockborrowing market. This ability to fail is valuable for markets also because it can prevent distortions in the stock-borrowing and lending market leading to serious pricing and liquidity distortions in the mainstream stock market. Regulators could alternatively consider progressive fines for settlement delays rather than a blanket removal of the alternative to fail. Importantly, regulation should focus on how best to generate liquidity and transparency in the stock-borrowing market.
References Abreu, D., Brunnermeier, M., 2002. Synchronization risk and delayed arbitrage. Journal of Financial Economics 66, 341–360. Abreu, D., Brunnermeier, M., 2003. Bubbles and crashes. Econometrica 71, 173–204. Autore, D., Billingsley, R., Kovacs, T., 2011. The 2008 short sale ban: liquidity, dispersion of opinion, and the cross-section of returns of US financial stocks. Journal of Banking & Finance 35, 2252–2266. Battalio, R., Schultz, P., 2011. Regulatory uncertainty and market liquidity: the 2008 short sale ban's impact on equity option markets. Journal of Finance 66, 2013–2053. Beber, A., Pagano, M., 2013. Short-selling bans around the world: evidence from the 2007–09 crisis. Journal of Finance 68, 343–381. Boehmer, E., Jones, C., Zhang, X., 2008. Which shorts are informed? Journal of Finance 63, 491–526.
516
V. Fotak et al. / Journal of Financial Economics 114 (2014) 493–516
Boehmer, E., Jones, C., Zhang, X., 2013. Shackling short sellers: the 2008 shorting ban. Review of Financial Studies 26, 1363–1400. Boni, L., 2006. Strategic delivery failures in U.S. equity markets. Journal of Financial Markets 9, 1–26. Boulton, T., Braga-Alves, M., 2010. The skinny on the 2008 naked short sale restrictions. Journal of Financial Markets 13, 397–421. Bris, A., Goetzmann, W., Zhu, N., 2007. Efficiency and the bear: short sales and markets around the world. Journal of Finance 62, 1029–1079. Chen, H., Singal, V., 2003. Role of speculative short sales in price formation: the case of the weekend effect. Journal of Finance 58, 685–705. Culp, C., Heaton, J., 2008. The economics of naked shorting. Regulation 31, 46–51. DeChow, P., Hutton, A., Meulbroek, L., Sloan, R., 2001. Short sellers, fundamental analysis, and stock returns. Journal of Financial Economics 61, 77–106. Dempster, P., Laird, N., Rubin, D., 1977. Maximum likelihood from incomplete data via the EM algorithm. Journal of the Royal Statistical Society B 39, 1–38. Desai, H., Thiagarajan, S.R., Ramesh, K., Balachandran, B., 2002. An investigation of the informational role of short interest in the Nasdaq market. Journal of Finance 57, 2263–2287. Diamond, D., Verrecchia, R., 1987. Constraints on short selling and asset price adjustment to private information. Journal of Financial Economics 18, 277–311. Diether, K., Lee, K., Werner, I., 2009. Short sale strategies and return predictability. Review of Financial Studies 22, 575–607. Edwards, A., Hanley, K., 2010. Short selling in initial public offerings. Journal of Financial Economics 98, 21–39. Evans, R., Geczy, C., Musto, D., Reed, A., 2009. Failure is an option: impediments to short selling and options prices. The Review of Financial Studies 22, 1955–1980. Fama, E., MacBeth, J., 1973. Risk, return, and equilibrium: empirical tests. Journal of Political Economy 81, 607–633. Geczy, C., Musto, D., Reed, A., 2002. Stocks are special too: an analysis of the equity lending market. Journal of Financial Economics 66, 241–269.
Hamilton, J., 1985. Uncovering financial market expectations of inflation. Journal of Political Economy 93, 1224–1241. Hasbrouck, J., 1993. Assessing the quality of a security market: a new approach to transaction-cost measurement. Review of Financial Studies 6, 191–212. Jones, C., Lipson, M., 1999. Price impacts and quote adjustment on the Nasdaq and NYSE/Amex. Unpublished working paper. Columbia Business School and University of Virginia. Kolasinski, A., Reed, A., Ringgenberg, M., 2013. A multiple lender approach to understanding supply and search in the equity lending market. Journal of Finance 68, 559–595. Kolasinski, A., Reed, A., Thornock, J., 2013. Can short restrictions result in more informed short selling? Evidence from the 2008 regulations. Financial Management 42, 155–181. Merrick, J., Naik, N., Yadav, P., 2005. Strategic trading behavior and price distortion in a manipulated market: anatomy of a squeeze. Journal of Financial Economics 77, 171–218. Miller, E., 1977. Risk, uncertainly, and divergence of opinion. Journal of Finance 32, 1151–1168. Morris, V., Goldstein, S., 2009. Guide to Clearance & Settlement. Lightbulb Press, New York. Pastor, L., Stambaugh, R., 2003. Liquidity risk and expected stock returns. Journal of Political Economy 111, 642–685. Putniņš, T., 2010. Naked short sales and fails-to-deliver: an overview of clearing and settlement procedures for stock trades in the USA. Journal of Securities Operations and Custody 2, 340–350. Schultz, P., 2000. Regulatory and legal pressures and the costs of Nasdaq trading. The Review of Financial Studies 13, 917–957. Stokes, A., 2009. In pursuit of the naked short. NYU. Journal of Law and Business 5, 1–61. Watanabe, A., Watanabe, M., 2008. Time-varying liquidity risk and the cross section of stock returns. Review of Financial Studies 21, 2449–2486. Welborn, J., 2008. The ‘phantom shares’ menace. Regulation 31, 51–61.
Contents lists available at ScienceDirect
Journal of Financial Economics journal homepage: www.elsevier.com/locate/jfec
Fails-to-deliver, short selling, and market quality$ Veljko Fotak a, Vikas Raman b, Pradeep K. Yadav c,n a
School of Management, University at Buffalo, The State University of New York, USA Warwick Business School, UK c Price College of Business, Division of Finance, University of Oklahoma, 307W. Brooks, Room 205A, Norman, Oklahoma 73019, USA b
a r t i c l e i n f o
abstract
Article history: Received 5 September 2011 Received in revised form 16 December 2013 Accepted 14 January 2014 Available online 17 August 2014
We investigate the aggregate market quality impact of equity shares that fail to deliver (hereafter “FTDs”). For a sample of 1,492 NYSE stocks over a 42-month period from 2005 to 2008, greater FTDs lead to higher liquidity and pricing efficiency, and their impact is similar to our estimate of delivered short sales. Furthermore, during the operative period of a Security and Exchange Commission (SEC) order mandating stock borrowing prior to short sales, the securities affected display relatively lower liquidity and higher pricing errors. Finally, we do not find any evidence that FTDs caused price distortions or the failure of financial firms during the 2008 financial crisis. & 2014 Elsevier B.V. All rights reserved.
JEL classification: G14 G18 Keywords: Naked short selling Short selling Failure to deliver
1. Introduction Trades in US stock markets are settled on a three-day cycle: for trades on day t, if the net delivery obligations of a clearing member are not fulfilled on day tþ3, any undelivered position becomes a “failure-to-deliver” (or FTD).
☆ The authors thank Leslie Boni, Tarun Chordia, John Griffin, Stewart Mayhew, Bill Megginson, David Musto, Narayan Naik, Adam Reed, Paul Schultz, Chester Spatt, “Vish” Viswanathan, Andriy Shkilko, and participants at the Western Finance Association Meetings, INQUIRE UK conference, Notre Dame Conference on Market Regulation, American Finance Association Meetings and seminars at Case-Western, Indian School of Business, the Fondazione Eni Enrico Mattei and the University of Oklahoma, for helpful comments and discussions. The authors gratefully acknowledge the support of the Center for Financial Research at the University of Koln, Germany, the Fondazione Eni Enrico Mattei, and the Institute for Quantitative Investment Research (INQUIRE), UK. The authors remain responsible for all errors. n Corresponding author at: Price College of Business, Division of Finance, University of Oklahoma, 307 W. Brooks, Room 205A, Norman, Oklahoma 73019, USA. Tel.: þ 1 405 325 5591; fax: þ1 405 325 5491. E-mail address: [email protected] (P.K. Yadav).
http://dx.doi.org/10.1016/j.jfineco.2014.07.012 0304-405X/& 2014 Elsevier B.V. All rights reserved.
Regulators and financial journalists have widely perceived FTDs, and specifically “naked” short sales, as having a very negative impact on markets.1 For example, a report in Time Magazine (“Watch out, they bite,” November 9, 2005), quoting estimates by former Under Secretary of Commerce Robert J. Shapiro, alleges that naked short selling has “cost investors $100 billion and driven 1,000 companies into the ground”; and an article in Euromoney (“Short selling: the naked truth,” December 2008) claims that, “Fails to deliver in the US equity market have exacerbated the sharp declines in share prices of financials. Although the SEC is clearing up the mess caused by naked short selling, more drastic measures might be needed to restore confidence.” In this context, this paper investigates the overall net impact (on day tþ1) of the change in the open interest of undelivered positions (i.e., the change in FTDs) generated
1 The term “naked short sale” is used (in this paper) to describe a short sale that, irrespective of any intent considerations, fails to deliver because of the short seller not making timely stock borrowing arrangements.
494
V. Fotak et al. / Journal of Financial Economics 114 (2014) 493–516
collectively by trades on day t (and observed on day t þ3). First, we analyze the effect of FTD changes on pricing efficiency and liquidity. Second, we examine whether FTD changes played a causal role in the major price declines or the demise of financial institutions during the 2008 financial crisis. Settlement on US stock exchanges is done electronically through the Depositary Trust and Clearing Corporation, its subsidiaries, and associated agencies (hereafter collectively referred to as the DTCC). Ownership records are largely held, tracked, and transferred electronically through the DTCC, with more than 99.9% of all trades involving only electronically held securities (Morris and Goldstein, 2009). The DTCC becomes the central counterparty of all dulymatched error-reconciled trades, and electronically checks and updates relevant stock ownership accounts. If the stock ownership accounts associated with a clearing member with a net delivery obligation do not actually include the stock needed for delivery on settlement day, the undelivered position becomes a FTD.2 FTDs can potentially arise from both short and “long” sales (i.e., sales of duly-owned stock). In a short sale, such a delivery shortfall arises when the stock is not borrowed in time to credit the stock ownership account by settlement day.3 In a long sale, a delivery shortfall arises, for example, when (duly-owned) sold stock has been lent out and not returned to the stock ownership account in time for settlement; or because the broker has not ensured electronic recording of a paper certificate prior to the trade; or in conjunction with oversubscribed securities at security issuance. There has been a strong regulatory focus on reducing FTDs. In January 2005, Regulation SHO introduced requirements to “locate” stock prior to every short sale to reduce FTDs due to stock-borrowing delays; it also forced firm stock-borrowing arrangements prior to short sales in “threshold-list” stocks (that is, securities with high levels of FTDs) unless existing FTDs were “closed-out.” In July/ August 2008, a temporary SEC order mandated stock borrowing arrangements prior to short selling in select financial stocks to “eliminate any possibility that naked short selling may contribute to the disruption of markets” (SEC Release 58166, 2008). After Lehman's collapse in September 2008, through Rule 204T (later made permanent), FTDs arising from short sales were virtually eliminated by requiring borrowing or purchasing by the broker by the morning after the failure day; FTDs arising from long 2 A delivery failure by an individual client account will not become a FTD at DTCC level if another client account held with the same clearing member lends the stock explicitly, or implicitly because of being due to receive delivery. Due to such “netting,” a FTD at DTCC level cannot generally be uniquely attributed to a specific trade or trader; nor is it possible to identify whether it originated from a short or long sale. 3 The Regulation SHO “locate” requirement introduced in January 2005 requires broker-dealers to ensure (prior to a short sale) that there are “reasonable grounds” to believe that the security could be borrowed for timely delivery, rather than identify a separate block of shares for each specific short sale. As a consequence, a lender may indicate to multiple potential borrowers that shares are available, but be unable to meet all demands. The requirement could also often be fulfilled by using published lists of easy-to-borrow securities (Welborn, 2008).
sales or market-maker FTDs were given three additional days. Similarly, in November 2012, new EU rules instituted a pre-trade borrowing requirement for short sales. Regulatory concerns have been further exacerbated by the widespread allegations and the extensive litigation about FTDs being used manipulatively by naked short sellers. Such “abusive” naked short selling is widely alleged to have contributed to the financial crisis by precipitating sharp price declines of financial firms.4 Notwithstanding extensive concerns, the SEC (in Report 450, March 2009) says that “there is hardly unanimity in the investment community or the financial media” on the dangers involved, and “despite its assertions regarding the potential of danger… the [SEC] Report can cite to no [relevant] bona fide studies” on the associated market impact. In the context of the strong regulatory focus, we begin by examining whether changes in FTDs are followed by changes in prices, pricing errors, intraday volatility, bid-ask spreads, and order imbalances. Our main sample consists of all the 1,492 NYSE ordinary common-share issues for which all relevant data are available over the period January 2005 to June 2008. For robustness, we replicate our core analysis on a similarly constructed sample of 2,381 Nasdaq ordinary common-share issues over the same period. First, we analyze liquidity and pricing efficiency for securities in different portfolios based on the number of FTDs. Second, we utilize ordinary least-squares (OLS) regressions and Granger-causality tests to investigate the link between FTD changes and market quality. Third, given the complex and endogenous interrelationships between market quality metrics and other market variables, we fit vector autoregressive models and test these relationships using impulse response functions. Fourth, we examine whether FTDs by market-makers affect securities differently from FTDs by other “public traders,” by using a proxy based on the reduction in FTDs subsequent to SEC Rule 204T in September 2008, which selectively precluded public-trader FTDs but not market-maker FTDs. We find that, during our sample period, FTDs affect about 95% of NYSE securities. For this NYSE sample, we conclude that an increase in FTDs equivalent to 0.1% of the number of outstanding shares leads to a 3% reduction in the magnitude of positive and absolute pricing errors; a 0.2% decline in intraday volatility; a 1.7% reduction in bid-ask spreads; and a 0.70% reduction in order imbalances. Each of these changes is statistically significant and not subject to significant subsequent corrections. Similar results are obtained in subsamples of securities affected by high levels of FTDs, or by persistent FTDs. Further, the beneficial 4 Over a two-year period leading up to the 2008 financial crisis, the SEC received more than 5,000 complaints in this regard (Wall Street Journal Asia, March 20, 2009), and a Factiva search shows over 4,600 printed English-language articles on the subject. Several investor associations and high-profile chief executive officers (CEOs) lobbied aggressively. The huge volume of litigation alleging associated stock price manipulation led to “naked” short selling being called the “Holy Grail….bigger than tobacco” for plaintiffs' lawyers (Stokes, 2009). Examples of lawsuits include the Biovail lawsuit against Stephen Cohen, Gradient, and others; the Overstock lawsuit against Rocker Partners, Gradient, and others; and the NFI lawsuit against Bank of America (the Specialist) and the Prime Brokers.
V. Fotak et al. / Journal of Financial Economics 114 (2014) 493–516
impact on pricing efficiency and liquidity is driven by FTDs of both market-makers and public traders. Overall, our results are consistent with failing-to-deliver traders acting as value arbitrageurs and liquidity suppliers, enhancing pricing efficiency and providing liquidity as needed. We observe that the regulatory focus on reducing FTDs has been particularly concerned with FTDs arising from short sales, despite extant literature documenting that short sales have a beneficial overall impact on pricing efficiency and liquidity by enabling value-traders to more effectively bring prices of overpriced securities in line with their true value, and facilitating financial intermediaries in providing liquidity more expeditiously.5 This suggests an underlying regulatory belief that short sales that fail to deliver have an impact on liquidity and pricing efficiency that is different from the corresponding impact of short sales that result in timely delivery. However, from an economic perspective, that should arguably not be the case. First, short sales that fail to deliver might not differ from timely delivered short sales at the time of the trade since a short seller need not definitively know on day t of the trade whether or not she will deliver or fail on settlement day tþ3. This is because that decision may rationally be taken only on day t þ3 on the basis of the rebate rates in the overthe-counter (OTC) stock-borrowing market on day tþ3.6 The short seller may want to borrow and deliver if these rebate rates are positive and fail if they are negative. Second, as explained by Culp and Heaton (2008), a FTD results either in automatic stock-borrowing at zero rebate rates from a pool of voluntary lenders under the DTCC Stock Borrow Program, or in forced stock-borrowing at zero rebate rates from a randomly assigned broker with a long stock position—and little incentive to enforce delivery by forcing a “buy-in” from the market, since the random assignment potentially changes every day (Putniņš, 2010). As far as the market is concerned, since the stock is being temporarily borrowed anyway (voluntarily or otherwise), there should not be any functional consequences arising from whether a short sale ultimately results in a FTD or not. Hence, any FTDs that arise from short sales should contribute to price discovery and liquidity on the day of trade in the same way as timely delivered short sales. In the context of the above, we extend our empirical analysis (on the impact of FTD changes on pricing efficiency and liquidity) to also examine the corresponding impact of delivered short sales, and test whether that impact is different. We find strong evidence that the impact on our market-quality metrics of daily changes in FTDs, and of
5 While there are some dissenting opinions, there is broad agreement behind an extensive literature finding that short selling leads to improvements in pricing efficiency and liquidity (Diamond and Verrecchia, 1987; Abreu and Brunnermeier, 2002, 2003; Miller, 1977; Bris, Goetzmann, and Zhu, 2007; Diether, Lee, and Werner, 2009; Boehmer, Jones, and Zhang, 2008). 6 Since stock-borrowing arrangements are ordinarily accompanied by same-day delivery, it is economically not rational for short sellers to borrow prior to the delivery date and pay extra borrowing fees (Geczy, Musto, and Reed, 2002). Evans, Geczy, Musto, and Reed (2009) show that short sellers fail to deliver when rebate rates are negative. Boni (2006) offers evidence of correlation between fails and borrowing costs, which she interprets as indicative of “strategic” fails.
495
daily changes in short interest net of FTDs (our proxy for delivered short sales), is similar in both magnitude and significance. To focus on FTDs more likely to originate from short sales, and to control for any omitted variables, we investigate the securities affected by an exogenous event: the imposition of a temporary SEC Emergency Order (Release 58166 dated July 15, 2008) that remained in force between July 21 and August 12, 2008. This order required pre-borrowing arrangements prior to all short sales in 19 selected financial stocks and hence did not directly affect any FTDs arising from “long” sales. We find that during the order validity period, reported FTDs for the affected securities reduce to about 12% of their pre-order average within one week, to about 2% of their pre-order average within three weeks, and finally, on the last day of the order period, reported FTDs are below the minimum reportable threshold for each and every affected security. We also find that securities affected by the order display a significant increase in intraday volatility, absolute pricing errors, and bid-offer spreads while the order was in force: i.e., the stock-borrowing prior to short sales mandate designed to reduce FTDs significantly worsened pricing efficiency and liquidity. While this natural experiment may not be generalizable because very few securities in the same industry and during a particularly turbulent time are affected, our results indicate that FTDs for these firms appear to arise largely from short sales, and contribute beneficially to market liquidity and pricing efficiency. In the context of the extensive concerns articulated in the press, and by investor groups and company CEOs, the final section of our empirical analysis investigates whether FTDs caused price crashes and distortions for specific securities during the 2008 financial crisis period.7 Since FTDs have been blamed for the price crashes of Bear Stearns Companies Inc. (“Bear Stearns”), Lehman Brothers Holdings Inc. (“Lehman”), Merrill Lynch & Co. Inc. (“Merrill”), and American Insurance Group (“AIG”), we test for high levels of FTD changes prior to the large price declines in the stock prices of those companies. We find that FTDs were too few for any significant stock price distortions on most days, and when FTDs did become abnormally high, it was after price declines, not before. In each case, FTDs were responding to information about the firms, rather than being responsible for triggering price declines. Overall, we find no evidence that FTDs played a causal role in the demise of financial institutions or in generating price distortions during the 2008 financial crisis. This research contributes to the literature on settlement systems and delivery failures. Fails are not actual “failures” of delivery, but delays with a median age of three days as per Boni (2006). Evans, Geczy, Musto, and Reed (2009) show that “the alternative to fail is valuable and important 7 Press sentiment is exemplified by statements such as “when Bear and Lehman made their final leap off the cliff of history, both undeniably got a push—especially in the form of a flat-out counterfeiting scheme called naked short selling” (Rolling Stone, October 2009). Concerned investor associations include The Movement for Market Reform, National Coalition against Naked Short Selling, and Coalition for the Reform of Regulation SHO. An example of a crusading high-profile CEO is Patrick Byrne of Overstock.com, which filed several lawsuits against alleged naked-short sellers and financial institutions they accused of facilitating “naked-shorting”.
496
V. Fotak et al. / Journal of Financial Economics 114 (2014) 493–516
to the pricing and trading of options.” Our results additionally and importantly show that the ability to fail has a beneficial impact on liquidity and pricing efficiency of equity markets as well. Earlier, Merrick, Naik, and Yadav (2005) have shown that the ability to fail is an important release valve for settlement-related pressures and manipulative distortions. Similarly, the difficulty of borrowing shares when short selling demand is high leads to costly frictions which negatively impact market quality; and the ability to fail is a release valve that helps to protect traders from the worst effects of these frictions: it is particularly valuable when stock-borrowing is so costly that short selling rebate rates become negative, which is exactly when liquidity is most needed in the stock-borrowing market (Evans, Geczy, Musto, and Reed, 2009).8 Hence, regulatory removal of the ability to fail for all public traders is debatable when progressive fines for settlement delays can arguably be effective without being an extreme solution. It is more important for regulators to focus on removing the economic incentives for delivery failures by improving the liquidity, transparency, and regulation of the stock-borrowing market. This research also contributes significantly to the short selling literature. First, extending the extant evidence on the market-quality benefits of short selling, our results suggest that short sales that fail to deliver are as beneficial for market quality as short sales that deliver on time. Second, our findings have important implications for the extensive press coverage of naked short selling, and the strong regulatory response to it. Finally, we extend the literature on the impact of short sales restrictions: the ban on FTDs in some stocks in July 2008 was effectively an additional constraint on short selling and our evidence shows that it had predictably negative consequences on market quality.9 The remainder of the paper is structured as follows. Section 2 outlines the data, the variables, and the measures utilized. Section 3 presents empirical results on the impact of FTD changes on pricing efficiency and liquidity. Section 4 investigates the role of FTD changes in creating price distortions during the 2008 financial crisis. Finally, Section 5 offers concluding remarks. 2. Data, variables, and measures 2.1. Data and sample The main sample consists of all NYSE listed common stocks in the Center for Research in Security Prices (CRSP) share codes 10 and 11 for which complete data are 8 These frictions are well documented by Kolasinski, Reed, and Ringgenberg (2013): They find that high levels of stock-borrowing demand lead to dramatic increases in loan fees and search costs related to supply constraints. 9 Several recent papers examine short selling restrictions in 2008: Boulton and Braga-Alves (2010) and Kolasinski, Reed, and Thornock (2013) examine aspects of the July/August 2008 Emergency Order that we investigate in this paper. Other papers—Battalio and Schultz (2011), Boehmer, Jones, and Zhang (2013), Autore, Billingsley, and Kovacs (2011), and Beber and Pagano (2013)—focus on the subsequent September 2008 short selling ban, which we do not examine.
available. By restricting the analysis to share codes 10 and 11, we exclude exchange-traded funds (ETFs) (in case FTDs are caused by delays in the creation of units) and other securities that are not common stocks. We thereby exclude securities classified by CRSP as certificates, American trust components, American Depositary Receipts (ADRs), shares of beneficial interest (SBIs), and units such as “depositary units” and “units of limited partnership interest,” as well as securities issued by other types of entities such as foreign firms, closed-end funds, and real estate investment trusts (REITs). We also restrict the sample to securities that are listed for at least six months during the sample period (in order to have adequate data to fit vector autoregressive models) and that have been trading for at least one year prior to the sample period. Having a year of trading data allows for the estimation of pricing errors as discussed later in this section. We further require that the number of shares outstanding does not vary by more than 10% over any single day, to control for the unusual volumes of FTDs around primary market transactions documented by Edwards and Hanley (2010); though our results are robust to resampling without this constraint. In view of the potentially confounding influence of the July 2008 Emergency Order and the restrictions on short selling and FTDs subsequently imposed in September and October 2008, we confine our analysis to the period up to June 30, 2008. The main sample thus consists of 1,492 NYSE securities over January 2005 to June 2008. For robustness tests, we construct a comparable sample of 2,381 Nasdaq securities for the same period. Short interest data are from www. shortsqueeze.com and short sales data (available from January 2005 onwards) are from the NYSE. The number of shares outstanding is from CRSP. Market quality measures are based on NYSE Trade and Quote (TAQ) data. 2.2. Outstanding fails-to-deliver ratio (OFR) and outstanding delivered shorts ratio (ODR) Our metric for FTDs is the Outstanding fails-to-deliver ratio (OFR), defined for each day t as the number of outstanding failed positions reported on day tþ3 scaled by the total number of shares outstanding of the firm. We obtain data on the number of shares outstanding from CRSP, while data on outstanding FTDs have been made available by the SEC under the Freedom of Information Act (FOIA) since March 22, 2004. Reported raw FTD data from the SEC represent the open interest of failed-to-deliver trades that occurred up to three days prior. For each day t, we look at the number of outstanding FTDs on day tþ3, as we are interested in investigating the impact of FTD changes on the market on the day on which the transaction actually occurs (day t), rather than on the day on which the resulting failures are disclosed (day t þ3). Hence, OFR on day t is equal to the number of FTDs for day t þ3, scaled by shares outstanding on day t. Our approach of scaling the number of FTDs by the number of shares outstanding is similar in spirit to previous literature that relies on measures of “relative short interest” (DeChow, Hutton, Meulbroek, and Sloan, 2001; Desai, Thiagarajan, Ramesh, and Balachandran, 2002; Chen and Singal, 2003), and mirrors recent studies using FTD-based data (for example,
V. Fotak et al. / Journal of Financial Economics 114 (2014) 493–516
Boulton and Braga-Alves, 2010). FTD changes are measured as the daily change in OFR. Over our sample period, the SEC reports the number of outstanding FTDs only for a trading day on which it exceeds 10,000 shares of a particular security. Accordingly, on a day on which no figure is reported for a particular security, we are unable to determine the true number of outstanding FTDs—we simply know that it is less than 10,000 shares. In the absence of reported data, we record the number of FTDs as zero. Our approach is consistent with extant literature using FTD data (for example, Boulton and Braga-Alves, 2010). Accordingly, reported figures somewhat understate the true extent of fails. To put this into perspective, we note that the average number of shares outstanding for the securities in our NYSE sample is about 212 million. Further, over the days on which nonzero FTDs are reported, the average number of reported FTDs is 97,068 shares, which is well in excess of the minimum reporting threshold. In terms of our tests regarding the market impact of FTDs, this data bias makes our findings more conservative, rather than less. First, by inducing a level of noise, it is lowering the statistical power of our tests. Second, it is arguably unlikely that securities with low levels of FTDs would suffer from distortions in market quality due to failed trades, given that we find that securities with high levels of FTDs do not display such problems. In order to examine whether claims regarding the market impact of “naked” short selling are valid, we construct a control variable, the Outstanding delivered shorts ratio, or ODR, which is the estimated short ratio net of FTDs. The Outstanding delivered shorts ratio is intended to be the open interest of timely delivered short sales positions scaled by the number of shares outstanding. We estimate the open interest of timely delivered short sales by the difference between the estimated short interest and the number of outstanding failed positions. The implicit assumption in this estimate of delivered short sales is that all FTDs arise from short selling. Additionally, while we have daily data on short sales from TAQ, short interest data are biweekly; accordingly, we calculate the short interest on a day within a biweekly period as the sum of the short interest reported at the start of the period and the cumulative daily changes in short interest since the start of that period. The daily change in short interest is calculated as the difference between the daily total short sales and an estimate of the daily short positions covered. Following Diether, Lee, and Werner (2009), we assume that the number of shares covered on each day of the biweekly interval is equal to the total short sales over the period minus the change in short interest over that period, divided by the number of trading days in that period. We use the daily change in ODR as our measure of timely delivered short sales. 2.3. Pricing error We define the “pricing error” on any day as the difference between the observed price on that day and the estimated information-efficient price for the same day. The unobservable daily estimate of the informationefficient random-walk or “fundamental” price of the security, a latent stochastic variable, is estimated for each sample
497
security using a Kalman-filter methodology as in Hamilton (1985). The procedure involves establishing two equations. The first equation dictates the evolution of the latent variable, and in our case we assume, in the spirit of Hasbrouck (1993), that the logarithm of the stock's underlying or information-efficient value, F(t), follows a random walk with a drift, μ, and a white noise innovation, ε(t), with mean zero and variance σ 2ε : FðtÞ ¼ μ þFðt 1Þ þ εðtÞ;
ε Nð0; σ 2ε Þ:
ð1Þ
The second equation relates the observed and latent variables, i.e., specifies the pricing error process. We assume that the pricing error Y(t) follows a meanreverting process around zero, with α, the rate of meanreversion, ranging between zero and one. Pricing errors correct fully in one period when α is equal to one, and do not correct at all when α is equal to zero:
ΔYðtÞ ¼ αYðt 1Þ þ ϕðtÞ;
ϕ Nð0; σ 2ϕ Þ:
ð2Þ
The observed log of stock price S(t) is the sum of the fundamental price and pricing error: SðtÞ ¼ FðtÞ þYðtÞ:
ð3Þ
Hence; SðtÞ ¼ μ þ ð1 αÞSðt 1Þ þ αFðt 1Þ þ θðtÞ;
θðtÞ ¼
ϕðtÞ þ εðtÞ: The Expectation Maximization (EM) algorithm (Dempster, Laird, and Rubin, 1977) is employed to compute the Maximum Likelihood (ML) estimate of the unobservable variable, F(t), based on data relating to the observed variable, S(t). Hamilton (1985) employs such an approach to estimate expected quarterly inflation, the latent variable, based on observed actual inflation. In exactly the same way, we utilize the observed daily stock prices to infer the daily unobserved fundamental price, and hence the daily pricing error, using daily closing price data from CRSP. The statespace representation of the system is as follows: Measurement equation: 2 3 SðtÞ 6 7 ½SðtÞ ¼ 1 0 0 4 FðtÞ 5 ð5Þ εðtÞ Transition equation: 2 3 2 SðtÞ ð1 αÞ α 6 FðtÞ 7 6 1 4 5¼4 0
εðtÞ
0
0
3 3 2 Sðt 1Þ θðtÞ 7 76 7 6 0 54 Fðt 1Þ 5 þ 4 εðtÞ 5 0 1 0
μ
32
ð6Þ
To test the bias and efficacy of the pricing error estimation process, we run 500 simulations of both fundamental price and pricing error for a hypothetical stock over 252 trading days assuming a range of volatility parameters and mean-reversion parameters. In each case, we add the fundamental price and the pricing error to arrive at the equivalent of a simulated “observed” price. Then, we run our Kalman-filter estimation procedure on this observed price series to determine our Kalman-filter estimate of the originally simulated fundamental price. Finally, we run a regression of changes in the originally simulated fundamental price on changes in our Kalman-filter estimate of that fundamental price. In each and every case, the regression intercept is not significantly different from zero, and
498
V. Fotak et al. / Journal of Financial Economics 114 (2014) 493–516
the regression slope is not significantly different from one; and the root mean square error in the estimated fundamental price is economically small in magnitude. We employ three measures of pricing efficiency related to pricing errors. First, we use duly-signed pricing errors as estimated. Second, we use the absolute values of pricing errors, which given the zero-mean distribution of pricing errors, are also a direct measure of the volatility of pricing errors. Finally, given that the impact of FTDs on pricing errors could be asymmetric, we use a dummy for positive pricing errors to distinguish between positive and negative pricing errors. 2.4. Liquidity measures and other variables The liquidity measures we use are relative order imbalance, bid-ask spread, and trading volume defined as in Table 1. Our measure of volatility is the intraday volatility computed as the daily standard deviation of five-minute stock returns. The definitions of all of our variables are based on extant empirical literature analyzing NYSE TAQ data, and are summarized in Table 1. 3. Empirical results: FTDs, pricing efficiency, and liquidity 3.1. Preliminary descriptive analysis of overall sample We sort securities into ten deciles based on mean OFR, computed over the entire sample period (January 1, 2005 to June 30, 2008). In Table 2, we report means, medians, and standard deviations of the main variables, for the overall sample and for the top and bottom deciles. We find that overall mean OFR is about 0.04%. By construction, OFR varies by decile: the mean is less than 0.01% in decile 1 and about 0.29% in decile 10. ODR is similarly higher in
decile 10 (13.35%) than in decile 1 (4.08%), with an almost monotonic positive relationship between mean OFR and mean ODR across (OFR-based) deciles of different securities. However, even though securities with higher mean OFR tend to also have a higher mean ODR, the average timeseries correlation between ODR and OFR of a particular security is, in contrast, extremely small—only 0.08. Securities in deciles 1 and 10 do not differ significantly in terms of mean pricing errors. Securities with higher OFR display significantly higher positive pricing errors and absolute pricing errors, which is consistent with FTDs intensifying when securities are overpriced, although we should not infer causality from this correlation. Similarly, higher OFR is associated with higher and positive order imbalances and higher intraday volatility, but lower spreads. Our results also indicate that FTDs are significantly more common for smaller firms, as average market capitalization for firms in decile 1 (USD 9.1 billion) is almost six times larger than for firms in decile 10 (USD 1.6 billion). As discussed in Section 2.2, FTD reporting is subject to a minimum threshold of 10,000 shares. It is reasonable to conjecture that this reporting threshold is relatively more onerous for securities with smaller market capitalization and, likely, a lower number of shares outstanding. That is not actually so. The proportion of security-days with no reported FTDs is 97.6% in OFR-decile 1, and drops monotonically across deciles to 48.0% for securities in decile 10; the daily average number of outstanding FTDs for the days on which FTDs exceed the 10,000 share-reporting threshold ranges from 29,031 for securities in decile 1 to 357,238 for securities in decile 10. Since securities with smaller mean market capitalization generally have higher mean OFR, they also have a relatively larger average number of reported FTDs (in number of shares) on the days on which
Table 1 Variable definitions. This table lists the variables used in the analysis and provides definitions. All variables are daily, unless otherwise specified. FTDs and short selling Outstanding fails-to-deliver ratio (OFR) Outstanding delivered shorts ratio (ODR) Pricing error Pricing error (PE) Negative pricing error (Negative PE) Positive pricing error (Positive PE) Absolute pricing error (Abs PE) Positive PE dum Liquidity Order imbalance (OIB)
For each day t, ratio of the number of outstanding fails-to-deliver reported on day t þ3, over total shares outstanding on day t. Ratio of the estimated number of outstanding delivered shorted shares over total shares outstanding.
The non-random walk component of a daily return series estimated using a Kalman-filter methodology. A variable set equal to the pricing error when pricing error is negative, to zero otherwise. A variable set equal to the pricing error when pricing error is positive, to zero otherwise. The absolute value of the pricing error. A binary variable set equal to one if pricing error is positive, to zero otherwise.
Positive OIB dum Spread Volume
The daily sum of the five-minute difference between the market value of shares traded in buyer-initiated trades and the market value of shares traded in seller-initiated trades, divided by total daily dollar trading volume. A binary variable set equal to one if OIB is positive and to zero otherwise. The daily average of the ratio of the difference between bid and ask and the mid-quote. Daily number of shares traded.
Other Price Return Intraday volatility Market value
The The The The
natural log of the daily average of the five-minute midquote. daily stock price return. daily average of standard deviation of the five-minute stock price return. number of shares outstanding multiplied by the closing price for the day.
V. Fotak et al. / Journal of Financial Economics 114 (2014) 493–516
499
Table 2 Summary statistics. All variables are as defined in Table 1. The Outstanding fails-to-deliver ratio (OFR) is computed for all NYSE common stock of US-based firms (CRSP share codes 10 and 11) listed for at least 18 months over the interval January 1, 2005 to June 30, 2008, and with no large daily change ( 410%) in the number of shares outstanding and for which all required data are available (n¼ 1,492). Securities are ranked by mean OFR and allocated to decile 1 (lowest) through 10 (highest). Daily statistics are computed by security, security averages are further averaged over deciles and for the entire sample. This table reports mean, median, and standard deviation for the sample, means for deciles 1 and 10, and the difference between those, along with results of a t-test for differences in means across deciles 1 and 10. “n”, “nn”, and “nnn” indicate significance at the 10%, 5%, and 1% level, respectively. Sample
Outstanding fails-to-deliver ratio Outstanding delivered shorts ratio Pricing error Positive pricing error Abs pricing error Price (log) Intraday volatility Return Spread Order imbalance Market value (US$ M) Obs
OFR decile means
Difference
Mean
Median
Std
OFR D1
OFR D10
OFR D10–D1
p-Value
0.04% 5.46% 0.15% 0.51% 1.18% 3.3001 0.21% 0.07% 0.18% 6.72% $8,223.33 1,492
o0.01 4.34% o 0.01% 0.10% 0.21% 3.4171 0.20% 0.06% 0.11% 6.92% $2,087.97 1,492
0.18% 5.32% 4.33% 1.47% 4.96% 0.7631 0.08% 0.11% 0.27% 5.71% $24,212.97 1,492
o 0.01 4.08% o 0.01% 0.37% 0.74% 3.1484 0.21% 0.05% 0.39% 4.00% $9,128.81 149
0.29% 13.35% 0.55% 1.06% 2.66% 2.9121 0.28% 0.03% 0.26% 8.26% $1,598.89 150
0.29% 9.27% 0.55% 0.68% 1.92% 0.2363 0.08% 0.02% 0.13% 4.26% $7,529.91 149
o 0.01nnn o 0.01nnn 0.35 o 0.01nnn o 0.01nnn 0.03nn o 0.01nnn 0.26 o 0.01nnn o 0.01nnn o 0.01nnn
FTDs are reported, and a lower proportion of days with no reported FTDs.
3.2. Portfolio approach As a first test of the relationship between FTD changes and subsequent market quality, on each day t in our sample spanning January 2005 to June 2008, we group the 1,492 securities into nine portfolios based on changes in OFR and ODR. We start by estimating the daily time-series standard deviation in OFR by security. On each day, we include securities with a one-standard deviation or greater decrease in OFR into a “FTD decrease” portfolio and, similarly, securities with a one-standard deviation or greater increase in OFR into a “FTD increase” portfolio. We include all remaining securities into a “FTD constant” portfolio. In order to control for the extent of timely delivered short sales, we replicate the same procedure on the basis of timely delivered short sales, proxied by changes in ODR, thus forming the portfolios “Delivered short sales decrease,” “Delivered short sales constant,” and “Delivered short sales increase”. Finally, we intersect those groups of securities, forming nine final portfolios. For each portfolio, we compute average changes in our metrics of price levels, intraday volatility, market liquidity, pricing errors, absolute pricing errors, and order imbalances for the following day (t þ1). All variables are standardized and winsorized (at three standard deviations) by security. We compute nextday average changes for the portfolios “FTD decrease, Delivered short sales constant” and “FTD increase, Delivered short sales constant” and we further test for differences between these averages. Our interest lies in changes in OFR. We form portfolios on changes in ODR to control for the impact of estimated delivered short sales on liquidity and pricing efficiency. We present results for the two key portfolios (the ones
including low and high levels of FTD changes while experiencing normal levels of delivered shorts), rather than for the full set of nine portfolios, primarily for the sake of brevity. Our focus on the portfolios with high or low levels of FTD changes and “normal” levels of delivered shorts, allows us to draw inferences about FTD changes under “normal” short selling conditions. Also, the samples with abnormal levels of both FTDs and estimated delivered shorts vary greatly in size, leading to comparisons of data samples of dramatically different sizes. The results of this analysis are presented in Table 3. For the portfolio “FTD increase, Delivered short sales constant,” we find that, on the following day, there is a 4.2% reduction in absolute pricing error, a 0.1% decrease in spreads, a 0.4% decrease in order imbalances, and a 0.67% decrease in intraday volatility—all results being statistically significant at the 1% level, except for the result on order imbalances, which is significant at 5%. We also find a decrease in price and pricing error, but estimates are not statistically significant. For the portfolio “FTD decrease, Delivered short sales constant,” we similarly observe a decrease in absolute pricing error, but the magnitude of the estimated effect is smaller (0.18%). We further investigate the differences in next-day metrics between the two portfolios. The “FTD increase” portfolio, compared to the “FTD decrease” portfolio, displays lower absolute pricing errors, spreads, order imbalances, and intraday volatility, with all results being statistically significant at the 1% level. This initial analysis offers evidence that large FTD changes are followed by next-day improvements in market liquidity and pricing efficiency. Given the limitations of this univariate framework, in the following section we employ OLS regressions to investigate the relationship between FTD changes and subsequent market quality while controlling for timely delivered short sales and relevant systematic factors.
500
V. Fotak et al. / Journal of Financial Economics 114 (2014) 493–516
Table 3 FTDs and market quality, portfolio approach. All variables are as defined in Table 1. All variables are standardized and winsorized by security. The sample includes all NYSE common stock of US-based firms (CRSP share codes 10 and 11), listed in the CRSP and TAQ databases for at least 18 months over the interval January 1, 2005 to June 30, 2008, and with no large changes ( 410%) in the number of shares outstanding. The resulting sample includes 1,492 securities. On each day t between January 2005 and June 2008, securities with a one-standard deviation or greater decrease in OFR are allocated to a “FTD decrease” portfolio and, similarly, securities with a one-standard deviation or greater increase in OFR to a “FTD Increase” portfolio. All remaining securities are allocated to a “FTD constant” portfolio. The same procedure is repeated on the basis of the intensity of delivered short selling, proxied by changes in ODR, thus forming the portfolios “Delivered short sales increase,” “Delivered short sales decrease,” and “Delivered short sales constant.” Those portfolios are intersected, forming nine final portfolios. Averages of next-day (tþ 1) changes in returns, return volatility, market liquidity, pricing errors, and order imbalances are computed. The table presents results for the “FTD increase, Delivered short sales constant” and the “FTD decrease, Delivered short sales constant” portfolios, along with results of a test for differences between them. p-Values are from two-sided t-tests. “n”, “nn”, and “nnn” indicate significance at the 10%, 5%, and 1% level, respectively. FTD increase, delivered short sales constant
p-Value
FTD decrease, delivered short sales constant
p-Value
Difference
p-Value
2.31% 0.28% 4.20% 0.10% 0.40% 0.67%
0.31 0.41 o 0.01nnn o 0.01nnn 0.01nn o 0.01nnn
0.15% 0.04% 0.18% o 0.01% 0.09% 0.01%
0.36 0.36 o0.01nnn 0.22 0.12 0.14
2.46% 0.24% 4.02% 0.10% 0.31% 0.66%
0.75 0.77 o 0.01nnn o 0.01nnn o 0.01nnn o 0.01nnn
ΔPrice(tþ 1) ΔPE(tþ 1) ΔAbs PE(t þ1) ΔSpread(t þ1) ΔOIB(tþ 1) ΔIntraday volatility (tþ 1) N (days)
844
849
3.3. Panel OLS regressions We next estimate panel regressions to investigate the relationship between FTD changes and next-day marketquality metrics. Accordingly, we estimate six separate regressions, with changes in pricing errors, absolute pricing errors, prices, intraday volatility, spreads, and order imbalances as responses. Our main explanatory variable is the previous-day change in OFR. Given the regulatory focus on reducing FTDs arising from short sales, and the welldocumented impact of short sales on liquidity and pricing efficiency, our empirical tests aim to control for the estimated impact of delivered short sales on market quality. Accordingly, we also add the previous-day change in ODR to the model. As further controls, we include, in each regression, previous-day changes of the dependent variable, to account for possible autocorrelations. In addition, to control for systematic effects, we include market averages of the changes in the same quality metrics. Finally, when modeling changes in pricing errors, we add interaction variables between previous-day changes in OFR and ODR and a binary variable identifying positive pricing errors, to allow for an asymmetric impact of fails on pricing errors. All regressions contain security fixed-effects. The models estimated are:
ΔPEi;t ¼ λ0;i þ λ1 ΔOFRi;t 1 þ λ2 ΔODRi;t 1 þ λ3 ΔOFRi;t 1 Positive_PE_Dumi;t 1 þ λ4 ΔODRi;t 1 Positive_PE_Dumi;t 1 þ λ5 ΔPEM;t þ λ6 ΔPEi;t 1 þ δi;t ð7Þ ΔPE_Volatilityi;t ¼ γ 0;i þ γ 1 ΔOFRi;t 1 þ γ 2 ΔODRi;t 1 þ γ 3 ΔPE_VolatilityM;t þ γ 4 ΔPE_Volatilityi;t 1 þ νi;t
ð8Þ
ΔVolatilityi;t ¼ δ0;i þ δ1 ΔOFRi;t 1 þ δ2 ΔODRi;t 1 þ δ3 ΔVolatilityM;t þ δ4 ΔVolatilityi;t 1 þ θi;t
ð9Þ
844
ΔSpreadi;t ¼ β0;i þ β1 ΔOFRi;t 1 þ β2 ΔODRi;t 1 þ β3 ΔSpreadM;t þ β4 ΔSpreadi;t 1 þ εi;t
ð10Þ
ΔOIBi;t ¼ κ 0;i þ κ 1 ΔOFRi;t 1 þ κ 2 ΔODRi;t 1 þ κ 3 ΔOIBM;t þ κ 4 ΔOIBi;t 1 þ ρi;t ð11Þ ΔLogmidi;t ¼ μi þ μ1 ΔOFRi;t 1 þ μ2 ΔODRi;t 1 þ μ3 ΔLogmidM;t þ μ4 ΔLogmidi;t 1 þ σ i;t :
ð12Þ
We estimate the above parameters with OLS regressions using daily data and standard errors are time-clustered. Variable definitions are included in Table 1; the i and t subscripts index, respectively, securities and days. The M subscript indicates a “market” equally weighted average computed over our entire sample. Δ is used to identify daily changes (for example, ΔOFRt ¼OFRt OFRt 1). Results for the overall sample are presented in Table 4. As reported, an increase in OFR is associated with a nextday decrease in absolute pricing errors, spreads, order imbalances, and intraday volatility (all significant at 1%). We find no significant relationship between increases in FTDs and price levels, nor with positive and negative pricing errors, when those are separately modeled. This second set of results is highly consistent with our previous findings: an increase in FTDs is related to improvements in liquidity and pricing efficiency. We recognize that the correlations documented in this analysis are not proof of causation. Hence, we attempt to gain insights into the direction of causality in our system by first testing for Granger-causality. For each metric of interest, we test both directions of Grangercausality: that is, we test whether previous-day FTD changes “Granger-cause” next-period market-quality metrics. We first find that FTD changes Granger-cause absolute pricing errors, spreads, order imbalances, and intraday volatility. Yet, when testing the other side of causality, we find it equally meaningful, as each of those variables Granger-causes FTD changes as well. FTD changes and market-quality metrics
V. Fotak et al. / Journal of Financial Economics 114 (2014) 493–516
501
Table 4 FTDs and market quality, OLS regressions. Coefficient estimates and p-values (in italics) from panel regressions with firm fixed effects and time-clustered standard errors. Metrics of market quality (in column headings) are regressed on lagged values of the predictors (listed in the first column). All variables are standardized and winsorized by security, and defined as in Table 1. The sample includes all NYSE common stocks of US-based firms (CRSP share codes 10 and 11), listed in the CRSP and TAQ databases for at least 18 months over the interval January 1, 2005 to June 30, 2008, and with no large changes ( 410%) in the number of shares outstanding. The resulting sample includes 1,492 securities and 1,001,656 firm-days. The “Market” variable on day t is obtained by averaging the related variable across all securities on day t. Δ is used to identify changes (for example, ΔOFRt ¼OFRt OFRt 1). “n”, “nn”, and “nnn” indicate significance at 10%, 5%, and 1% level, respectively.
ΔOFR(t 1) ΔODR(t 1) ΔPricing error (t 1) ΔOFR(t 1) Positive PE(t 1) ΔODR(t 1) Positive PE(t 1)
ΔPricing error(t)
ΔAbs PE(t)
ΔPrice(t)
ΔIntraday volatility(t)
ΔSpread(t)
ΔOIB(t)
0.0019 0.4127 0.1932 o 0.01nnn 0.3489 o 0.01nnn 0.0048 0.1375 0.4240 o 0.01nnn
0.0069 o 0.01nnn 0.2733 o 0.01nnn
o 0.0001 0.9718 0.0042 0.03nn
0.0068 o0.01nnn 0.0721 o0.01nnn
0.0125 o0.0nnn 0.1008 o0.01nnn
0.0093 o 0.01nnn 0.2599 o 0.01nnn
ΔAbs PE(t 1)
0.3763 o 0.01nnn
ΔPrice(t 1)
0.0255 o0.01nnn
ΔIntraday volatility(t 1)
0.1937 o0.01nnn
ΔSpread(t 1)
0.3214 o0.01nnn
ΔOIB(t 1) ΔMarket pricing error(t)
0.4480 o 0.01nnn 0.8559 o 0.01nnn
ΔMarket abs PE(t)
0.8363 o 0.01nnn
ΔMarket price(t 1) ΔMarket intraday volatility(t) ΔMarket spread(t) ΔMarket OIB(t)
Granger-cause each other, and the whole system displays feedback effects over time. Hence, it is necessary for us to investigate the impact of FTD changes within a statistical model that allows us to control for intertemporal interdependency amongst all metrics. Accordingly, in the next section, we present results from vector autoregressive (VAR) models, duly accounting for endogenous interrelationships between marketquality metrics, FTD changes, and timely delivered short sales. 3.4. Vector autoregressive framework In a third set of tests, we control for endogenous interrelationships using three vector autoregressive (VAR) models, with additional exogenous variable(s) added as controls. The system of equations underlying each of these models is formally described in Table 5. In VAR Model 1, the variables included are changes in OFR, ODR, Pricing error, Intraday volatility, Spread, and Order imbalance. We add, as predictors in the pricing error equation, (1) an interaction between lagged changes in OFR and a lagged binary variable set equal to one when pricing error is positive, and (2) an interaction between lagged changes in ODR and
0.9971 o0.01nnn 0.9150 o0.01nnn 0.8751 o0.01nnn 0.8016 o 0.01nnn
a lagged binary variable set equal to one when pricing error is positive. Accordingly, we are able to separately estimate the impact of FTD changes and delivered short sales on pricing error when the pricing error is positive in contrast to when the pricing error is negative. We add two more predictors to each of the equations in which the changes in OFR or ODR are the response variables: (3) a lagged binary variable set equal to one when order imbalance is positive and zero otherwise, and (4) a lagged binary variable set equal to one when pricing error is positive and zero otherwise. Finally, in each equation we add, as an exogenous variable, the market-wide equally weighted average of the dependent variable, to control for possible systematic effects. In VAR Model 2, we replace Pricing error by Absolute pricing error; given that we do not model pricing error here, we do not include the interaction-variables described above in points (1) and (2). VAR Model 3 is similar to Model 2, but we replace Pricing error by Price. We estimate VAR Models 1, 2, and 3 separately for each security. The results we report in Table 5 are obtained by first standardizing all variables by subtracting the securityspecific mean and dividing by the security-specific standard
502
V. Fotak et al. / Journal of Financial Economics 114 (2014) 493–516
Table 5 FTDs and market quality, VAR analysis. Results for the estimation of three vector autoregressive models of order one used to investigate the impact of changes in OFR and ODR: VAR with Pricing error (Model 1), VAR with Pricing error volatility (Model 2), and VAR with Price (Model 3). All variables are defined in Table 1 and are standardized and winsorized by security. The “entire market” sample includes all common stocks of US-based firms (CRSP share codes 10 and 11) with NYSE as the primary exchange, listed in the CRSP and TAQ databases for at least 18 months over the interval January 1, 2005 to June 30, 2008, and with no large changes (410%) in the number of shares outstanding. For the “Most FTDs” sample, securities from the sample described above are ranked by mean OFR and allocated on that basis to decile 1 (lowest) through 10 (highest); the securities included in decile 10 constitute the “Most FTDs” sample. The VAR models are defined below; the subscripts M and t indicate, respectively, the market-wide average of the variable of interest and day t; Δ is used to identify changes (for example, ΔOFRt ¼ OFRt OFRt 1). Panels A and B are extracts of estimates of the parameters in Models 1, 2, and 3 above and report results pertaining to the impact of, respectively, OFR and ODR. Reported parameter estimates are averages of parameters estimated by security. The p-values from two-sided t-tests are in italics below the parameter estimate. “n”, nn “ ”, and “nnn” indicate significance at the 10%, 5%, and 1% level, respectively. Panels C and D present the effect on the market-quality metrics included in Models 1, 2, and 3 of an increase in Outstanding FTD ratio (OFR) and Outstanding delivered shorts ratio (ODR) equivalent to 10 basis points of the number of outstanding shares. The panels report the impact on the response variable as a proportion of the mean of the response variable. Percentages presented in bold are significant at 10% or lower. Panels E and F are extracts of estimates of the parameters in Models 1, 2, and 3 above and report results pertaining to the impact of OFR. Panel E presents results from a subset of the “entire market” data set discussed in Panels A–D; the sample includes all common stocks of US-based firms (CRSP share codes 10 and 11) with NYSE listed as the primary exchange, listed in the CRSP and TAQ databases for at least 18 months over the interval January 1, 2005 to June 30, 2008, with no large changes (410%) in the number of shares outstanding and that appear, for at least one day, on the NYSE-published list of Threshold Securities. Panel F presents results from a data set including Nasdaq securities: all common stocks of US-based firms (CRSP share codes 10 and 11) with Nasdaq listed as the primary exchange, available in the CRSP and TAQ databases, listed for at least 18 months over the interval January 1, 2005 to June 30, 2008, and with no large changes (410%) in the number of shares outstanding. Reported parameter estimates are averages of parameters estimated by security. Significance is tested employing a cross-sectional estimate of the standard error of the parameter estimate. The p-values from two-sided t-tests are in italics below the parameter estimate. “n”, “nn”, and “nnn” indicate significance at the 10%, 5%, and 1% level, respectively. Model1: ΔY1t ¼ c1 þ β1ΔY1t 1 þ M1t þ S1t þ S2t þ ε1t Model2: ΔY2t ¼ c2 þ β2ΔY2t 1 þ M2t þ S1t þ ε2t
ε1t i:i:d: Nð0; ΩÞ
ε2t i:i:d: Nð0; ΩÞ
Model3: ΔY3t ¼ c3 þ β3ΔY3t 1 þ M3t þ S1t þ ε3t ε3t i:i:d: Nð0; ΩÞ 1 1 1 0 0 0 OFRt OFRt OFRt C C C B B B ODRt C B ODRt C B B ODRt C C C C B B B C C C B B B PEt C B B PE Volatilityt C B Pricet C Y1t ¼ B C; Y2t ¼ B C; Y3t ¼ B C B Volatilityt C B V olatilityt C B Volatilityt C C C C B B B C C B Spreadt C B B Spread Spread t t A A A @ @ @ OIBt OIBt OIBt 3 3 3 2 2 2 ψ1 ΔOFRM;t ψ7 ΔOFRM;t ψ13 ΔOFRM;t 7 7 7 6 ψ2 ΔODR 6 6 ψ8 ΔODRM;t 7 7 6 6 6 ψ14 ΔODRM;t 7 M;t 7 7 7 6 6 6 7 7 7 6 6 6 ψ3 ΔPE ψ9 ΔPEVolatility ψ15 ΔPrice M;t M;t M;t 7 7 7 6 6 6 M1t ¼ 6 7; M2t ¼ 6 7; M3t ¼ 6 7 6 ψ4 ΔVolatilityM;t 7 6 ψ10 ΔVolatilityM;t 7 6 ψ16 ΔVolatilityM;t 7 7 7 7 6 6 6 6 ψ5 ΔSpreadM;t 7 6 ψ11 ΔSpreadM;t 7 6 ψ17 ΔSpreadM;t 7 5 5 5 4 4 4 ψ6 ΔOIBM;t ψ12 ΔOIBM;t ψ18 ΔOIBM;t 3 2 τ1 Positive OIB dumt 1 þ τ2 Positive PE dumt 1 7 6 τ3 Positive OIB dum t 1 þ τ4 Positive PE dumt 1 7 6 7 6 7 6 0 7 S1t ¼ 6 7 6 0 7 6 7 6 5 4 0 0
3 0 7 6 0 7 6 7 6 6 τ5 ðPositive PE dumt 1 ΔOFRt 1 Þ þ τ6 ðPositive PE dumt 1 ΔODRt 1 Þ 7 7 6 S2t ¼ 6 7 0 7 6 7 6 5 4 0 2
0 ΔX t ¼ X t X t1
Panel A: Impact of FTDs
A: Entire market 1,407 Securities B: Most FTDs 150 Securities A: Entire market 1,407 Securities B: Most FTDs 150 Securities A: Entire market 1,407 Securities B: Most FTDs 150 Securities
ΔPE
ΔPE (incremental effect when lag PE 40)
0.0045 0.20 0.0451 o 0.01nnn
0.0168 o 0.01nnn 0.0785 o 0.01nnn
ΔAbs PE
ΔPrice
ΔIntraday volatility
ΔSpread
ΔOIB
0.0002 0.70 o 0.0001 0.99
0.0054 o 0.01nnn 0.0121 0.11 0.0071 o 0.01nnn 0.0189 0.01nn 0.0049 0.02nn 0.0133 0.08n
0.0126 o0.01nnn 0.0204 o0.01nnn 0.0095 o0.01nnn 0.0168 o0.01nnn 0.0103 o0.01nnn 0.0192 0.05nn
0.0093 o 0.01nnn 0.0130 0.31 0.0071 0.04nn 0.0115 0.37 0.0068 0.05nn 0.0140 0.28
0.0077 o 0.01nnn 0.0361 o 0.01nnn
V. Fotak et al. / Journal of Financial Economics 114 (2014) 493–516
503
Table 5 (continued ) Panel B: Impact of delivered short selling
A: Entire market 1,407 Securities B: Most FTDs 150 Securities A: Entire market 1,407 Securities B: Most FTDs 150 Securities A: Entire market 1,407 Securities B: Most FTDs 150 Securities
ΔPE
ΔPE (incremental effect when lag PE40)
0.5414 o 0.01nnn 0.8416 o 0.01nnn
1.3759 o 0.01nnn 1.4300 o 0.01nnn
ΔAbs PE
ΔPrice
0.8516 o 0.01nnn 0.7913 o 0.01nnn 0.0221 o 0.01nnn 0.0196 0.04nn
ΔIntraday volatility
ΔSpread
ΔOIB
0.9128 o 0.01nnn 0.9759 o 0.01nnn 0.9594 o 0.01nnn 1.0141 o 0.01nnn 0.9126 o 0.01nnn 0.9872 o 0.01nnn
0.2267 o 0.01nnn 0.0690 0.06n 0.2104 0.09n 0.0615 0.09n 0.2106 o 0.01nnn 0.0762 0.04nn
0.7665 o 0.01nnn 0.6021 o 0.01nnn 0.7629 o 0.01nnn 0.6035 o 0.01nnn 0.7628 o 0.01nnn 0.6079 0.01nn
Panel C: Impact of a change in Outstanding FTD ratio (OFR) equal to 10 basis points of the number of outstanding shares Overall sample Response variable Positive PE Abs PE Price Intraday volatility Spread OIB
Model 1 (PE)
Model 2 (Abs PE)
Most FTDs sample Model 3 (Price)
Model 1 (PE)
3.24%
Model 2 (Abs PE)Model 3 (Price)
2.59% 3.00%
0.18% 1.68% 0.73%
3.84% o 0.01% 0.17% 1.77% 0.55%
0.24% 1.27% 0.53%
0.12% 0.70% 0.35%
0.18% 0.58% 0.31%
o 0.01% 0.13% 0.66% 0.38%
Panel D: Impact of a change in Outstanding delivered shorts ratio (ODR) equal to 10 basis points of the number of outstanding shares Overall sample Response variable Positive PE Abs PE Price Intraday volatility Spread OIB
Model 1(PE)
Model 2 (Abs PE)
Most FTDs sample Model 3 (Price)
Model 1 (PE)
4.52%
Model 2 (Abs PE)Model 3 (Price)
2.39% 6.78%
3.49%
0.66% 0.58% 1.22%
0.01% 0.63% 0.58% 1.22%
ΔPE
ΔPE (incremental effect when lag PE 40)
ΔAbs PE
0.0409 o 0.01nnn
0.0857 o 0.01nnn
0.63% 0.62% 1.23%
0.49% 0.09% 0.96%
0.51% 0.08% 0.96%
0.01% 0.50% 0.10% 0.97%
ΔIntraday volatility
ΔSpread
ΔOIB
0.0176 0.02nn 0.0223 o 0.01nnn 0.0171 0.02nn
0.0158 0.01nn 0.0137 0.02nn 0.0146 o 0.01nnn
0.0142 0.23 0.0150 0.25 0.0178 0.17
ΔIntraday volatility
ΔSpread
ΔOIB
0.0053 0.02nn 0.0059 0.09n 0.0025 0.48
0.0048 0.07n 0.0042 0.09n 0.0045 o 0.01nnn
0.0054 0.04nn 0.0029 0.46 0.0004 0.93
Panel E: Impact of FTDs, Threshold Securities
163 Securities
ΔPrice
0.0268 0.02nn
175 Securities
0.0024 0.24
175 Securities Panel F: Impact of FTDs, Nasdaq sample
2223 Securities 2983 Securities 2983 Securities
ΔPE
ΔPE (incremental effect when lag PE 40)
0.0069 0.10
0.0162 o 0.01nnn
ΔAbs PE
ΔPrice
0.0135 o 0.01nnn 0.0019 0.02nn
504
V. Fotak et al. / Journal of Financial Economics 114 (2014) 493–516
deviation, and then winsorizing at three standard deviations from the mean. Based on an analysis of the Schwarz information criteria, we determine that, for all models, it is most appropriate to use a VAR of order one. In order to draw inferences about the true population parameters, we average coefficient estimates obtained for each security as in Fama and MacBeth (1973). In significance testing, we use cross-sectional estimates of standard errors. For robustness, we also estimate all three models without standardizing variables, but we face problems with the convergence of our maximum likelihood estimation algorithm, likely driven by the extreme differences in magnitude across our variables. For those securities for which we manage to obtain results without standardizing variables, our results are qualitatively similar to those reported. We estimate VAR Models 1, 2, and 3 for both our “entire market” sample and for a subset of the sample containing only the 10% of securities with the highest level of FTDs based on mean OFR over the sample period, January 2005 to June 2008 (the same subsample described as “decile 10” in Table 2). We focus on this subset of securities with the highest FTDs to find whether the impact of FTD changes differs when they occur most frequently. We find that the sign and the significance of the various inter-relationships involved are economically reasonable for both samples. For compactness, ease of interpretation, and to preserve the focus on the specific equations of interest to us, we report in Table 5 only those relationships that are relevant to the issues addressed in the paper. In particular, we report, for all three models and for both samples, the coefficient estimates related to the effects of FTD changes and timely delivered short sales on each of the market-quality metrics. We present estimated coefficients and related statistical significance in Table 5, Panels A and B. First, and most importantly, irrespective of the measure used as proxy for pricing efficiency or liquidity, increases in FTDs lead to higher liquidity and improve pricing efficiency. In particular: 1. Increases in FTDs significantly reduce positive pricing errors, consistent with those traders acting as value arbitrageurs when stocks are overpriced (statistically significant at 1%). 2. Increases in FTDs significantly reduce absolute pricing errors, consistent with an increase in pricing efficiency (statistically significant at 1%). 3. Increases in FTDs significantly reduce intraday volatility, consistent with improved market stability (significant at 1% in two models, at 5% in two models, at 10% in one model, and not significant in the sixth). 4. Increases in FTDs significantly reduce order imbalances, consistent with an improvement in liquidity (significant at 1% for the overall sample, but not for the sample of securities with most fails). 5. Increases in FTDs reduce spreads and thus lower trading costs (significant at 1% in five models and at 5% in the sixth). 6. The impact of changes in FTDs on prices is not significant overall.
Second, we find that the direction and significance of all results relating to the impact of FTD changes is qualitatively
similar to the direction and significance of all results relating to the impact of delivered short sales. The main differences lie in the fact that delivered short sales appear to have a statistically significant impact on prices, while the results are not statistically significant for FTD changes. Our main analysis does not include any measure of trading volume, as trading volume is highly correlated with other market-quality metrics. Nevertheless, for robustness, we add to the VAR models a measure of trading volume based on the daily number of trades for the security of interest. The core results describing the relationship between FTD changes and metrics of market liquidity and pricing efficiency are unaffected, while we find that an increase in OFR is related to lower subsequent trading volume. We also obtain similar results by utilizing a different metric of trading volume based on the monetary value of shares traded daily. The results are not included in tables for brevity. 3.5. FTD changes vs. timely delivered short sales Overall, when comparing the impact of FTD changes with the impact of delivered short sales, the direction and significance of the results are virtually identical, with the exception of the impact on stock prices. That said, the coefficients measuring the impact on market quality in Table 5, Panels A and B are, in most cases, larger in magnitude for delivered short sales relative to FTD changes. This is arguably what we should expect given that the vector autoregressive models we employ are based on standardized variables, and the average standard deviation of delivered short sales is much greater than that of FTD changes, as reported in Table 2. Hence, a one standard deviation change in delivered short sales predictably results in a much greater volume of trades relative to a one-standard deviation change in FTDs and should arguably have a greater impact on market quality. In this context, we next examine the economic significance of the market-quality impact of an equal size (in terms of the number of shares outstanding) of FTD changes and delivered short sales. The results are reported in Table 5, Panels C and D for the impact of FTD changes and delivered short sales, respectively. Based on the Model 1 estimates for the “entire market” sample, we find that an increase in FTDs corresponding to 10 basis points of the total number of outstanding shares— approximately equal to a one-standard deviation increase in total short interest—leads to a 3.2% reduction in the magnitude of positive pricing errors; a 3.2% reduction in absolute pricing errors; a 0.2% point decline in intraday volatility; a 1.7% reduction in spreads; and a 0.7% reduction in order imbalance. The impact on prices, aside from not being statistically significant, is tiny, with a reduction of less than 0.01%. In comparison, again based on the “entire market” sample, we find that delivered short sales corresponding to 10 basis points of the total number of outstanding shares leads to about a 4.5% reduction in the magnitude of positive pricing error; a 6.8% reduction in absolute pricing error; a 0.6% decline in intraday volatility; a 0.6% reduction in spreads; a 1.2% reduction in order imbalance. The impact on prices is again small, with a reduction of approximately 0.01%. We test for the statistical significant of differences between the coefficients estimated for OFR (Panel C) and ODR (Panel D)
V. Fotak et al. / Journal of Financial Economics 114 (2014) 493–516
and find that those are, in all cases, not statistically significant.10 Overall, the economic significance of the estimated impact of FTD changes and delivered short sales is of roughly similar magnitude. Inferences based on the sample with the highest FTD changes are weaker (the same quantum of FTD changes leads to smaller market-wide reactions). This suggests that a further increase in FTDs has a weaker impact when existing levels of outstanding FTDs are higher. The bottom-line is that the market-quality impact of FTD changes is economically significant and very similar to that of timely delivered short sales. Both types of trades appear to have a positive effect on market quality, first by enhancing pricing efficiency through correction of security overvaluation and reduction of intraday volatility, and second by providing and improving liquidity through a reduction of order imbalances and spreads.
3.6. Impulse response functions To further investigate the impact of FTD changes on pricing efficiency and liquidity, we compute impulse response functions based on the vector autoregressive models of Table 5. We estimate accumulated impulse response function parameters for each security and then present cross-sectional means of the parameter estimates. We construct confidence intervals using cross-sectional estimates of standard errors. We present the results related to accumulated impulse response functions depicting the impact of a one-standard deviation increase in OFR on pricing error volatility, intraday volatility, spreads, and order imbalances in Fig. 1. We opt to utilize accumulated rather than orthogonalized impulse response functions because, consistent with econometric literature in this regard, the latter are extremely sensitive to the ordering of variables in the model, thus adding a level of arbitrariness to the estimation procedure. For comparison, we present a similar set of impulse response functions for the impact of delivered shorts in Fig. 2. As the impulse response functions depict, the impact of FTD changes on market-quality metrics documented in Table 5 occur over the following day. On subsequent days, accumulated responses are mostly flat, indicating no further correction or reversal. Given the very close similarities in the estimated VAR coefficients, we only report impulse response functions related to the VAR models with pricing error volatility, and omit the other models including pricing errors and prices. We observe very similar patterns for changes in ODR, thus providing additional evidence that delivered short sales have a similar impact on market quality.11
10 We do not separately report all the estimated t-statistics related to the tests for significance because all tests uniformly fail to reject the null hypothesis that the mean difference between estimated coefficients is zero. 11 Given the different magnitude of the estimated coefficients, which reflect the scaling issues described in Section 3.5, the scales used in the vertical axes in the plots in Figs. 1 and 2 differ. Such unavoidable differences in scaling render a visual comparison of the magnitude of the estimated responses and of the related confidence intervals difficult: while the confidence intervals associated with OFR appear wider than those associated with ODR, they are, in reality, narrower.
505
3.7. Robustness check: threshold-list securities The impact of FTD changes could be different when failures are persistent or long-lived. Accordingly, we separately focus on a sample of securities which displays persistent FTDs—securities that appear on the NYSE Threshold List. In accordance to the SEC Regulation SHO, “threshold securities” are securities for which the aggregate FTD position for five consecutive settlement days at a registered clearing house totals 10,000 shares or more and equals at least 0.5% of the total number of shares outstanding for the issuer. From our sample, we select all securities which appear on the NYSE Threshold List for at least one day during the period spanning January 1, 2005 to June 30, 2008: we identify 175 such securities. Accordingly, we re-estimate the VAR model for this subset of securities with the addition of a binary variable, equal to one for each day on which a security is included on the threshold list, as a predictor in each equation. Results, reported in Table 5, Panel E, are largely consistent with what was previously reported: coefficient estimates are of similar magnitude to those obtained with the data subset containing the securities with the highest level of FTDs. While we do not report the coefficients on the “threshold” binary variable in tables for brevity, we note that inclusion on the threshold list is associated with higher spreads, higher intraday volatility, and higher prices.12 Overall, these results are consistent with inclusion on the threshold list hampering liquidity (as, under regulation SHO, shorting without pre-borrowing of securities on the threshold list is curtailed), while leading to higher prices, perhaps due to the resulting buying pressure of traders forced to close their non-delivered positions.
3.8. Robustness check: Nasdaq securities As a further robustness test, we re-estimate the VAR model with a sample of securities with Nasdaq as a primary exchange. We apply the same filters in creating the sample as we did for our overall, NYSE-based sample: we focus on securities with share codes 10 and 11 in CRSP, for which we have at least 18 months of continuous trading data and with no significant changes in the number of shares outstanding; we obtain a full sample of 2,983 securities. Results are presented in Table 5, Panel F. The total number of securities for which we are able to estimate the models presented in Table 5, Panel F ranges from 2,983 to 2,223. Estimation is not possible for some securities because of inadequate number of positive pricing errors, or negative pricing errors, or order imbalances. In this data subset, similar to our NYSE results, we find that increases in FTDs lead to a next-day decline in positive pricing errors and absolute pricing errors, intraday volatility, and in spreads, but we also find an increase in prices. We also find a decline in order imbalances, but results are statistically significant in only one of three models. However, in this context, it is 12 This test is based on a time-series analysis and, thus, should be interpreted as indicative of changes to market quality affecting a specific security at threshold-list inclusion.
506
V. Fotak et al. / Journal of Financial Economics 114 (2014) 493–516
Fig. 1. Plots of accumulated impulse response functions depicting the impact of a one-standard deviation shock in the Outstanding fails-to-deliver ratio (OFR). The VAR model used for estimation is Model 2 and the sample is the “entire market,” as described in Table 5. The response variables in the four panels are, respectively, Absolute pricing error, Intraday volatility, Spread, and Order imbalance, as defined in Table 1. The horizontal axes are time periods (days), following the initial shock (day 0). All variables are standardized and winsorized by security. Impulse response coefficients are estimated by security; mean values are depicted. 5% confidence intervals are computed using cross-sectional standard error estimates. Panel A: Absolute PE, Panel B: Intraday volatility, Panel C: Spread and Panel D: Order imbalance.
relevant to recognize data limitations, particularly liquidity metrics, for Nasdaq securities.13 Most importantly in the context of this paper, while NYSE has played the dominant role in our sample period in the supply of liquidity for stocks with NYSE as their primary exchange, this has not been so for stocks with Nasdaq as their primary exchange. For example, Diether, Lee, and Werner (2009) document that, while NYSE accounted for almost 80% of shares sold short in NYSE stocks, Nasdaq accounted for less than 35% of the shares sold short in Nasdaq securities. Hence, our estimated order imbalances can potentially be a noisy proxy of the overall order imbalances of Nasdaq stocks.
13 The fragmentation of trading across several different platforms and systems on Nasdaq has led in the past to difficulty in observing and reacting to short-term changes in market variables. For example, quote adjustment has been documented to be slower (Jones and Lipson, 1999) and time stamps of trades and quotes on Nasdaq could not always be matched (Schultz, 2000). More recently, Pastor and Stambaugh (2003) and Watanabe and Watanabe (2008) do not use Nasdaq data in their pricing of liquidity studies apparently because of problems in using Nasdaq-based liquidity measures.
3.9. FTD changes: market-makers vs. public traders The SEC has regarded the use of FTDs by market-makers as bona fide and positive for market liquidity. SEC Report “Key Points about Regulation SHO,” April 11, 2005, updated 2008 says: “because it may take a market maker considerable time to purchase or arrange to borrow the security, a market maker engaged in bona fide market making, particularly in a fast-moving market, may need to sell the security short without having arranged to borrow shares.” In a similar vein and context, SEC Report No. 450, March 18, 2009, states that the SEC “has repeatedly stressed the fact that the practice can provide needed market liquidity in certain circumstances”. It is therefore conceivable that the beneficial effects on market quality of FTDs arise due to such market-making activities, and that there are no beneficial effects from FTDs of public traders. We define the term “public trader” in this context to mean all market participants other than registered market-makers. Notably, it is these public traders that have been the primary focus of negative media and regulatory attention. If there are no beneficial effects of public-trader FTDs, the estimated FTD-related improvement in market quality will be relatively stronger for securities that have a relatively greater proportion of market-maker FTDs. To gain insights into the impact of
V. Fotak et al. / Journal of Financial Economics 114 (2014) 493–516
507
Fig. 2. Plots of accumulated impulse response functions depicting the impact of a one-standard deviation shock in the Outstanding delivered shorts ratio (ODR). The VAR model used for estimation is Model 2 and the sample is the “entire market,” as described in Table 5. The response variables in the four panels are, respectively, Absolute Pricing error, Intraday volatility, Spread, and Order imbalance, as defined in Table 1. The horizontal axes are time periods (days), following the initial shock (day 0). All variables are standardized and winsorized by security. Impulse response coefficients are estimated by security; mean values are depicted. 5% confidence intervals are computed using cross-sectional standard error estimates. Panel A: Absolute PE, Panel B: Intraday volatility, Panel C: Spread and Panel D: Order imbalance.
public-trader FTDs relative to the impact of market-maker FTDs, we classify securities based on the proportion of FTDs originating from market-makers. In order to accomplish that, we employ an exogenous event—an SEC ban in September 2008 that affected only public-trader FTDs but not market-maker FTDs: SEC Temporary Rule 204T (later made permanent). SEC Rule 204T required all market participants other than registered market-makers to purchase or borrow securities to close out their FTD position by the beginning of day tþ4, effectively banning FTDs for anyone other than a market-maker. Hence, we compute average FTDs across our entire sample of NYSE securities for two periods: pre204T from January 1, 2005 to June 30, 2008; and post-204T from January 1, 2009 to December 31, 2010. We intentionally omit the period July to October 2008 because of multiple short selling rule changes, and omit November and December 2008 to allow outstanding FTD positions to fully clear and thereby enable accurately gauging the impact of post-204T FTDs. Consistent with a 2009 s report, we find a drop in average OFR of approximately 71% across all securities, suggesting that, on average, about 29% of all FTDs were due to market-makers prior to September 2008.
For each sample security, we use the proportional change in mean OFR between the two subperiods as our proxy for FTDs by public traders. Our underlying assumption here is that the number of market-maker FTDs is fairly constant over time.14 We accordingly rank securities on the basis of this proportional decline and allocate securities to terciles. Table 6 presents results for the two terciles corresponding to “high proportion of estimated marketmaker trades” (i.e., securities with the lowest decline in mean OFR) and “low proportion of estimated marketmaker trades” (i.e., securities with the highest decline in mean OFR). For the “low proportion of estimated market-maker trades” subsample, despite the sample size being much smaller than the overall sample, all previously documented results remain strong and statistically significant, with the exclusion of the coefficient estimates associated with order imbalance. For the “high proportion of estimated 14 A brief example can be useful in clarifying our methodology: if, for a given security, average OFR in the pre-204T period is 1% and OFR in the post-204T period is 0.25%, then we estimate that 75% of FTDs are due to public traders and 25% are due to market-makers.
508
V. Fotak et al. / Journal of Financial Economics 114 (2014) 493–516
Table 6 Impact of FTDs for securities with high/low proportion of FTDs by market-makers. This table provides results for the estimation of three vector autoregressive models of order one: VAR with Pricing error (Model 1), VAR with Pricing error volatility (Model 2), and VAR with Price (Model 3), as described in Table 5. All variables are standardized and winsorized by security, and are as defined in Table 1. The sample includes all NYSE common stocks of US-based firms (CRSP share codes 10 and 11), listed in the CRSP and TAQ databases for at least 18 months over the interval January 1, 2005 to June 30, 2008, and with no large changes ( 410%) in the number of shares outstanding. The resulting sample includes 1,492 securities. Mean OFR is computed for all securities included in the sample for the period preceding the introduction of SEC Rule 204T (January 1, 2005 to June 30, 2008) and for a period following (January 1, 2009 to December 31, 2010). Securities are ranked according to the proportional change in mean OFR. Securities are then allocated to three quantiles—those with the highest decline in mean OFR are allocated to the “Low proportion of estimated market-maker trades” sample, while those with the lowest decline in mean OFR are allocated to the “High proportion of estimated market marker trades” sample. Reported parameter estimates are averages of parameters estimated by security. Significance is tested employing a cross-sectional estimate of the standard error of the parameter estimate. The p-values from two-sided t-tests are in italics below the parameter estimates. “n”, “nn”, and “nnn” indicate significance at the 10%, 5%, and 1% level, respectively. Low/high proportion of estimated market-maker trades A: Low 340 Securities B: High 329 Securities A: Low 340 Securities B: High 340 Securities A: Low 340 Securities B: High 340 Securities
ΔPE
ΔPE (incremental effect when lag PE40)
0.0216 o 0.01nnn 0.0094 0.30
0.0287 o0.01nnn 0.0033 0.74
market-maker trades” subsample, we observe statistically significant parameter estimates for the impact on spreads, but not for the other variables. Our overall results indicate that, if anything, the impact of FTD changes is stronger for the subset of securities with a relatively low involvement of market-makers. Hence, at the very least, our results are inconsistent with the hypothesis that the beneficial effects of FTD changes arise entirely from market-makers. Our intuition is that, while market-maker trades contribute to an improvement in liquidity (and, hence, lead to lower spreads), they do not lead to an improvement in pricing efficiency—which is consistent with market-makers not engaging in price arbitrage. 3.10. The market-quality impact of restrictions on FTDs As discussed in the introduction, FTDs can plausibly originate from both short and long sales. In this subsection, we use a natural experiment based on an exogenous regulatory change to focus on FTDs most likely arising from short sales. In particular, we look at the imposition on July 15, 2008 by the SEC of an Emergency Order selectively mandating borrowing requirements prior to short sales and applicable only to the stocks of a select group of 19 publicly traded financial institutions from July 21 to August 12, 2008. The SEC order required that “no person may effect a short sale in these securities… unless such person or its agent has borrowed or arranged to borrow the security or otherwise has the security available to borrow in its inventory prior to effecting such short sale.” (SEC Release 58166, 2008). The only regulatory change affecting these stocks during this period was this order mandating firm
ΔAbs PE
ΔPrice
ΔIntraday volatility
ΔSpread
ΔOIB
0.0018 0.09n 0.0007 0.52
0.0085 0.03nn 0.0003 0.95 0.0114 o0.01nnn 0.0070 0.15 0.0090 0.02nn 0.0050 0.30
0.0152 o 0.01nnn 0.0105 o 0.01nnn 0.0142 o 0.01nnn 0.0119 o 0.01nnn 0.0147 o 0.01nnn 0.0127 o 0.01nnn
0.0082 0.15 0.0108 0.05nn 0.0079 0.16 0.0104 0.11 0.0074 0.21 0.0099 0.13
0.0169 o 0.01nnn 0.0036 0.47
stock-borrowing arrangements prior to executing short sales. Any FTDs arising from “long” sales should clearly not be affected by this order. The overall effect of the SEC order in relation to FTDs should therefore be driven by FTDs arising from short sales. We employ an event study methodology to investigate the variation in FTDs and in market quality around this mandate requiring stockborrowing arrangements prior to short sales. Data for 17 of the 19 securities (the “event sample”) affected by the order are available from CRSP. We construct a matched control sample: for each of our 17 affected securities, we identify the firm that is not affected by the order, but shares the same four-digit Standard Industrial Classification (SIC) code and has the closest market capitalization (least absolute deviation) as of January 1, 2008. Overall, the securities affected by the ban have a higher mean and median market capitalization, but the differences are not statistically significant. Then, we analyze the number of reported FTDs (duly scaled by the number of shares outstanding) and relevant market-quality variables for both event and control samples, over the following periods: (1) a Pre-order period from January 1 to July 14, 2008 ending on the day prior to the announcement of the Emergency Order; (2) the Order period—July 21 to August 12, 2008—for which the Emergency order was in force; and (3) a Post-order period from August 13 to September 9, 2008 following the order and ending before Lehman's bankruptcy. We do not include the four trading days on July 15, 16, 17, and 18, 2008, in the Pre-order period to avoid any potentially confounding effects arising from the announcement of the order on July 15, 2008. We report our results on FTDs in Table 7. Reported FTDs (as a proportion of shares outstanding) for the 17 securities
V. Fotak et al. / Journal of Financial Economics 114 (2014) 493–516
509
Table 7 FTDs and short sales. This table presents the average number of fails-to-deliver (FTDs) scaled by the contemporaneous number of shares outstanding for “Event sample” and “Control sample.” Securities in the “Event sample” are 17 stocks that were subject to the imposition of the SEC Order requiring stock-borrowing arrangements prior to short sales between July 21, 2008 and August 12, 2008, i.e., the “Order period.” While the order affected 19 securities, data are available for 17 of these: Bank of America Corporation, Barclays, Bear Stearns Companies Inc., Citigroup Inc., Credit Suisse Group, Deutsche Bank Group AG, Allianz SE, Goldman Sachs Group Inc., Royal Bank ADS, HSBC Holdings PLC ADS, J.P. Morgan Chase & Co., Merrill Lynch & Co., Inc., Mizuho Financial Group, Inc., Morgan Stanley, UBS AG, Freddie Mac, and Fannie Mae. “Control sample” securities are 17 market capitalization and industry-matched stocks that were not subject to any relevant regulatory changes during the “Order period.” The “Pre-order period” refers to the interval January 1, 2008 to July 14, 2008. The “Post-order period” refers to the interval August 13, 2008 to September 9, 2008. The number of FTDs scaled by shares outstanding is reported in basis points and computed on a daily basis over the interval January 1, 2008 to August 12, 2008. Reported p-values are for a test of differences in mean FTDs (scaled by shares outstanding) between the indicated order subperiods and the “Pre-order period” and between “Event sample” and “Control sample”. “n”, “nn”, and “nnn” indicate significance at the 10%, 5%, and 1% level, respectively. Event companies Period
Mean (bps)
Pre-order Order, 1st week (Order, 1st week) – Pre-order Order, 2nd week (Order, 2nd week) – Pre-order Order, 3rd week (Order, 3rd week) – Pre-order Order, Last day (Order, Last day) – Pre-order Post-order (Post-order) – (Order, 3rd week)
4.78 0.57 4.22 0.17 4.61 0.12 4.67 0.00 4.78 1.12 1.01
Control companies
p-Value
o 0.01nnn o 0.01nnn o 0.01nnn nnn
o 0.01
0.03nn
in our sample average 4.78 basis points (“bp”) during the pre-order period. After the order, FTDs reduce rapidly to about 0.57 bp, or 12% of their pre-order average within one week; to about 0.12 bp, or 2% of the pre-order average, within three weeks; and, on the last day of the order period, reported FTDs are below the reporting threshold of 10,000 shares for each and every affected security. In contrast, reported FTDs for control sample securities increase during the first week of the ban, slightly decrease over the second week of the ban, and show no significant change during the third week of the ban. As soon as the temporary order period expires, reported FTDs of affected securities again increase monotonically and significantly. The difference in reported FTDs between event and control firms is positive and significant prior to the mandate period, negative and significant during the mandate period, and again positive and significant after the mandate is lifted. Given that reported FTDs cannot fall immediately to zero because of the time required for fails outstanding from prior to the SEC order to clear (Boni, 2006), this natural experiment suggests that, at least for these particular securities and around this specific period, a very large portion of FTDs originated from short sales.15 We report our results on market quality in Table 8. To test the impact of the order on market quality, we analyze Absolute pricing error, Return, Intraday volatility, Spread, and Order imbalance over each of our three subperiods. We average each of these five variables across securities to obtain a daily mean for the affected and control samples. We compare differences in means 15 For our larger NYSE sample, we also find that new FTDs on day tþ 3 are significantly (p-value⪡0.01) and positively related only to the daily trading volume arising from short sales on day t, and not the daily trading volume arising from “long” sales on day t.
Mean (bps) 0.47 2.79 2.32 0.46 0.01 0.48 0.01 0.23 0.24 0.18 0.31
p-Value
o0.01nnn 0.72 0.21 o0.01nnn 0.51
Event-Control Mean (bps)
p-Value
4.31 2.23 6.54 0.29 4.60 0.36 4.68 0.23 4.55 0.95 1.31
o 0.01nnn o 0.01nnn o 0.01nnn 0.53 o 0.01nnn 0.43 o 0.01nnn 0.61 o 0.01nnn 0.04 o 0.01
between the order-period (July 21 to August 12, 2008) and the “pre- and post-order” (January 1 to July 14 and August 13 to September 9) periods for both sample and control securities. Our results are presented in Table 8. Comparing securities during the benchmark “pre- and post-order” period, we note that “event sample” securities display larger pricing errors, lower returns, higher intraday volatility, and lower spreads than “control sample” securities. While these results are indicative of the difficulty of obtaining perfect matches, we mitigate the impact of these differences by employing a difference-in-difference methodology: we compute the difference between the “order” and the “pre- and post-order” periods for both “event sample” and “control sample” securities and then compute the difference between those differences. We find that, while both samples display higher absolute pricing errors, intraday volatility, and spreads during the “order” period, the increase for the “event sample” securities is larger and the difference-in-difference tests are all statistically significant—at the 1% level for absolute pricing errors and for intraday volatility and at the 10% level for spreads. The impact of the SEC order on returns and order imbalances is not statistically significant. In robustness tests (that are not reported for brevity), we replicate a similar analysis excluding the period subsequent to the order from the benchmark, and obtain virtually identical results. Boulton and Braga-Alves (2010) also document a negative impact of the July 2008 Emergency Order on liquidity, i.e., higher bid-ask spreads and lower trading volume relative to a matched sample of financial firms. However, an SEC Office of Economic Analysis (OEA) memorandum dated January 14, 2009 titled “Analysis of the July Emergency Order Requiring a Pre-Borrow on Short Sales” examines selected market-quality variables around the July 2008 Emergency Order, but is unable to conclude statistically
510
V. Fotak et al. / Journal of Financial Economics 114 (2014) 493–516
Table 8 The impact of restrictions on FTDs imposed by the SEC between July 21, 2008 and August 12, 2008. The table presents means of Absolute pricing error, Return, Intraday volatility, Spread, and Order imbalance, as defined in Table 1. The SEC mandated preborrowing prior to short selling 19 securities during the period spanning July 21 to August 12, 2008 (the “Order” period). The Order affected 19 securities, but data are available for 17 of these (the “Event sample”): Bank of America Corporation, Barclays, Bear Stearns Companies Inc., Citigroup Inc., Credit Suisse Group, Deutsche Bank Group AG, Allianz SE, Goldman Sachs Group Inc., Royal Bank ADS, HSBC Holdings PLC ADS, J. P. Morgan Chase & Co., Merrill Lynch & Co., Inc., Mizuho Financial Group, Inc., Morgan Stanley, UBS AG, Freddie Mac, and Fannie Mae. Securities in the “Control sample” are market capitalization and industry-matched stocks that were not subject to any relevant regulatory changes during the “Order” period. Means during the “Order” period are compared to variable means computed over a period surrounding the “Order,” the “Pre- and Post-order” periods, as indicated in the table. The p-values from two-sided t-tests for the significance of mean differences are in italics below the parameter estimates. “n”, “nn”, and “nnn” indicate significance at the 10%, 5%, and 1% level, respectively.
Abs PE
Pre- and post-order (Jan 1–Jul 14 and Aug 13–Sep 9, 2008)
Order (Jul 21–Aug 12, 2008)
Difference (Order – Pre- and postorder)
20.47%
28.08%
1.34%
1.48%
7.61% o0.01nnn 0.14%
19.13% o 0.01nnn
26.60% o0.01nnn
0.24 7.47% o0.01nnn
0.37%
0.27%
Control sample
0.14%
0.46%
0.10% 0.93 0.59%
Difference (EventControl) Event sample
0.24% 0.06n
0.73% 0.07n
0.49 0.49% 0.26
0.33%
0.44%
0.11%
0.29%
0.37%
o0.01nnn 0.07%
0.03% o 0.01nnn
0.08% o0.01nnn
o0.01nnn 0.04% o0.01nnn
0.10%
0.15%
0.17%
0.21%
0.05% o0.01nnn 0.04%
0.07% o 0.01nnn
0.06% o0.01nnn
o0.01nnn 0.01% 0.07n
1.78%
0.16%
Control sample
0.18%
0.34%
1.94% 0.33 0.53%
Difference (EventControl)
1.59% o 0.01
0.18% 0.90
0.71 1.41% 0.37
Event sample Control sample
Return
Intraday volatility
Difference (EventControl) Event sample
Control sample
Spread
Difference (EventControl) Event sample Control sample
OIB
Difference (EventControl) Event sample
significant differences between their “event” and “control” samples for any of the market-quality variables they examine. There can be several reasons for the difference in our conclusions. First, except for returns (for which our conclusions are the same), the OEA do not examine the variables that we analyze all through this paper, or their variable is defined in a materially different way: (1) they do not examine absolute pricing errors and order imbalances; (2) as against our intraday volatility measure (computed daily over five-minute intervals), they examine daily closeto-close or open-to-close volatility (over several weeks); and (3) as against our relative spreads (i.e., dollar spreads
scaled by price), they examine dollar spread (which also means that their averages implicitly give higher weight to higher priced securities). Second, while we use a control sample of financial firms that are as closely matched to the affected securities as possible, the OEA uses two control samples—a sample of all financial firms and a sample of all nonfinancial firms—none of which are specifically matched with the affected securities. Third, the OEA uses a benchmark period prior to the Order that is constructed by randomly selecting only 20 trading days over the previous nine months. The small number of days and the randomness may arguably increase the standard errors of their
V. Fotak et al. / Journal of Financial Economics 114 (2014) 493–516
estimates, besides making it impossible for us to replicate their results or more fully explain any differences. The longer benchmark period we utilize, our closely matched construction of the control sample, and the higher frequency of some of our variables, are all reasons why our tests are likely to have greater power, and hence potentially accounts for the fact that we detect significance in tests in which the OEA may not. Overall, our results indicate that mandating stockborrowing arrangements prior to executing a short sale leads to a sharp drop in overall FTDs to almost zero (as it would if FTDs arose largely from short sales over the sample period), but at the same time, there is also a significant adverse impact on the stock's liquidity and its price discovery process. Banning FTDs should also arguably lead to an upward pressure on stock-lending rates as short sellers scramble to make stock-borrowing arrangements at short notice, and thereby also concomitantly increase costs for all short sellers. Kolasinski, Reed, and Thornock (2013) and the OEA memo cited above document that this is indeed the case with rebate rates for order-affected securities becoming negative on the first day of order-validity and remaining negative for several days thereafter. While our results clearly show that restricting FTDs reduces market quality, this could result directly from the absence of FTDs, or because informed short selling decreases due to consequential increase in stock-lending rates, or both. In
511
this context, we examine short interest of our event and control sample securities over our Pre-order period and Order-period, and find, consistent with similar findings of the OEA memo cited above, that the short interest of event sample securities over the Order-period is not significantly different from the short interest over the Pre-order period: the mean short interest for event firms reported on July 15 is 5.07% (of shares outstanding) and, by the end of the order period, on August 15, it is 5.45%; the slight increase is not statistically significant and neither is it if we measure it relative to the control sample. Given that FTDs arising from short sales decrease sharply during the ban while the short interest does not change significantly, we conjecture that the adverse market-quality impact is likely to be related at least partially to the absence of failing-to-deliver short sales. Irrespective, our results show that restricting FTDs reduces market quality. That said, our results in this subsection should be interpreted or generalized with caution. First, this Emergency Order affected only a few securities, all from one industry, and during a fairly unique period of time. Second, while Kolasinski, Reed, and Thornock (2013) note that it is sensible to analyze the impact of a short selling regulation during a crisis, as that is exactly when short selling regulation is most likely to be operational, the OEA memo cited above suggested a “Hawthorne effect,” whereby “firms could have been on their best behavior during the
Fig. 3. Plot of Outstanding fails-to-deliver ratio and Return index related to Bear Stearns Companies Inc., common stock (ticker: BSC) against calendar date. The Return index is set to 1 on January 1, 2008; Return indext ¼ Σ ij ¼ 1 ð1þ RBSC;j Þ RBSC;j is the observed total return for BSC on day j, from the CRSP database.
512
V. Fotak et al. / Journal of Financial Economics 114 (2014) 493–516
Table 9 Outstanding fails-to-deliver ratio (OFR) of Bear Stearns (ticker: BSC) and of Lehman Brothers Holdings Inc. (ticker: LEH) in 2008. The Outstanding fails-to-deliver ratio (OFR) is computed as the ratio of outstanding fails-to-deliver and shares outstanding. Index OFR is calculated as the equal-weighted average of OFR of common stocks of four firms that are matched on primary SIC and market capitalization as of the end of the fiscal year 2007 to BSC in Panel A and to LEH in Panel B. Tests for significance and related t-statistics are computed using the standard error of the OFR difference over the time interval January 1, 2008 to February 15, 2008 for BSC and over the time interval January 1, 2008 and ending 20 trading days prior to September 9, 2008 for LEH; p-values are two-sided. “n”, “nn”, and “nnn” indicate significance at the 10%, 5%, and 1% level, respectively. Panel A: Bear Stearns Date 03/03/08 03/04/08 03/05/08 03/06/08 03/07/08 03/10/08 03/11/08 03/12/08 03/13/08 03/14/08 03/17/08 03/18/08 03/19/08 03/20/08 03/24/08 03/25/08 03/26/08 03/27/08 03/28/08 03/31/08
Panel B: Lehman Brothers Holdings Inc.
BSC stock price (USD)
BSC OFR
Index OFR
Difference in OFR
p-Value
79.86 77.32 77.17 75.80 69.90 70.08 62.30 62.97 61.58 57.00 30.85 2.00 5.91 5.26 6.39 11.25 10.94 11.21 11.23 10.78
0.0580% 0.0069% 0.0154% 0.0634% 0.0574% 0.0082% 0.0854% 0.5284% 0.3175% 0.8979% 5.8383% 5.3300% 4.9694% 3.2489% 3.9550% 5.2488% 1.9806% 2.3933% 2.7779% 1.1950%
0.0003% 0.0000% 0.0062% 0.0008% 0.0017% 0.0001% 0.0010% 0.0249% 0.0041% 0.0155% 0.0174% 0.0004% 0.0031% 0.0180% 0.0030% 0.0240% 0.0076% 0.0110% 0.0000% 0.0000%
0.0577% 0.0069% 0.0092% 0.0626% 0.0558% 0.0081% 0.0845% 0.5035% 0.3134% 0.8824% 5.8209% 5.3296% 4.9663% 3.2310% 3.9520% 5.2248% 1.9730% 2.3823% 2.7779% 1.1950%
0.52 o 0.01nnn o 0.01nnn 0.66 0.47 o 0.01nnn 0.64 o 0.01nnn o 0.01nnn o 0.01nnn o 0.01nnn o 0.01nnn o 0.01nnn o 0.01nnn o 0.01nnn o 0.01nnn o 0.01nnn o 0.01nnn o 0.01nnn o 0.01nnn
time of the Emergency Order because of the likelihood of intense scrutiny by regulators of failed positions”. 4. Empirical results: FTDs and the 2008 financial crisis In the wake of the 2008 financial crisis, it has been widely alleged that short sellers have, by failing to deliver, caused or accelerated sharp declines in stock prices, particularly of financial firms, in order to profit from eventual collapse of the financial institutions involved. For example, in relation to Bear Stearns, former Under Secretary of Commerce Robert J. Shapiro has been quoted as saying that “… the pace of the collapse of the stock price was clearly accelerated by the enormous naked short sale activity” (Euromoney, December 2008). Similarly, Richard Fuld, former CEO of Lehman Brothers, in his October 2008 testimony before the US House of Representatives Committee on Oversight and Reform, stated that “naked short selling […] dealt a critical, if not fatal, blow to Bear Stearns” and “contributed significantly” to the collapse of Lehman Brothers. In this context, we examine whether significant increases in FTDs preceded (and hence potentially triggered) the price crashes associated with the four major casualties of the 2008 financial crisis, i.e., Bear Stearns, Lehman, AIG, and Merrill, or did significant increases in FTDs take place after these price crashes and in response to negative public news. The first large financial casualty of the 2008 financial crisis was Bear Stearns, the fifth largest investment banking firm in the nation at the time of its demise. We analyze OFR on the days preceding and immediately following the
Date 08/25/08 08/26/08 08/27/08 08/28/08 08/29/08 09/02/08 09/03/08 09/04/08 09/05/08 09/08/08 09/09/08 09/10/08 09/11/08 09/12/08 09/15/08 09/16/08 09/17/08
LEH stock price (USD)
LEH OFR
Index OFR
Difference in OFR
p-Value
13.45 14.03 14.78 15.87 16.09 16.13 16.94 15.17 16.20 14.15 7.79 7.25 4.22 3.65 0.21 0.30 0.13
0.0010% 0.0017% 0.0000% 0.0070% 0.0008% 0.0045% 0.0038% 0.0009% 0.0000% 0.0023% 0.0786% 0.4232% 1.6291% 2.3673% 2.0212% 2.1495% 3.6041%
0.0001% 0.0000% 0.0016% 0.0027% 0.0016% 0.0015% 0.0000% 0.0038% 0.0095% 0.0005% 0.0055% 0.0084% 0.0161% 0.0167% 0.0202% 0.0777% 0.0797%
0.0009% 0.0017% 0.0016% 0.0043% 0.0008% 0.0030% 0.0038% 0.0029% 0.0095% 0.0018% 0.0731% 0.4148% 1.6130% 2.3506% 2.0010% 2.0718% 3.5244%
0.89 0.79 0.79 0.50 0.90 0.63 0.55 0.64 0.13 0.78 o 0.01nnn o 0.01nnn o 0.01nnn o 0.01nnn o 0.01nnn o 0.01nnn o 0.01nnn
dramatic loss of market value that led to the demise of the firm. Fig. 3 provides a time-line about the crisis. Outstanding suspicions about liquidity problems at Bear Stearns were reported in the media from March 10 onwards, along with news that the company's management was repeatedly denying rumors about such problems. The first major price-crash took place on Friday, March 14, when the price per share dropped from $57 to $30 after a 9 a.m. announcement that Bear Stearns would receive an unprecedented loan from the Federal Reserve System; and two days later, on Sunday, March 16, JP Morgan Chase proposed buying Bear Stearns for $2 per share (Wall Street Journal, March 15, 2008; J.P. Morgan News Release, March 16, 2008). When markets opened on March 17, a second major price-crash materialized, and the price dropped to a close of $4.81. We compute OFR for Bear Stearns on each trading day from January 1 to March 28, 2008. We also compute, as a control variable, an equal-weighted average OFR for four other financial institutions—Raymond James Financial Corporation, Ameritrade Holdings Corporation, Ameriprise Financial Inc., and Charles Schwab Corporation —with the same primary SIC code as Bear Stearns, and with the closest market value as of the end of the fiscal year 2007. We test for the statistical significance each day of the difference, i.e., the “abnormal” OFR of Bear Stearns. To construct a t-statistic for the difference in means, we compute the standard error of this difference over the period January 1 to February 15, 2008. Results are presented in Table 9, Panel A. Even though negative media attention started on March 10, Table 9, Panel A and Fig. 3 show that, while OFR for Bear
V. Fotak et al. / Journal of Financial Economics 114 (2014) 493–516
513
Fig. 4. (A) Plot of Outstanding fails-to-deliver ratio and Return index related to Lehman Brothers Holdings Inc. (ticker: LEH) against calendar date. The Return index is set to 1 on January 1, 2008; Return indext ¼ Σ ij ¼ 1 ð1 þ RLEH;j Þ. RLEH,j is the observed total return for LEH on day j, from the CRSP database. (B) Detail of plot of Outstanding fails-to-deliver ratio and Return index related to Lehman Brothers Holdings Inc. (ticker: LEH) against calendar date. The Return index is set to 1 on January 1, 2008; Return indext ¼ Σ ij ¼ 1 ð1þ RLEH;j Þ . RLEH,j is the observed total return for LEH on day j, from the CRSP database.
Stearns was greater than for the benchmark securities, its small magnitude (below 0.1%) indicates that fails were unlikely to have a significant economic impact up to March 11. While OFR did increase significantly on March 12 and 13, the increase was still relatively tiny from an economic perspective relative to the total number of shares outstanding, to the typical overall short volume, and to what took place on or after March 14: OFR was only 0.32% of shares outstanding until market close on March 13. OFR increased to 0.90% on March 14, but increased massively
only on March 17, reaching 5.84%. Importantly, given that the Fed announcement was at the start of trading on March 14, even the (relatively small) increase in OFR on March 14 was clearly subsequent to the public release of tangibly negative news in the form of the announcement and the consequent immediate precipitous price-drop. By the time OFR spiked on March 17, the company was already in an open distress sale. The evidence clearly shows that the abnormal incidence of increases in FTDs did not precede the price decline but followed it; and the decline in stock
514
V. Fotak et al. / Journal of Financial Economics 114 (2014) 493–516
price was triggered by other well-identified negative economic news. Consistent with our previous results, short sellers appear to be facilitating price discovery, rather than increasing pricing errors. The second notable casualty of the financial crisis of 2008 was Lehman. In Fig. 4A we present the relationship between OFR and stock price for the period from January 2008 to Lehman's bankruptcy on September 15, 2008. We present a closer look at the period surrounding Lehman's bankruptcy in Fig. 4B. It is clear that OFR remained minuscule in magnitude all through till September 10; and till September 9, the firm's stock price had already lost approximately 87% of its value as of the beginning of the year. There were some spikes in OFR over the period, but these were too small to be economically relevant. For example, OFR increases on June 27 just prior to an 11% drop in share price on the next trading day of June 30, but this spike is just 0.06%, and ostensibly linked, according to an article in Bloomberg (March 19, 2009), to a rumor about Barclays PLC buying Lehman for a 25% discount to market value. The biggest single-day price drop of about 45% occurred on September 9, again following news that talks with the Korea Development Bank (previously rumored to be considering a 25% stake in Lehman) had failed. The OFR on the days immediately surrounding the dramatic price fall on September 9, 2008, is presented in Table 9, Panel B. The table indicates very low OFR, below 0.01%, in the days leading to September 9. While OFR increases on September 9, it still remains less than 0.09% of shares outstanding. Abnormal OFR increased more dramatically only after September 10, well after widespread coverage of negative news about Lehman and the associated price crash. On September 11, as shareholders rejected a management rescue plan, the stock price fell by an additional 42% and OFR further increased to 1.63%. Over the following days, talks of a possible acquisition by Bank of America and Barclays failed, triggering further declines in stock price and an increase in OFR to 2.37%. Lehman announced its
bankruptcy on September 15, and OFR increased beyond 3.6% on September 17. In sum, our analysis shows that the incidence of increases in FTDs occurred only after the large price drops on and after September 9, not before. We report the relationship between OFR and the stock price for Merrill and AIG in Figs. 5 and 6, respectively. In both cases, the stock price declines were fairly gradual through the year and not accompanied by any significant increase in FTDs. For Merrill, OFR reached its highest value of only 0.09% on October 10, 2008, well after Merrill had lost most of its value. AIG had also lost about 40% of its market value by the end of August 2008, and the largest single-day price drop was on September 15, 2008 when Standard & Poor's (S&P) cut AIG's credit rating. Following the announcement, the company's stock price dropped by about 60% and, on the following day, OFR climbed to its highest value up to that point, at 0.13%; OFR reached its highest value only a fortnight later on September 29, 2008, and even this highest value was only 0.14%. Given that OFR remained so low all through the financial crisis period for both firms, we do not engage in any further statistical testing. 5. Conclusions Recent regulatory action has focused on eliminating settlement day delivery failures in equity markets, particularly delivery failures arising from short sales, even though the alternative to fail has been shown to be valuable and important for pricing and liquidity (Evans, Geczy, Musto, and Reed, 2009; Merrick, Naik, and Yadav, 2005); and notwithstanding the extensive evidence on the beneficial effect of short sales on market quality (Diamond and Verrecchia, 1987; Abreu and Brunnermeier, 2002, 2003; Miller, 1977; Bris, Goetzmann, and Zhu, 2007; Diether, Lee, and Werner, 2009; Boehmer, Jones, and Zhang, 2008). In this context, we investigate the aggregate impact on pricing efficiency and liquidity of changes in FTDs, i. e., the market-quality impact of the set of trades over the days that collectively affect the open interest of undelivered positions.
Fig. 5. Plot of Outstanding fails-to-deliver ratio and Return index related to Merrill Lynch & Co., Inc. (ticker: MER) against calendar date. The Return index is set to 1 on January 1, 2008; Return indext ¼ Σ ij ¼ 1 ð1þ RMER;j Þ. RMER,j is the observed total return for MER on day j, from the CRSP database.
V. Fotak et al. / Journal of Financial Economics 114 (2014) 493–516
515
Fig. 6. Plot of Outstanding fails-to-deliver ratio and Return index related to American International Group (ticker: AIG) against calendar date. The Return index is set to 1 on January 1, 2008; Return indext ¼ Σ ij ¼ 1 ð1 þ RAIG;j Þ . RAIJ,j is the observed total return for AIG on day j, from the CRSP database.
For samples of both NYSE and (in robustness checks) Nasdaq securities, we find that increases in FTDs lead to a reduction in pricing errors, intraday volatility, spreads, and order imbalances. Further, consistent with our expectation that the effect on market quality of a short sale will not depend on whether or not the short sale finally results in timely delivery at settlement, we also find that the positive impact of FTD changes on pricing efficiency and liquidity is similar to the beneficial impact of our estimate of delivered short sales. We then focus specifically on FTDs most likely arising from short sales by examining the effect of a July 2008 temporary SEC order restricting FTDs by mandating stock-borrowing arrangements prior to short sales in select financial securities. We find that, while the SEC order banning FTDs arising from (naked) short sales did lead (as expected) to a drastic reduction in FTDs, it also led to a significant increase in absolute pricing errors, relative bid-ask spreads, and intraday volatility. These findings are consistent with other studies documenting a negative impact on market quality associated with regulatory restrictions on short sales in general, for example, Battalio and Schultz (2011), Boehmer, Jones, and Zhang (2013), Autore, Billingsley, and Kovacs (2011), and Beber and Pagano (2013). In the context of the widespread allegations that FTDs arising from (naked) short sales contributed to the collapse of major financial firms, we find that large abnormal increases in FTDs took place after and not before their major stock price declines and associated negative news; and hence, FTDs were responding to information about the firms, rather than triggering the observed price declines. Overall, our results indicate that, contrary to media and regulatory perceptions about naked short selling, FTDs did not precipitate the collapse of major financial firms in 2008. Our findings have important implications for regulatory policy. While we do not advocate tolerance for FTDs tied to
manipulative episodes, we find no evidence indicating that FTDs, in aggregate, systematically and manipulatively precipitated price declines, even in the extreme situation of the 2008 financial crisis. Instead, the gently regulated FTD regime that existed after Regulation SHO up to mid-2008 was net beneficial for pricing efficiency and market liquidity. In this context, the removal of the ability to fail for the vast majority of market participants is very questionable, since the ability to fail arguably reduces stock-borrowing costs at the time when such costs are the highest, and thereby protects traders from the lack of liquidity sometimes seen in the less regulated stockborrowing market. This ability to fail is valuable for markets also because it can prevent distortions in the stock-borrowing and lending market leading to serious pricing and liquidity distortions in the mainstream stock market. Regulators could alternatively consider progressive fines for settlement delays rather than a blanket removal of the alternative to fail. Importantly, regulation should focus on how best to generate liquidity and transparency in the stock-borrowing market.
References Abreu, D., Brunnermeier, M., 2002. Synchronization risk and delayed arbitrage. Journal of Financial Economics 66, 341–360. Abreu, D., Brunnermeier, M., 2003. Bubbles and crashes. Econometrica 71, 173–204. Autore, D., Billingsley, R., Kovacs, T., 2011. The 2008 short sale ban: liquidity, dispersion of opinion, and the cross-section of returns of US financial stocks. Journal of Banking & Finance 35, 2252–2266. Battalio, R., Schultz, P., 2011. Regulatory uncertainty and market liquidity: the 2008 short sale ban's impact on equity option markets. Journal of Finance 66, 2013–2053. Beber, A., Pagano, M., 2013. Short-selling bans around the world: evidence from the 2007–09 crisis. Journal of Finance 68, 343–381. Boehmer, E., Jones, C., Zhang, X., 2008. Which shorts are informed? Journal of Finance 63, 491–526.
516
V. Fotak et al. / Journal of Financial Economics 114 (2014) 493–516
Boehmer, E., Jones, C., Zhang, X., 2013. Shackling short sellers: the 2008 shorting ban. Review of Financial Studies 26, 1363–1400. Boni, L., 2006. Strategic delivery failures in U.S. equity markets. Journal of Financial Markets 9, 1–26. Boulton, T., Braga-Alves, M., 2010. The skinny on the 2008 naked short sale restrictions. Journal of Financial Markets 13, 397–421. Bris, A., Goetzmann, W., Zhu, N., 2007. Efficiency and the bear: short sales and markets around the world. Journal of Finance 62, 1029–1079. Chen, H., Singal, V., 2003. Role of speculative short sales in price formation: the case of the weekend effect. Journal of Finance 58, 685–705. Culp, C., Heaton, J., 2008. The economics of naked shorting. Regulation 31, 46–51. DeChow, P., Hutton, A., Meulbroek, L., Sloan, R., 2001. Short sellers, fundamental analysis, and stock returns. Journal of Financial Economics 61, 77–106. Dempster, P., Laird, N., Rubin, D., 1977. Maximum likelihood from incomplete data via the EM algorithm. Journal of the Royal Statistical Society B 39, 1–38. Desai, H., Thiagarajan, S.R., Ramesh, K., Balachandran, B., 2002. An investigation of the informational role of short interest in the Nasdaq market. Journal of Finance 57, 2263–2287. Diamond, D., Verrecchia, R., 1987. Constraints on short selling and asset price adjustment to private information. Journal of Financial Economics 18, 277–311. Diether, K., Lee, K., Werner, I., 2009. Short sale strategies and return predictability. Review of Financial Studies 22, 575–607. Edwards, A., Hanley, K., 2010. Short selling in initial public offerings. Journal of Financial Economics 98, 21–39. Evans, R., Geczy, C., Musto, D., Reed, A., 2009. Failure is an option: impediments to short selling and options prices. The Review of Financial Studies 22, 1955–1980. Fama, E., MacBeth, J., 1973. Risk, return, and equilibrium: empirical tests. Journal of Political Economy 81, 607–633. Geczy, C., Musto, D., Reed, A., 2002. Stocks are special too: an analysis of the equity lending market. Journal of Financial Economics 66, 241–269.
Hamilton, J., 1985. Uncovering financial market expectations of inflation. Journal of Political Economy 93, 1224–1241. Hasbrouck, J., 1993. Assessing the quality of a security market: a new approach to transaction-cost measurement. Review of Financial Studies 6, 191–212. Jones, C., Lipson, M., 1999. Price impacts and quote adjustment on the Nasdaq and NYSE/Amex. Unpublished working paper. Columbia Business School and University of Virginia. Kolasinski, A., Reed, A., Ringgenberg, M., 2013. A multiple lender approach to understanding supply and search in the equity lending market. Journal of Finance 68, 559–595. Kolasinski, A., Reed, A., Thornock, J., 2013. Can short restrictions result in more informed short selling? Evidence from the 2008 regulations. Financial Management 42, 155–181. Merrick, J., Naik, N., Yadav, P., 2005. Strategic trading behavior and price distortion in a manipulated market: anatomy of a squeeze. Journal of Financial Economics 77, 171–218. Miller, E., 1977. Risk, uncertainly, and divergence of opinion. Journal of Finance 32, 1151–1168. Morris, V., Goldstein, S., 2009. Guide to Clearance & Settlement. Lightbulb Press, New York. Pastor, L., Stambaugh, R., 2003. Liquidity risk and expected stock returns. Journal of Political Economy 111, 642–685. Putniņš, T., 2010. Naked short sales and fails-to-deliver: an overview of clearing and settlement procedures for stock trades in the USA. Journal of Securities Operations and Custody 2, 340–350. Schultz, P., 2000. Regulatory and legal pressures and the costs of Nasdaq trading. The Review of Financial Studies 13, 917–957. Stokes, A., 2009. In pursuit of the naked short. NYU. Journal of Law and Business 5, 1–61. Watanabe, A., Watanabe, M., 2008. Time-varying liquidity risk and the cross section of stock returns. Review of Financial Studies 21, 2449–2486. Welborn, J., 2008. The ‘phantom shares’ menace. Regulation 31, 51–61.