Time-varying return predictability in South Asian equity markets

Author’s Accepted Manuscript Time-varying Return Predictability in South Asian Equity Markets Md Lutfur Shamsuddin

Rahman,

Doowon

Lee,

Abul

www.elsevier.com

PII: DOI: Reference:

S1059-0560(16)30353-7 http://dx.doi.org/10.1016/j.iref.2016.12.004 REVECO1334

To appear in: International Review of Economics and Finance Received date: 21 January 2016 Revised date: 31 October 2016 Accepted date: 12 December 2016 Cite this article as: Md Lutfur Rahman, Doowon Lee and Abul Shamsuddin, Time-varying Return Predictability in South Asian Equity Markets, International Review of Economics and Finance, http://dx.doi.org/10.1016/j.iref.2016.12.004 This is a PDF file of an unedited manuscript that has been accepted for publication. As a service to our customers we are providing this early version of the manuscript. The manuscript will undergo copyediting, typesetting, and review of the resulting galley proof before it is published in its final citable form. Please note that during the production process errors may be discovered which could affect the content, and all legal disclaimers that apply to the journal pertain.

Time-varying Return Predictability in South Asian Equity Markets Md Lutfur Rahman, Doowon Lee*, Abul Shamsuddin Newcastle Business School The University of Newcastle University Drive, Callaghan, NSW 2308, Australia *

Corresponding author. [email protected]

Abstract Time-varying return predictability in four South Asian stock markets is examined using the wild-bootstrapped automatic variance ratio test and price delay measures. Strong evidence of predictability is found in aggregate market and size-sorted portfolio returns. The crosssectional variation in return predictability is inversely related to firm size and trading frequency, while the time variation in return predictability is related to market conditions— the level of equity market development, liquidity, volatility, automation of trading mechanism and financial crises. These results strongly corroborate Lo’s (2004) adaptive market hypothesis, and are robust to controlling for thin trading, changes in data frequency, and use of alternative return predictability measures.

JEL Classification: F30, G14, G15 Keywords: Return predictability, Adaptive market hypothesis, South Asian stock markets, Variance ratio, Price delay

1. Introduction The efficient market hypothesis (EMH) states that a market is infomationally efficient when asset prices fully and instantly reflect available information (Fama, 1970). Proponents of the EMH argue that market participants update their expectations about an asset’s intrinsic 1

value in response to new information arriving in asset markets. Hence, any deviation of an asset’s price from its intrinsic value is quickly exploited by arbitrageurs, and consequently mispricing cannot persist. This assertion may hold if there are no limits to arbitrage, all market participants have equal access to information, and they process information rationally. In practice, there are limits to arbitrage, and company insiders and institutional investors have more access to superior information than individual investors. Furthermore, investors do not always behave rationally (Shiller, 1983; Shleifer & Vishny, 1997). These attributes are more prevalent in emerging equity markets than in developed equity markets (see Morck, Yeung, & Yu, 2000; Lai, Ng, & Zhang, 2014). Thus, contrary to the EMH, there is a priori reason to believe that equity prices in emerging markets may not fully and instantly reflect available information. Even in the context of developed markets, many researchers have refuted the EMH (see Lo & MacKinlay, 2011). This study contributes to the ongoing debate over market efficiency in the context of emerging markets. We examine stock return predictability in four South Asian countries – Bangladesh, India, Pakistan and Sri Lanka.1 These countries have been selected for several reasons. In recent years, they have become attractive destinations for foreign portfolio investment (FPI). For example, in 2014, net FPI to these countries was US$13,683 million which was 66% of the net FPI to all lower middle income countries (World Bank, 2016).2 However, very little is known regarding informational efficiency of these markets. Compared to developed markets, the South Asian markets have a much lower market capitalization-to-GDP ratio and liquidity, restrictive capital flows, and less efficient trading mechanisms. These markets also differ from each other in terms of the level of equity market development, institutional settings, 1

The terms ‘return predictability’ and ‘market inefficiency’ are used interchangeably in this paper. Although the South Asian region consists of eight countries, we do not include Afghanistan, Bhutan, the Maldives and Nepal in our sample because data for individual stock prices are not available for them. 2 A country with a per capita GNI US$1,046 to $4,125 is classified as a ‘lower middle income’ country by the World Bank. All the four South Asian countries belong to this category. As of 2014, per capita GNI was US$1,560 in India, 1,400 in Pakistan, 3,650 in Sri Lanka and 1,080 in Bangladesh (World Bank, 2016).

2

information technology adoption, and corporate governance practices. Financial liberalisation of the South Asian markets over the last two decades has potentially exposed them to financial crisis. Thus, these markets are good candidates for studying the impact of diverse market characteristics on return predictability. 3 Also, findings of this paper may provide important policy directions in order to: firstly, promote informational efficiency; and secondly, help investors formulating investment strategies based on evidence of time-varying return predictability. Although return predictability is extensively investigated in developed markets, those located in South Asia are underrepresented in the literature. More importantly, the literature on emerging market efficiency is limited in a number ways. First, return predictability is typically examined ignoring its variation over time. Second, return predictability evidence is based on conventional efficiency tests that exhibit size distortion and low power, particularly in a small sample. Third, while most studies examine autocorrelation in stock returns, there is lack of studies examining the speed of market-wide information incorporation into stock prices. Fourth, return predictability is examined without properly addressing emerging market features, for example, infrequent trading and the greater prevalence of micro-cap and smallcap stocks.4 Fifth, most studies on emerging market efficiency are based on stocks that are listed until the end of the sample period, which may result in survivorship bias. This study overcomes all of these limitations.

3

Although it is claimed that factors such as level of equity market development (Shamsuddin & Kim, 2010), efficient trading mechanism (Martens, 1998), high liquidity (Chordia, Roll, & Subrahmanyam, 2008), shortselling facility (Chang, Luo, & Ren, 2014), improved corporate governance (Lagoarde-Segot & Lucey, 2008) and openness to foreign investors (Bae, Ozoguz, Tan, & Wirjanto, 2012) do have an impact on the degree of return predictability, the empirical evidence is mixed and conflicting. Furthermore, there is a lack of consensus regarding time-variation in return predictability and its determinants. 4 Firm size is important in investigating return predictability because there are many examples of the inverse relationship between return predictability and firm size (Fama, 1991). This anomaly is explained in terms of infrequent trading (Lo & MacKinlay, 1988), less coverage by financial analysts (Hong, Lim, & Stein, 2000) and high transaction costs (Griffin, Kelly, & Nardari, 2010) which are commonly associated with small stocks.

3

More specifically, this paper contributes to the existing literature in a number of ways. First, the wild bootstrapped automatic variance ratio (WBAVR) (Kim, 2009) test is used as a measure of market efficiency. This test statistic provides robust measure of market efficiency by removing the problem of arbitrary lag-order selection, and exhibits superior size and power properties relative to conventional efficiency tests that are commonly used in the literature (see Griffin, Kelly, & Nardari, 2010). Second, in emerging markets, common information is typically incorporated into stock prices with delay due to infrequent trading (see Harvey, 1995). We address this issue by using price delay (Hou & Moskowitz, 2005) as our second efficiency measure, and examining the sensitivity of the price delay measure to trading frequency. The price delay measure focuses on the delay to which stock price incorporates market-wide information. Third, we apply return predictability measures on overlapping subsamples of stock returns, allowing an examination of time variation in return predictability. This analysis is relevant with reference to the fast-changing South Asian emerging markets (see Romero-Torres, Wells, & Selwyn-Khan, 2013). Fourth, variation in market characteristics in the South Asian countries provides an appropriate context to verify Lo’s (2004) assertion that time variation in return predictability depends on market conditions. To this end, we run a regression of the return predictability measure on market condition variables such as the level of equity market development, liquidity, the extent of information technology adoption, stock market cycle and financial crises, among others. These potential determinants have not been considered by earlier studies in explaining return predictability in emerging markets. Finally, we use a sample that is free of survivorship bias and filtered for infrequent trading, outliers and IPO-related anomalies. Overall, we find strong evidence of inefficiency in aggregate market and size-sorted portfolio returns for daily data even after controlling for thin trading. The level of market efficiency is found to be inversely related to firm size and trading frequency. The degree of 4

return predictability is less prevalent in India and Pakistan compared to that in Bangladesh and Sri Lanka. Overlapping subsample analysis suggests that the level of market efficiency varies over time in an oscillatory fashion, which is analogous to observations for the US and other developed stock markets (Lo, 2004; Kim, Shamsuddin, & Lim, 2011; Urquhart & Hudson, 2013; Kim & Shamsuddin, 2015). We find that time-varying return predictability is related to market conditions and institutional settings, which supports Lo’s (2004) adaptive market hypothesis. The rest of the paper proceeds as follows. Section 2 presents a review of the relevant literature. Section 3 describes data and computational details. In section 4, we focus on measures of return predictability and methodology associated with examining time-varying predictability. Empirical results are discussed in section 5. We present robustness checks in section 6, and finally, section 7 concludes the paper with a summary of the main themes covered here.

2. Literature review The EMH receives strong support until the 1970s as the literature finds that stock price changes are essentially random (Fama, 1970). However, in the 1980s, researchers provide convincing evidence of return predictability particularly in the short-horizon, which cannot be solely attributed to market microstructure biases (for example, non-trading and bid-ask bounce) or time-varying expected returns (Lo & MacKinlay, 1988; Conrad & Kaul, 1989). Behaviourists argue that psychological biases prevent investors from reacting rationally to new information that leads to significant return autocorrelation over the short-horizon (Barberis, Shleifer, & Vishny, 1998; Hong & Stein, 1999). Lo (2004) makes an attempt to reconcile the EMH with the behavioural arguments and proposes a new paradigm, known as the adaptive market hypothesis (AMH). The basic tenet 5

of the AMH is that market efficiency is time-varying. The AMH is based on Simon’s (1955) concept of bounded rationality which indicates that individual investors, due to their lack of computational abilities, tend to follow a satisficing rather than an optimising approach to decision-making. Market participants compete, learn from their mistakes and adapt to changing market conditions that eventually result in a satisficing outcome. Thus, the AMH views market efficiency from an evolutionary viewpoint, while the EMH is based on an allor-nothing notion of efficiency. While the EMH claims that the markets are in general efficient and unpredictable, the AMH argues that return predictability and informational efficiency can coexist in a market over time. According to the AMH, trading on past information can be successful or unsuccessful based on a particular market environment (for example, the degree of competition, profit-making opportunities and market participants’ adaptability to changing market conditions). However, the EMH claims that past information cannot be used to formulate profitable trading strategy. Lo (2005) maintains that the notion of investors’ irrationality is consistent with time-varying market efficiency arising from investors’ continuing efforts to adapt to changing market conditions. A number of studies examine time-varying return predictability and find strong support for the AMH. Kim et al. (2011) find that changing market conditions drive return predictability over time in the US. The authors show absence of return predictability during stock market crashes, low predictability during market bubbles and high predictability in periods of fundamental political and economic crises. Kim and Shamsuddin (2015) show that time variation in return predictability is also influenced by firm size. In particular, largecapitalization portfolios exhibit lower return predictability since the early 1980s, possibly in response to implementation of advanced information technology and reductions in transaction costs. The studies by Ito, Noda, and Wada (2014, 2016) and Noda (2016) use a non-Bayesian time-varying vector autoregressive model to examine change in the degree of market 6

efficiency over time. Ito et al. (2016) report that the US market is mostly efficient from 1871 to 2012 except when economic crises erupt. The authors attribute this outcome to changes in individual investors’ risk aversion. Noda (2016) claims that time-varying return predictability in the Japanese market is linked to the oil crisis and the asset price bubble. Urquhart, Gebka, and Hudson (2015) show that profitability from a moving average-based trading strategy declines in the post-1987 period in the US, the UK and Japan. Other studies on developed and emerging markets indicate that dynamic stock return behaviour can be better explained by the AMH than the EMH (see Abdmoulah, 2010; Lim, Luo, & Kim, 2013; Urquhart & Hudson, 2013; Urquhart & McGroarty, 2014). Time-varying return predictability is also found in foreign exchange market (Katusiime, Shamsuddin, & Agbola, 2015), international precious metal markets (Charles, Darné, & Kim, 2015) and energy markets (Khediri & Charfeddine, 2015), among others. These studies essentially argue that time variation in return predictability is a function of economic and political factors. For instance, foreign exchange return predictability decreases in response to central bank’s interventions, but increases due to the imposition of market regulation and emergence of a financial crisis (Katusiime et al., 2015). On the other hand, tightening monetary policy, tensions in metal-producing countries, and supply shocks, among others, are found to influence return predictability in precious metal markets (Charles et al., 2015). Traditionally, emerging markets show high average returns, weak relationship with developed markets and a high degree of return predictability (Bekaert & Harvey, 2002). However, contrary to conventional beliefs, Griffin et al. (2010) find that emerging markets are at least as efficient as developed markets. Although emerging markets in Europe and South East Asia have received some attention in the literature, the South Asian markets are largely ignored. The scanty literature on the South Asian markets provides mixed evidence of market efficiency. Poshakwale (2002) reports non-linear serial dependence in returns for an 7

equal weighted portfolio of 100 large stocks in India. Results of the study may be subject to idiosyncratic noises since efficiency measures are applied to individual stock returns. Abeysekera (2001) finds asymmetric return predictability for short and long-horizon data in Sri Lanka. This study applies conventional efficiency tests (for example, autocorrelation test, unit root test and runs test), which do not have satisfactory size and power properties (Fama, 1991). These two studies may suffer from survivorship bias as delisted stocks are not included in the sample. While the Bangladeshi stock market exhibits short-horizon return predictability prior to the 1997 stock market crash, return predictability disappears after the crash (Islam & Khaled, 2005). The authors examine autocorrelation in daily, weekly and monthly market index returns without taking into account the potential effects of nonsynchronous or thin trading on their measure of return predictability. Based on monthly stock market indices, Chaudhuri and Wu (2003) find structural breaks in return autocorrelations in the South Asian stock markets. Griffin et al. (2010) reveal significant return predictability in these markets, where the predictability appears to be associated with portfolio characteristics and information environment. The last three studies employ the Lo and MacKinlay (1988) variance ratio test, which is subject to size distortion particularly in a small sample. Moreover, none of these studies examine time-varying return predictability and its sources.

3. Data description This study concentrates on the equity markets of Bangladesh, India, Pakistan and Sri Lanka. Characteristics of these markets are summarised in Table A1 in the Appendix. These markets are at different stages of development. India and Pakistan are secondary emerging markets, whereas, Bangladesh and Sri Lanka are frontier markets according to the Financial Times Stock Exchange (FTSE) country classification review (FTSE, 2015). Only India has achieved a credit rating of ‘investment’ grade while the other three markets are rated as ‘speculative’. The Indian market is the largest in the region with 7,248 securities listed at the 8

end of 2014. The number of listed securities in the other three countries was 557 in Pakistan, 274 in Bangladesh and 294 in Sri Lanka. During the same year, stock traded as a percentage of GDP (a measure of liquidity) was 37.65% in India while this ratio was less than 20% in the other three countries. The level of information technology adoption (such as, automated trading and settlement, and internet-based trading) is higher in India and Pakistan compared to that in Bangladesh and Sri Lanka. Short-sale and stock lending facilities are still prohibited in Bangladesh and Sri Lanka, but they are allowed in India and Pakistan. Only India has a substantial derivative market in the South Asian region. Although Pakistan did introduce an equity derivative market in 2001, the other two markets are yet to initiate a derivative market.5 The sample period ranges from January 1995 to December 2013. Reliable market data for individual stocks for all of these markets are available only from the mid-1990s. Stock price and market capitalization data are collected from Thomson Reuters DataStream. We use value-weighted portfolio returns in both daily and weekly frequencies. The sample includes all common stocks (both active and delisted) in the major stock exchanges of these countries. DataStream adds some texts to the security’s name field to indicate the type of security.6 Name fields for all the securities were searched extensively to eliminate non-common equity. Eventually we exclude duplicates, mutual funds, bonds, debentures, preferred stocks, unit trusts and other non–common equity. Financial stocks are excluded to ensure exclusion of mutual funds and investment trusts. We apply separate filters to eliminate secondary listing, non-voting shares and rights. Delisted common stocks are included to avoid survivorship bias. Excluding delisted stocks can understate return

5

Information, summarized in this paragraph, is obtained from the FTSE market assessment matrix (FTSE, 2015), World Development Indicators (World Bank, 2016), and respective stock exchange websites. 6 See Appendix Table B.1 in Griffin et al. (2010, p. 3272) for a list of the texts.

9

predictability since such stocks are predominantly small, thinly traded particularly just before being delisted, and thereby likely to show higher return predictability. A firm is included in the sample three months after its listing on the relevant stock exchange. This filter removes stock return outliers that may arise from IPO-related information asymmetries. We implement two more filters to exclude outliers. First, if daily and weekly returns ( ) are greater than 200%, then the return is set to missing. Second, in the cases of both daily and weekly returns if , then both

and

or

and (

)(

)

are considered to be missing values. 7 Since imputing

missing data can induce autocorrelation (Anderson, Eom, Hahn, & Park, 2013), this study excludes stocks with the missing values to avoid spurious return predictability. Finally, a stock is included in the sample if it has both daily closing price and market capitalization data. Since infrequent trading may lead to spurious return predictability, this study examines a full sample and two subsamples obtained by implementing price-change filters. The full sample includes all stocks; while the 25% price-change filter and the 50% price-change filter include stocks that are traded on at least 25% and 50% of the total trading days in a year, respectively.8 The full sample includes 2,774 stocks from India, 275 from Pakistan, 284 from Bangladesh and 253 from Sri Lanka. The number of stocks in the full samples and the subsamples obtained by implementing price-change filters are reported in Panel A of Table A2 in the Appendix. This panel also reports segregation of the total sample into active and delisted stocks.

7

These filters are also used in Griffin et al. (2010). Griffin et al. (2010) use 30% and 75% price-change filters. However, applying the 75% price-change filter in the South Asian markets eliminates the majority of stocks from the sample. To ensure that the number of stocks in the usable sample is sufficient, this study employs 25% and 50% price-change filters. 8

10

Return predictability measures are employed on returns of aggregate market and sizesorted portfolios. The use of portfolio returns instead of individual stock returns mitigates idiosyncratic noises associated with individual stock returns and allows us to condition our analysis for portfolio characteristics such as firm size. Annually rebalanced size-sorted portfolios are constructed. All the stocks are sorted on the basis of year-end market capitalization and divided into size quintile portfolios in the subsequent year. Panel B of Table A2 reports the average number of stocks for each size-sorted portfolio.

We first calculate simple price return on an individual stock as price change over the time period t-1 to t divided by price at t-1. The value weighted portfolio return is computed in the following manner: ∑

[ ( where,

(

))]

(1)

is continuously compounded portfolio return and

is the simple return for stock i at time t; and

( ) is the natural log function;

represents weight of stock i at time t,

which is the proportion of an individual stock’s market capitalization relative to total market capitalization of the portfolio. We calculate weekly returns by summing log daily returns from each Wednesday to Tuesday. This approach aims to circumvent possible day-of-theweek bias in stock returns.

4. Measures of return predictability This study applies the wild bootstrapped version of the automatic variance ratio (AVR) test (Kim, 2009) and price delay measure (Hou & Moskowitz, 2005) to assess return predictability.

4.1 Wild bootstrapped AVR test

11

The variance ratio test developed by Lo and MacKinlay (1988) is based on the idea that variance of uncorrelated increments of a time series, such as stock price, is proportional to the sampling intervals. Thus, the variance ratio for kth return horizon is the ratio of k-period return variance to 1/k times one-period return variance. If rt is a security return at time t, the variance ratio of rt for a holding period of k is as follows: k 1 Var[rt (k )] i  VR (k )  1  2 1   (i) , kVar[rt ] k i 1 

(2)

where, ρ(i) is the autocorrelation coefficient of rt for ith lag. Thus, variance ratio is a linear combination of autocorrelation coefficients associated with declining weights. To implement Lo and MacKinlay’s (1988) variance ratio test, a set of alternative holding periods (lag intervals, k) is typically selected by researchers without statistical justification. Choi (1999) proposes the automatic variance ratio (AVR) test where the optimal lag length, k, is determined by a fully data-reliant procedure. Under the null hypothesis of serially uncorrelated returns, the AVR test of Choi (1999) takes the following form: (̂)

√

̂[

(̂)

] √ → (

)

(3)

where T is the sample size, ̂ is the optimum choice of k and variance ratio (VR) is calculated as (̂)

∑

( ̂ ) ̂( ),

(4)

where ̂ is sample autocorrelation coefficient and

( ) is a weighting function.

Choi’s (1999) AVR test depends on the assumption that return series is i.i.d (independent and identically distributed). Using an extensive Monte Carlo experiment, Kim (2009) shows that the AVR test exhibits serious size distortion in a small sample and heteroskedastic return.

12

He proposes a wild bootstrapped AVR test (WBAVR) that does not show size distortion and has desirable power properties under conditional heteroskedasticity. The wild bootstrap method is a resampling technique that approximates the sampling distribution of a statistic using the following procedure: firstly, a bootstrap sample is formed of T observations; secondly, the AVR statistic is calculated for each bootstrapped sample; and thirdly, the first and second steps are repeated many times to form a bootstrap distribution of AVR statistics.9 The two-tailed p-value is computed as the proportion of absolute values of AVR statistics from the bootstrapped sample and absolute values of AVR statistics derived from real data. The WBAVR test is applied to both daily and weekly returns. Daily data is more informative but argued to be greatly affected by microstructure biases, such as non-trading and bid-ask bounce. On the other hand, weekly data is less susceptible to market microstructure biases. There is mounting evidence of higher autocorrelation in high frequency returns (Patro & Wu, 2004). Using both daily and weekly data enables us to examine whether return predictability is related to data frequency, even after conditioning efficiency measures on trading frequency and firm size.

4.2 Price delay Price delay investigates whether information contained in lagged market returns is quickly incorporated in contemporaneous stock returns. It is therefore a measure of sensitivity of stock returns to past market-wide information. Local delay provides a measure of stock price delay with respect to local market returns, whereas, global delay considers both local and global market returns. To calculate delay, we first estimate the unrestricted models in which returns on each size-sorted portfolio are regressed on current and lagged market returns following Hou and Moskowitz (2005): 9

The number of bootstrap iterations is set to 500 in this study following Kim (2009).

13

∑

(5) ∑

where,

∑

,

is the return on size-sorted portfolio i at time t,

global market portfolios, respectively and

and

(6)

are returns on local and

is the random error term. We calculate return

on local market portfolio using Eq. (1) for each South Asian market. Global market return is calculated from the FTSE all-world index.10 Eq. (5) and (6) are the unrestricted models for calculating local and global delay, respectively. These models include four lagged market returns since four-week time is substantial for stock prices to respond to market-wide information even in a thinly traded market (Hou & Moskowitz, 2005). Next, the restricted models are constructed by setting the coefficients of lagged market returns to zero: (7)

(8) The unrestricted and restricted models provide the basis for calculating price delay. While the unrestricted model uses current and lagged market returns to explain variation in sizesorted portfolio returns, the restricted model uses only current market returns to explain variation in size-sorted portfolio returns. If stock prices imbed market-wide information immediately, coefficients of lagged market returns (

and

different from zero while those of current market returns (

) will not be significantly and

) will be statistically

different from zero. As a result, explanatory power (R-square) of the unrestricted and 10

The FTSE all-world index includes 47 countries including India and Pakistan. However, India and Pakistan contribute a negligible 1.05% and 0.01%, respectively, to global market capitalization of this index. Thus, double counting is not likely to distort the delay results.

14

restricted models will be similar. However, if stock prices do not instantaneously reflect market-wide information, coefficients of lagged market returns will be significantly different from zero and the unrestricted model will exhibit greater explanatory power (higher Rsquare) relative to the corresponding restricted model. This is because lagged market returns explain a part of the variation in size-sorted portfolio returns. Accordingly, a delay measure can be constructed which compares the proportion of variation of current size-sorted portfolio returns explained by past market returns from the two models. More specifically, Delay =

(9)

If R2 in the unrestricted model is greater than R2 in the restricted model, price delay will take a higher value implying delayed incorporation of market-wide information in stock prices. Conversely, price delay closer to zero indicates less delay and greater market efficiency. Griffin et al. (2010) argue that the unrestricted model may have an artificially inflated explanatory power due to increased number of regressors. Hence, they use adjusted R2 instead of R2 to calculate delay (

) as given in the equation below: (10)

We calculate price delay using weekly returns. Daily or intra-day returns may show higher volatility in price delay across size-sorted portfolios and over time due to market microstructure noise in high-frequency data. Also, historical intra-day data are not available for all stocks in our sample. Price delay based on monthly data, on the other hand, should show little variation across stocks since most of the stocks incorporate information within a month. Hou and Moskowitz (2005) also argue that lower frequency data, such as, monthly 15

data can produce greater error in estimating price delay. Thus, this study uses weekly returns to calculate the price delay measures.11

4.3 Measuring time-varying return predictability The concept of time-varying return predictability is introduced by Lo (2004) within the framework of the Adaptive Market Hypothesis (AMH). He argues that investors suffer from “bounded rationality” and exhibit a satisficing rather than optimizing behaviour. Due to inadequate information and limited cognitive ability, investors use heuristics or rule of thumb to attain satisficing investment outcomes. Nonetheless, a satisficing outcome does not come automatically; instead it emerges through a process of evolution involving trial and error. This process ensures “the survival of the fittest”. Lo (2005) argues that investors are intelligent, learn from their mistakes and adaptable to changing market conditions. This evolutionary perspective indicates that profit-making opportunities arise from time to time, however, they disappear as investors exploit them. This cycle continues based on investors’ demographics, institutional factors and market conditions. Therefore, the AMH suggests that the market may oscillate between efficiency and inefficiency over time. We apply the WBAVR test on overlapping subsample windows to capture potential variation of return predictability over time. The use of overlapping subsamples helps to mitigate data snooping bias. We use fixed 24-month window period that advances monthly. Use of this window period is consistent with other studies (Kim et al., 2011; Lim et al., 2013). The first overlapping window ranges from May 1995 to April 1997.12 The second one starts from June 1995 and spans to May 1997. This process continues until the end of the sample period and consequently we obtain 201 overlapping subsample windows. 11

Weekly returns are also used by Hou and Moskowitz (2005) and Bae et al. (2012). Like Hou and Moskowitz (2005), we also use daily returns to check the robustness of the results. 12 The first subsample window does not start on the first trading day of January 1995 since the study excludes the first 90-day trading history.

16

We calculate the AVR test statistic for each overlapping subsample. Since the AVR test statistic is a scaled version of variance ratios, it is used as a measure of return predictability. We investigate the association between degree of return predictability and market conditions using the following ordinary least squares (OLS) regression model. |

|

where |

(11) | is the absolute value of variance ratio statistic for portfolio i for the subsample

window ended at month t;

is a column vector of market condition variable;

of return predictability that cannot be explained by parameters; and

is a random error term.

is the level

is a row vector of corresponding

includes three continuous variables, these

being market development, liquidity, volatility, and three dummy variables, i.e., automated trading, financial crisis and market downturn to indicate market conditions. 13 Since the dummy variables are not constructed for an overlapping window, it is possible that a dummy variable may take the value of 1 for an overlapping window where a particular market condition prevails only a month or two in that window. To mitigate this problem, if a specific market condition lasts for at least four consecutive months in a window then the corresponding dummy variable takes a value of 1 for that window, and 0 if otherwise.

5. Empirical results and discussion 5.1 Descriptive statistics 13

See Table A3 in the Appendix for definitions of the variables. To match with the left-hand-side variable in Eq. (11), we first take monthly observations of the continuous variables, then calculate moving average of the variables for the corresponding overlapping window. Taking a moving average may introduce serial correlation in the respective explanatory variables. In this circumstance, disturbances of the autoregressive structure of an |. Consequently, explanatory variable can be correlated with the error-terms from the regression model for | slope coefficient and standard error from ordinary least squares (OLS) regression can be biased in a small sample that may lead to over-rejection of the null hypothesis of no explanatory power (Nelson & Kim, 1993). To avoid this problem, we run an augmented regression model (ARM) proposed by Amihud and Hurvich (2004). However, our result from the ARM is not qualitatively different from the OLS regression results. Hence, the OLS results with heteroskedasticity and autocorrelation consistent (HAC) t-statistics are presented in Table 6.

17

Table 1 reports mean returns of aggregate market and size-sorted portfolios for both daily and weekly frequencies. Mean returns are calculated for the full sample and the two subsamples obtained by applying the 25% and 50% price-change filters. For market portfolios (reported in Panel A), both Pakistan and Sri Lanka generate the same mean daily returns (0.10%), followed by India (0.09%) and Bangladesh (0.08%) in the full sample. In weekly frequency, highest mean return is observed in Pakistan (0.48%), followed by Sri Lanka (0.46%), India (0.42%) and Bangladesh (0.35%) in the full sample. In general, the results show that portfolio returns for the full sample are smaller than those for subsamples obtained by applying the 25% and 50% price-change filters. This is expected because implementing trading filters excludes thinly traded stocks with zero returns on the non-traded days. However, we also observe lower mean returns for the 50% price-change filter compared to that of the full sample and 25% price-change filter, which may arise due to exclusion of infrequently traded stocks that generate high returns during the days these stocks are being traded. Thus, the relationship between mean returns and filter size is a priori indeterminate. From Panel B of Table 1, we find a negative association between mean returns and firm size. The smallest portfolio generates the highest return, which declines when there is an increase in firm size. This result holds for all the countries, all price-change filters and both data frequencies. A similar result is well documented in the literature for developed markets (Fama & French, 1992; Chordia & Swaminathan, 2000). Consistent with aggregate market portfolios, size-sorted portfolios do not show any association between mean return and filter size. In untabulated results, we find that both daily and weekly return series deviate from normality (based on the Jarque-Bera test) and the distributions are leptokurtic. The returns on aggregate market portfolios of India and Pakistan are negatively skewed, however, positive 18

skewness is found for the other two countries. With regard to size-sorted portfolios, large portfolios tend to show negative skewness (except Bangladesh) whereas small portfolios show positive skewness. 5.2 Return predictability 5.2.1 Return predictability in full sample period We first investigate return predictability for the whole sample period using the wild bootstrapped automatic variance ratio (WBAVR) test. This analysis assumes that the level of return predictability is static over time. Table 2 presents the results for aggregate market portfolios. For daily returns, the results reveal that the AVR test statistics are statistically significant for the full sample of stocks as well as for the subsamples obtained by employing two different price-change filters. This finding holds for all the four countries and therefore the random walk hypothesis is unambiguously rejected. Automatic variance ratio statistics for weekly returns are statistically significant for all the countries except India. This result is also robust for different trading filters. Table 3 documents AVR statistics and wild bootstrapped p-values for size-sorted portfolios. Results for daily and weekly returns are on Panel A and Panel B, respectively. In the case of daily returns, we report significant AVR statistics for all size-sorted portfolios which reject the random walk hypothesis (RWH). For weekly returns, the RWH cannot be rejected in the context of the largest size quintile portfolio for India and Pakistan irrespective of price-change filters. However, randomness in returns is rejected for all the size-sorted portfolios in Bangladesh and Sri Lanka. We also find that test statistics are the lowest in magnitude for the largest quintile portfolios, and they gradually increase with the decrease in size of the stocks. This result is consistent with the findings in the context of developed markets, revealing higher return predictability for small stock portfolios compared to their 19

large counterparts (Conrad & Kaul, 1989; Lo & MacKinlay, 1988, 1990; Hendershott & Seasholes, 2014). We discover a number of interesting findings from return predictability analysis of the whole sample period. First, positive and statistically significant AVR statistics (except in the cases of weekly returns on aggregate market portfolio in India and weekly large quintile portfolio returns in India and Pakistan) signify the presence of significant positive autocorrelations in daily and weekly returns. Evidence of positive autocorrelation in shorthorizon returns is consistent with the empirical literature (Lo & MacKinlay, 1988; Conrad & Kaul, 1989; Patro & Wu, 2004).14 Our result corroborates Anderson et al.’s (2013) partial price adjustment explanation. The authors argue that return autocorrelation may arise as transaction takes place without fully incorporating information due to market frictions. Second, although AVR statistics predominantly decline when a trading filter is implemented, they remain statistically significant even in the case of the 50% price-change filter, suggesting that return predictability evidence is not spurious due to infrequent trading. Third, return predictability evidence in daily returns cannot be solely attributed to microstructure biases since the RWH is also rejected in the case of weekly returns with a few exceptions in India and Pakistan.15 Fourth, among the four countries, return predictability appears to be less evident in India and Pakistan. This result is expected since the India and Pakistan stock markets are ahead of the other two South Asian markets in terms of market size, liquidity, accessibility to foreign investors, availability of stock lending and short-sale facilities,

14

A range of explanations are provided for positive autocorrelation in portfolio returns. The market microstructure effect, such as nonsynchronous trading and bid-ask bounce (Lo & MacKinlay, 1990), transaction cost (Mech, 1993), time-varying risk premia (Conrad & Kaul, 1989) and partial price adjustment (Anderson et al., 2013) are the pre-eminent reasons for return autocorrelations. 15 Autocorrelation in daily returns may arise due to nonsynchronous trading, bid-ask bounce or other microstructure biases (Anderson et al., 2013). However, microstructure bias is less prevalent in weekly and other longer horizon returns compared to daily returns.

20

efficient trading mechanism and developed equity market. 16 Previous studies argue that allowing foreign investors access to domestic equity market (Kim & Singal, 2000), developed equity market and appropriate regulatory framework (Kim & Shamsuddin, 2008), liquidity and good corporate governance (Lagoarde-Segot & Lucey, 2008), short-selling facility and strong regulatory enforcement (Shamsuddin & Kim, 2010), and automated trading mechanisms (Naidu & Rozeff, 1997) help investors to respond to new information quickly. To summarize, we find strong evidence of return predictability in the South Asian markets except for a few cases of no-predictability in weekly returns. This finding can be taken as evidence against market efficiency since return predictability generally exists in both daily and weekly returns and at different trading frequencies. This result contradicts the findings of Chaudhuri and Wu (2003) and Islam and Khaled (2005) who report efficiency in the Pakistan and Bangladesh stock markets, respectively. Our finding of less prevalent return predictability in India and Pakistan is at odds with Griffin et al.’s (2010) study, which reports a relatively higher degree of return predictability in India and Pakistan compared to Bangladesh. This discrepancy may be attributable to the use of different return predictability measures.17 Besides, Islam and Khaled (2005) do not take into account the potential effect of infrequent trading in daily data, while Chaudhuri and Wu’s (2003) results are derived from monthly data. 5.2.2 Time-varying return predictability

16

More specifically, the Bangladesh and Sri Lanka markets are smaller and less liquid compared to those of India and Pakistan. For instance, during the period 1995 to 2014, market capitalization as a percentage of GDP (average) was 57.17% in India and 20.68% in Pakistan. However, it was less than 20% in the other two countries. During the same period, value of share traded as a percentage of GDP (average) was 62.82% in India, 39.89% in Pakistan, 5.68% in Bangladesh and 3.17% in Sri Lanka (World Bank, 2016). As discussed in section 3, the Bangladesh and Sri Lanka markets suffer from regulatory impediments (for instance, a ban on stock lending and short-selling), lack of a derivative market and absence of an efficient trading mechanism (FTSE, 2015). 17 While we use the AVR test, all of these three studies use Lo and MacKinlay’s (1988) variance ratio test. Apart from this, Islam and Khaled (2005) use unit root test and the Box-Pierce test; Chaudhuri and Wu (2003) apply the unit root test which takes into account structural breaks.

21

We apply the WBAVR test on 24-month overlapping subsamples to capture time variation in return predictability. The test is conducted on both aggregate market and sizesorted portfolio returns for daily and weekly frequencies. Table 4 reports the percentage of total number of overlapping windows with evidence of significant return predictability at the 5% significance level. In the cases of aggregate market portfolios, Bangladesh exhibits the lowest percentage of significant windows (22.39%), while Sri Lanka represents the largest proportion of significant windows (76.12%) in daily frequency in the full sample. However, for weekly returns, the smallest percentage of significant windows is found in India (5.97%) followed by Bangladesh (12.93%), Pakistan (25.87%) and Sri Lanka (39.30%) in the full sample. With regard to size-sorted portfolios, we find that percentage of significant windows is lowest for the largest portfolio, and this proportion increases when firm size decreases. We further find that the larger the price-change filter, the smaller the percentage of significant windows. These results suggest that return predictability is negatively associated with firm size and trading frequency. The overlapping subsample results apparently differ from the whole sample period results as the latter shows strong evidence of market inefficiency, whereas the former indicates oscillations between efficiency and inefficiency over time. The results for the whole sample period are on par with the prevalence of return predictability in overlapping subsamples for medium- and small-sized portfolios in daily data. Moreover, the whole sample period includes a large number of observations, which may lead to the overrejection of the null hypothesis of no predictability at the conventional significance level of 5% (Neal, 1987). We plot p-values of WBAVR test statistics for weekly returns in Figure 1. The shaded areas in the graphs correspond to a significance level of 5% or less. A p-value above the shaded area indicates the absence of significant return predictability. For aggregate market portfolios (reported in Panel A), we find that all the markets except Sri Lanka exhibit lack of 22

predictability in the late-1990s. Capital market liberalization and automation of trading mechanisms during the 1990s may have contributed to this result. Previous studies provide evidence of less inefficiency following the implementation of an automated trading system (Naidu & Rozeff, 1994) and capital market liberalization (Kim & Singal, 2000). The prevalence of market inefficiency at the end of 2001 in all the countries is accompanied by economic and political turmoil following the 9/11 terrorist attacks. Although efficiency in these markets appears to be unaffected by the 1997 Asian financial crisis, the markets display significant predictability during 2008-2009 which can be attributed to the global financial crisis (GFC) (2007-2009) and subsequent economic slowdown. 18 This result may appear conflicting but it is economically rational. South Asian countries have stronger economic linkages with the GFC-ridden countries such as the US and the UK than the Asian crisisridden countries of South East Asia. The South Asian countries’ major sources of foreign portfolio investment (FPI) and the largest trading partners are the US and the UK (Kumar, 2012; Mehta & Yadav, 2013). These two countries were greatly affected by the GFC (see Claessens, Dell’Ariccia, Igan, & Laeven, 2010), which exerted a contagion effect on the South Asian economies that eventually contributed to decreased equity market efficiency during the GFC. On the other hand, the Asian financial crisis had a diminutive impact on the South Asian markets because of these nations’ weak economic linkages with the Asian crisisridden countries in South East Asia such as Thailand, Indonesia, South Korea, Malaysia and the Philippines (see Johnson, Boone, Breach, & Friedman, 2000; Mitton, 2002).19

18

Greater return predictability during economic crisis is also reported by Kim et al. (2011) and Ito et al. (2016), among others. 19 For example, portfolio investment from the US and the UK to Pakistan was US$-154.3 million in 2008, which coincides with the peak of the GFC. In the same year, portfolio investment from the five South East Asian countries to Pakistan was US$1.0 million. During 1997–2001, Pakistan’s major sources of FPI were the US, UAE and Hong Kong whereas there was no cross-country portfolio capital flow between Pakistan and the South East Asian countries (see State Bank of Pakistan, www.sbp.org.pk/ecodata/index2.asp#external). In the case of India, in 2008, about 60% of the registered foreign institutional investors were from the US and the UK. This trend was similar during the late 1990s (see Securities and Exchange Board of India Bulletin, www.sebi.gov.in/sebiweb/home/list/4/30/0/0/SEBI-Bulletin). The US and the UK are also the largest export destinations for the South Asian countries. In 1997 (2008), Bangladesh exported 46% (34%), India 22% (16%),

23

The Pakistan stock market reverts to efficiency from 2003 onwards, which coincides with the removal of short-selling restrictions, implementation of internet-based trading, and improvement in corporate governance and disclosure practices. Market inefficiency in Bangladesh during 2005-2006 is accompanied by political turmoil resulting from the demand for a free and fair general election. The Sri Lankan stock market displays significant predictability until 1999 with an improvement in efficiency during the next two years. However, the market again shows evidence of high predictability in 2007-2009, with a trend towards efficiency after 2009. Such oscillating patterns in return predictability may be credited to the ongoing civil war throughout the 1990s, the cease fire agreement in 2001, resumption of the civil war after 2006 until it ended in 2009. Consistent with aggregate market portfolios, p-values for size-sorted portfolios (reported in Panel B) present evidence of time-varying predictability attributable to different events and shocks discussed in the preceding paragraphs. Overall, we find that the return series conform to random walk in some overlapping subsamples and deviate from random walk in other overlapping subsamples. This result strongly supports the adaptive market hypothesis (AMH) and is consistent with Lo’s (2004) notion of “context dependent and dynamic” market efficiency.

5.3 Price delay Price delay measures the delay with which stock price incorporates market-wide information. Higher delay suggests that stock price incorporates market-wide information

Pakistan 29% (21%), and Sri Lanka 55% (45%) of their total exports to the US and the UK. However, total exports to the five South East Asian countries were much lower. In 1997, Bangladesh exported 0.28%, India 8.55%, Pakistan 5.85% and Sri Lanka 0.45% of their total exports to the five South East Asian countries (see Observatory of Economic Complexity, http://atlas.media.mit.edu/en/). As far as foreign direct investment (FDI) is concerned, the South Asian countries receive the largest FDI inflows from the US, Japan, and the UK, but negligible inflows from the South East Asian countries (Sahoo, Nataraj, & Dash, 2006).

24

slowly, indicating a departure from market efficiency. Delay is calculated for weekly returns on five size-sorted portfolios. Table 5 presents price delay measures for size-sorted portfolios calculated from Eq. (9).20 For the smallest quintile portfolio, local price delay is highest for India (0.1814), followed by Pakistan (0.1201), Bangladesh (0.0358) and Sri Lanka (0.0134) in the full sample. The results for global price delay reveal a very similar pattern in all the countries for the smallest quintile portfolio. However, the result is nearly reversed for the large quintile portfolio with India exhibiting the lowest local delay (0.0001), followed by Pakistan (0.0004), Sri Lanka (0.0013) and Bangladesh (0.0019). Thus, cross-country variation in price delay is significantly affected by firm size. Price delay is larger for small-cap portfolios, and it generally declines with an increase in firm size. This result holds for all the countries regardless of which price-change filter is used. Therefore, large stocks incorporate market-wide information more quickly compared to small stocks even after controlling for trading frequency. This finding: firstly, complements the WBAVR test result of an inverse relationship between return predictability and firm size; and secondly, is consistent with the findings of Chordia and Swaminathan (2000), Hou and Moskowitz (2005), and Hou (2007). Price delay decreases with filter size. For example, in the case of India, local price delay for the smallest quintile portfolio is 0.1814 in the full sample; and 0.1312 and 0.0911 in the subsamples obtained by applying the 25% and 50% price-change filters, respectively. We find a similar pattern for the other size-sorted portfolios in the sample countries. Implementing trading filters reduces price delay by excluding thinly traded stocks that sluggishly incorporate market-wide information. This outcome is compatible with the findings from developed markets (Chordia & Swaminathan, 2000; Hou & Moskowitz, 2005). 20

Table 5 shows the price delay measure of Hou and Moskowitz (2005).We also calculate price delay following Griffin et al. (2010) as presented in Eq. (10) and obtain qualitatively similar results, which are available from the corresponding author on request.

25

Global price delay also exhibits an inverse relationship with firm size and filter size. Not surprisingly, global delay is generally higher than local delay for all size-sorted portfolios. This finding holds for all but the Indian stock market. For example, in Bangladesh, local delay for the smallest portfolio is 0.0358 compared to corresponding global delay of 0.1265 in the full sample. The result is intuitive because local information is more readily available and less costly to obtain, so it is incorporated into stock prices more quickly compared to global information. Global information diffusion is facilitated by greater participation of foreign investors in the local market (Bae et al., 2012), but the South Asian markets experience a limited participation by foreign investors, which explain the delay in global information diffusion. To examine time variation in price delay, we calculate price delay for two-year overlapping subsamples from May 1995 to December 2013. We run rolling regression to obtain R2 for each window of 104 weeks length (2 years). The window is moved forward by 4 weeks to obtain the next subsample and this process is repeated until the end of the sample period. Price delay measures, calculated from rolling regressions, are plotted in Figure 2. We find that the delay to which stock prices respond to market-wide information is time-varying. In particular, we report that the smallest quintile portfolio not only shows a larger price delay but also a greater fluctuation in price delay over time. In contrast, for most overlapping subsamples, price delay for the largest quintile portfolio is close to zero in magnitude, showing little or no time variation. In general, the time path of price delay for the mid-cap portfolio lies between those for the large-cap and small-cap portfolios. We do not see any synchronised movements in price delay across countries in terms of peaks, troughs or longterm trends. For example, for the smallest quintile portfolio, India demonstrates a rising trend in price delay but other countries show oscillations in price delay without any secular trend.

26

This finding indicates that investors may be able to exploit market inefficiency by adopting a country rotation investment strategy (along with size-based investment strategies).

5.4 Determinants of return predictability The degree of return predictability ( |

|) is explained in terms of market and

institutional conditions based on Eq. (11). |

| is regressed on a number of continuous and

dummy variables representing market and institutional conditions. We include three continuous variables: the ratio of market capitalization to industrial production–a measure of market development; 21 the ratio of trading volume to market capitalization–a measure of liquidity in the market; and market volatility. We use realised volatility as a measure of market volatility, which is calculated by taking log of the square root of the sum of squared daily returns for a month. 22 We expect a decline in return predictability with the development of an equity market. Shamsuddin and Kim (2010) report that there is negative correlation between return predictability and equity market development in a cross-country setting. In the case of liquidity, we intend to examine two competing hypotheses. On one hand, the literature provides evidence that increased liquidity facilitates arbitrage trading activity that results in enhanced market efficiency (Chordia et al., 2008). On the other hand, liquidity can impede market efficiency when it is an indicator of uninformed or noise trading. Bloomfield, O’hara, and Saar (2009) assert that noise trading increases trading volume and market depth but hinders the market’s ability to adjust to new information due to adverse selection and pricing errors (deviation of transaction price from its true value). 23 The third explanatory variable, volatility, is expected to have a negative association with the degree of return 21

The ratio of market capitalization to GDP is a commonly used indicator of equity market development (e.g., Shamsuddin & Kim, 2010). However, we use industrial production instead of GDP because the latter is unavailable in monthly frequency. 22 This is a non-parametric ex-post estimate of the return volatility. A similar measure is also used by Welch and Goyal (2008) and Kim et al. (2011), among others. 23 Market development and liquidity appear to be highly correlated in all the four countries. To address this problem, we orthogonalise the liquidity variable. More specifically, liquidity is regressed on market development variable, and the resulting residual series is used as the liquidity measure.

27

predictability. Gu and Finnerty (2002) argue that higher volatility, arising from greater speculative activities, results in a low degree of return autocorrelation. Apart from these variables, three dummy variables are included in our model. First, we include automated trading dummy to represent the level of information technology adoption. This dummy takes value of 1 once the automated trading mechanism is implemented in the respective markets, and 0 before that. Automation of the trading mechanism is likely to reduce information and transaction cost that in turn accelerates information incorporation into stock prices and decreases return predictability (see Naidu & Rozeff, 1994). Second, a dummy variable for the Asian financial crisis and global financial crisis is used to examine the impact of crises on the degree of return predictability. This variable is constituted by combining the Asian financial crisis and GFC. It takes a value of 1 during the GFC and Asian financial crisis, and 0 if otherwise. Third, in order to examine the relationship between return predictability and market states, we include a market ‘downturn’ dummy. We define market state as ‘downturn’ when the cumulative monthly return in a particular overlapping window is negative.24 This variable takes a value of 1 during market downturn, and 0 if otherwise. Table 6 presents the regression results. As expected, the degree of return predictability decreases with the level of market development in Pakistan, but remains unchanged in the other three countries. Return predictability reacts positively to liquidity in all the markets except in Bangladesh. This result is inconsistent with Chordia et al.’s (2008) argument of greater efficiency associated with higher liquidity, however, it is in line with the notion that noise trading increases liquidity and reduces the market’s ability to respond to new information (Bloomfield et al., 2009). Market volatility reduces return predictability in all the four countries and the effect is statistically significant.

24

This is analogous to Cooper, Gutierrez, and Hameed (2004) who use the 36-month cumulative market return to identify the market state.

28

The degree of return predictability negatively responds to automation of the stock exchanges. The coefficient is statistically significant in India, Bangladesh and Sri Lanka. This result supports Naidu and Rozeff (1994) and Gu and Finnerty (2002) in the cases of Singapore and US markets, respectively. Financial crisis adversely affects market efficiency in the South Asian countries except in India, which is consistent with the contention that investors tend to overreact to common information during a financial crisis (Lim, Brooks, & Kim, 2008). Kim et al. (2011) provide similar findings for the US market. Return predictability decreases during market downturns in India, but it is invariant to this market state in the other three countries. In order to examine effect size of the independent variables, we first multiply regression coefficients with standard deviation of the corresponding variable, and then resulting estimates are expressed as a percentage of average |

| . The result reveals that one

standard deviation increase in market development causes a 35% reduction in average |

|

in Pakistan. These rates are 14% and 11% in India and Sri Lanka, respectively but they are statistically insignificant. We also find that one standard deviation rise in liquidity results in more than 100% increase in average |

| in India and more than 70% rise in Pakistan and

Sri Lanka. Finally, one standard deviation rise in volatility leads to a decline in average |

| by 23% in India, 32% in Pakistan, 33% in Sri Lanka, and 17% in Bangladesh. Among

the three variables, therefore, the contribution of liquidity in explaining |

| is the highest.

In summary, we find that market development, return volatility, automated trading and market downturn contribute to decline in return predictability. On the other hand, liquidity and financial crisis result in greater predictability in stock returns. Nonetheless, the

29

directional relationship between return predictability and its determinants are not robust in all these countries.25

6. Robustness checks 6.1 Equal-weighted returns Since our previous analysis is based on value-weighted portfolios, we apply the WBAVR test on equal-weighted market and size-sorted portfolios. While we do not tabulate these results in this paper, they can be obtained from the corresponding author on request. In general, evidence of return predictability is more pronounced for equal-weighted portfolios than value-weighted portfolios in terms of both the magnitude of AVR statistics and their statistical significance.26 This result is consistent with the literature (Lo & MacKinlay, 1988) in the context of developed markets. Value-weighted portfolios provide higher weights to large capitalization stocks that are more liquid and actively traded in the market, facilitating diffusion of information more quickly.

6.2 Non-overlapping subsamples In Section 5 the results from overlapping subsamples are presented as evidence of timevarying return predictability. As a robustness check, we calculate wild bootstrapped p-values of AVR statistics and price delay for two-year non-overlapping subsamples for weekly returns. Wild bootstrapped p-values for AVR statistics and local price delay measures for non-overlapping subsamples are available from the corresponding author on request. Consistent with overlapping subsample results, it is found that the p-values for AVR statistics and price delay generally vary over time in an oscillatory fashion. Price delay of the large-cap 25

To facilitate cross-country comparison of the results, the regression model includes explanatory variables that are relevant to all the four countries. The relevance of some country-specific variables for explaining return predictability is previously discussed in this paper, but they are not included in the regression model. 26 Since the AVR statistic is one plus weighted sum of autocorrelations of the optimum order, a higher AVR statistic indicates a higher degree of return predictability and vice versa.

30

portfolio is small in magnitude with negligible time variation, which is consistent with that of overlapping subsamples. Overall, the time paths of these efficiency measures strongly corroborate Lo’s (2004) AMH. AVR statistics are also calculated but not presented for oneyear non-overlapping subsamples for the daily returns obtained by applying the 25% pricechange filter. The corresponding AVR test results show that Sri Lanka has significant return predictability in 58% of the subsamples, followed by India (42%), Bangladesh (21%) and Pakistan (21%). This cross-country variation is in line with the overlapping subsample results for daily returns reported in Panel A of Table 4.

6.3 Ranks- and signs-based variance ratio tests As robustness checks, Wright’s (2000) ranks- and signs-based VR tests are also utilized. These are non-parametric tests and based on exact sampling distribution that addresses the problem of size distortion. Let, rt is a stock return series, T is the sample size and xt (rt ) is rank of rt . The standardized ranks (

√

)(

)

(

and

and

( )

) where

can be defined as

( ( )

)

is the standard normal cumulative function. The

test statistics for the ranks-based tests are

(

( where,

(

∑

) ∑

(

∑

) ∑

and

) (

(

)(

)

) (

(

)(

)

)

)

(12)

,

represent the test statistics for the standardized ranks

respectively. While the series

(13) and

,

represents linear transformation of the ranks, the series

is the inverse normal. Both series are standardized to have a sample mean of zero and sample variance of 1.

31

Signs of the returns also can be utilised for an exact variance ratio type test even when conditional heteroskedasticity is present. For a time series, rt, let u(rt, q)= 1(rt >q)-0.5. Therefore, u(rt, 0) will be 1/2 if rt is positive and -1/2 if rt is negative. The signs-based variance ratio statistic can be expressed as follows:

( )

where,

(

(

∑

) ∑

) (

(

)(

)

)

,

(14)

(rt, 0).

The test statistics of ranks- and signs-based tests have the same exact sampling distribution and by simulating these exact sampling distributions, critical values for the tests can be computed. Since qualitatively similar results are obtained for both daily and weekly returns, the results from weekly returns are reported.27 Panel A and Panel B of Table 7 show the results of the ranks- and signs-based tests for aggregate market portfolios, respectively. The tests are applied for four different lags, for example 2, 4, 8, and 16.28 The random walk hypothesis is rejected for all the countries at the 1% significance level for all the four lags. This result holds for both the ranks- and signs-based tests and for all the price-change filters. This finding is at odds with the WBAVR test results for India because the latter provides some evidence of no return predictability in weekly returns. Table 8 reports ranks- and signs-based variance ratio test results for size-sorted portfolios. These tests are applied for all the five size-based portfolios; however, the results are reported only for the largest, medium and smallest quintile portfolios to conserve space. The findings for size-sorted portfolios are similar to those of aggregate market portfolios. The random walk hypothesis is rejected by both the ranks- and signs-based tests for all the four countries. This result holds for all the size-sorted 27 28

We do not report R2 test statistics because the results for R1 and R2 statistics are similar. Similar lags are also used in Hoque, Kim, and Pyun (2007).

32

portfolios and for all the price-change filters. Nevertheless, as reported earlier, the WBAVR test results provide some evidence of no return predictability particularly for large portfolios. Thus, the ranks- and signs-based test results indicate a slight deviation from the WBAVR test results. This discrepancy may arise because non-parametric tests are in general less powerful than their parametric counterparts.

6.4 Daily price delay We also compute daily price delay for value-weighted size-sorted portfolios, and obtain results that are generally analogous to that of weekly price delay. The results are not tabulated in this paper, but they are available from the corresponding author on request. In general, the inverse relationship of price delay with firm size and filter size holds in daily data. We also find that daily price delay is greater than weekly price delay in the case of India. However, this result does not hold for the other three countries. Higher global delay compared to local delay reported in weekly results is not prevalent in daily data. The presence of non-zero price delay even after implementing the 50% price-change filter indicates that infrequent trading is not the sole contributor to price delay in daily returns.

6.5 An alternative measure of volatility In subsection 5.4, we have explained time-varying return predictability in terms of market condition variables, including realised volatility—a non-parametric measure of the return variation. As a robustness check, we now use a parametric measure of volatility, which is conditional volatility, estimated from a GARCH (1,1) model of monthly returns on aggregate market portfolio. We find that GARCH volatility is negatively related to the degree of return predictability in Pakistan and Sri Lanka, but positively related to return

33

predictability in Bangladesh and India. The latter result is inconsistent with the findings obtained from the use of realised volatility. However, our other key findings remain robust to the use of conditional volatility rather than realised volatility. For instance, (i) return predictability decreases with the level of market development and automation of stock exchange; (ii) return predictability is higher in times of financial crises and greater market liquidity; and (iii) return predictability is lower during an equity market downturn. 29

7. Conclusion The South Asian countries have initiated measures to develop their stock markets since the 1990s. However, these markets’ progress with regard to informational efficiency has largely remained unexplored. In general, return predictability studies on emerging markets may be imprecise due to the use of statistical tests with unsatisfactory size and power properties, the thin-trading bias, and ignoring time variation in market efficiency. This study overcomes these limitations. We examine time-varying equity return predictability in the four South Asian countries. The wild bootstrapped automatic variance ratio test and price delay measure are employed for the whole sample period (January 1995 to December 2013) and overlapping subsamples. A regression analysis is conducted to examine how degree of return predictability is linked to changing market conditions and institutional settings. We find strong evidence of time variation in market efficiency. The null hypothesis of no autocorrelation in daily returns is rejected in the whole sample period, but more than 50% of the overlapping subsamples are found to have statistically insignificant autocorrelation (except Sri Lanka). This finding is helpful in reconciling contradictory return predictability

29

We do not report the detail results to conserve space; however they can be obtained from the corresponding author on request.

34

results reported in the literature. For instance, Chaudhuri and Wu (2003) and Islam and Khaled (2005) find support for the random walk hypothesis in Pakistan and Bangladesh, respectively, while Griffin et al. (2010) report significant predictability in these two countries. This result is apparently conflicting but may be a natural consequence of timevariation in return predictability. We further find that return predictability is negatively associated with firm size and trading frequency. We observe that return predictability is related to market and institutional development. The India and Pakistan markets exhibit fewer incidents of significant return predictability compared to Bangladesh and Sri Lanka. Market size, efficient trading mechanism, and shortsale facility may have contributed to this outcome. The regression results also show that degree of return predictability decreases with market development and automation of the stock exchanges. However, contrary to the findings in developed markets, return predictability appears to react positively to liquidity which may indicate greater prevalence of noise traders in these markets (see Bloomfield et al., 2009). In general, we find that the EMH cannot adequately explain the dynamics of equity return predictability. Our finding of oscillation between market efficiency and market inefficiency over time supports Lo’s (2004) adaptive market hypothesis because the AMH asserts that the degree of return predictability varies across market states whereas the EMH is based on an all-or-nothing notion of market efficiency. Lo (2004), however, has not developed any directional hypothesis regarding the relationship between return predictability and the specific market states. Our paper empirically determines these relationships. Future research may focus on developing directional hypotheses between return predictability and diverse market conditions.

35

This study’s results have important policy implications. First, the evidence of market oscillation between efficiency and inefficiency over time suggests that arbitrage opportunities can arise from time to time depending on market conditions. This information may help investors to formulate market timing strategies. Second, given that the sensitivity of return predictability to market conditions varies across our selected countries, it may be possible to devise a country rotation investment approach. Third, given the finding that equity market development is conducive to market efficiency, regulators and policy-makers may consider relaxing regulatory barriers to equity issuance and cross-border capital flows to increase market size and liquidity. In addition, regulatory measures may be undertaken to ensure investor protection, adopt efficient trading mechanisms, facilitate quick diffusion of information, and empower investors to quickly respond to such information.

APPENDIX Table A1 Characteristics of the South Asian equity markets based on the FTSE market assessment matrix FTSE market classification World criteriaBank country rating Creditworthiness Market classification

India Pakistan Lower Middle Lower Investment Speculative Middle Secondary Secondary Emerging Emerging Market and regulatory environment Formal regulatory authority Pass Pass Fair treatment of minority Not met Pass actively monitors stock market Foreign ownership restrictions Pass Pass shareholders Significant restriction on the Pass Pass Free and well-developed Pass repatriation of capital andequity Pass Custody and settlement market income Incidence of failed trade Pass Pass Sufficient competition for high Pass Pass Settlement periodservices T+2 T+2 quality custodian Stock lending is permitted Pass Restricted Dealing landscapes Sufficient liquidity to support Pass Pass Reasonable transaction cost – Pass Pass substantial global investment Short-selling is allowed Restricted Restricted implicit and explicit Efficient trading mechanism Restricted Pass 36

Sri Lanka Lower Speculative Middle Frontier

Bangladesh Lower Speculative Middle Frontier

Pass Restricted Restricted Pass Restricted

Pass Restricted Pass Pass Restricted

Pass Pass T+3 Not met

Pass Pass T+1, T+3 Not met

Restricted Pass Not met Restricted

Not met Pass Not met Restricted

Developed derivatives market

Derivative Restricted Not met Size of market 3,079,225 73,585 7,248 557

Not met

Not met

Market capitalization US $m (at 23,665 34,131 Number of listed companies (at 294 274 31 Dec, 2014) Source: FTSE (2015). Pass: Qualify market classification criteria, Restricted: Partial failure in 31 Dec, 2014) fulfilling the criteria, Not met: Does not meet market classification criteria. World Bank country ranking is based on per capita gross national income.

Table A2 Number of stocks included in the full sample and subsamples Panel A: Total number of stocks included in the sample Full sample 25% Price-change filter Active Deli Total Active Delist Total sted ed India 2444 330 2774 2313 307 2620

50% Price-change filter Acti Deliste Total ve d 206 240 2309 9 124 18 142

Pakista 194 81 275 166 46 212 n Sri 232 21 253 194 05 199 91 01 92 Lanka Bangla 223 61 284 218 49 267 191 21 212 desh Panel B: Average number of stocks included in each size-sorted portfolio Full sample 25% Price-change filter 50% Price-change filter La 4 3 2 S Lar 4 3 2 Smal Larg 4 3 2 rg m ge l e e all India 32 32 329 329 32 287 288 2 28 287 258 25 25 259 9 9 9 8 8 9 9 8 Pakistan 38 39 39 39 38 26 27 2 27 27 20 21 21 21 7 Sri Lanka 37 38 37 38 37 24 24 2 24 24 14 14 14 14 4 Banglade 36 36 36 36 36 29 30 3 30 29 24 25 24 25 sh 0

S m all 25 8 20 14 24

Table A3 Variable definitions Panel A: Continuous variables Market Ratio of market capitalization to industrial production development Liquidity Ratio of trading volume to market capitalization Volatility Log of square root of the sum of daily stock return squares over a month Panel B: Dummy variables Automated This variable takes a value of 1 after implementing automated trading mechanism, trading and 0 before that. The automation date is June 1997 in India, May 1998 in Pakistan, August 1998 in Bangladesh and July 1997 in Sri Lanka. Financial crisis This variable takes a value of 1 for the Asian financial crisis (July 1997 – December 1998) and the global financial crisis (December 2007 – March 2009), and 0 if otherwise. 37

Market downturn

Market downturn refers to a period when the 36-month cumulative return on aggregate market portfolio is negative. This variable takes a value of 1 during market downturn, and 0and if otherwise. Market capitalization, industrial production trading volume data are taken from Global Financial Data (www.globalfinancialdata.com). Dates for automation of trading mechanisms are taken from the websites of respective stock exchanges.

References Abdmoulah, W. (2010). Testing the evolving efficiency of Arab stock markets. International Review of Financial Analysis, 19(1), 25-34. Abeysekera, S. P. (2001). Efficient markets hypothesis and the emerging capital market in Sri Lanka: Evidence from the Colombo Stock Exchange–A note. Journal of Business Finance & Accounting, 28(1‐2), 249-261. Amihud, Y., & Hurvich, C. M. (2004). Predictive regressions: A reduced-bias estimation method. Journal of Financial & Quantitative Analysis, 39(4), 813-841. Anderson, R. M., Eom, K. S., Hahn, S. B., & Park, J.-H. (2013). Autocorrelation and partial price adjustment. Journal of Empirical Finance, 24, 78-93. Bae, K.-H., Ozoguz, A., Tan, H., & Wirjanto, T. S. (2012). Do foreigners facilitate information transmission in emerging markets? Journal of Financial Economics, 105(1), 209-227. Barberis, N., Shleifer, A., & Vishny, R. (1998). A model of investor sentiment. Journal of Financial Economics, 49(3), 307-343. Bekaert, G., & Harvey, C. R. (2002). Research in emerging markets finance: looking to the future. Emerging Markets Review, 3(4), 429-448. Bloomfield, R., O’hara, M., & Saar, G. (2009). How noise trading affects markets: An experimental analysis. Review of Financial Studies, 22(6), 2275-2302. Chang, E. C., Luo, Y., & Ren, J. (2014). Short-selling, margin-trading, and price efficiency: Evidence from the Chinese market. Journal of Banking & Finance, 48, 411-424. Charles, A., Darné, O., & Kim, J. H. (2015). Will precious metals shine? A market efficiency perspective. International Review of Financial Analysis, 41, 284-291. Chaudhuri, K., & Wu, Y. (2003). Random walk versus breaking trend in stock prices: evidence from emerging markets. Journal of Banking & Finance, 27(4), 575-592. Choi, I. (1999). Testing the random walk hypothesis for real exchange rates. Journal of Applied Econometrics, 14(3), 293-308. Chordia, T., Roll, R., & Subrahmanyam, A. (2008). Liquidity and market efficiency. Journal of Financial Economics, 87(2), 249-268. Chordia, T., & Swaminathan, B. (2000). Trading volume and cross‐autocorrelations in stock returns. The Journal of Finance, 55(2), 913-935. Claessens, S., Dell’Ariccia, G., Igan, D., & Laeven, L. (2010). Cross-country experiences and policy implications from the global financial crisis. Economic Policy, 25(62), 267293. Conrad, J., & Kaul, G. (1989). Mean reversion in short-horizon expected returns. Review of Financial Studies, 2(2), 225-240. Cooper, M. J., Gutierrez, R. C., & Hameed, A. (2004). Market states and momentum. The 38

Journal of Finance, 59(3), 1345-1365. Fama, E. F. (1970). Efficient capital markets: A review of theory and empirical work. The Journal of Finance, 25(2), 383-417. Fama, E. F. (1991). Efficient capital markets: II. The Journal of Finance, 46(5), 1575-1617. Fama, E. F., & French, K. R. (1992). The cross‐section of expected stock returns. The Journal of Finance, 47(2), 427-465. FTSE (2015). FTSE market assessment matrix. Retrieved June 3, 2015 from http://www.ftse.com/products/downloads/Asia-Pacific_latest.pdf Griffin, J. M., Kelly, P. J., & Nardari, F. (2010). Do market efficiency measures yield correct inferences? A comparison of developed and emerging markets. Review of Financial Studies, 23(8), 3225-3277. Gu, A. Y., & Finnerty, J. (2002). The evolution of market efficiency: 103 years daily data of the Dow. Review of Quantitative Finance & Accounting, 18(3), 219-237. Harvey, C. R. (1995). Predictable risk and returns in emerging markets. Review of Financial Studies, 8(3), 773-816. Hendershott, T., & Seasholes, M. S. (2014). Liquidity provision and stock return predictability. Journal of Banking & Finance, 45, 140-151. Hong, H., Lim, T., & Stein, J. C. (2000). Bad news travels slowly: Size, analyst coverage, and the profitability of momentum strategies. The Journal of Finance, 55(1), 265-295. Hong, H., & Stein, J. C. (1999). A unified theory of underreaction, momentum trading, and overreaction in asset markets. The Journal of Finance, 54(6), 2143-2184. Hoque, H. A., Kim, J. H., & Pyun, C. S. (2007). A comparison of variance ratio tests of random walk: A case of Asian emerging stock markets. International Review of Economics & Finance, 16(4), 488-502. Hou, K. (2007). Industry information diffusion and the lead-lag effect in stock returns. Review of Financial Studies, 20(4), 1113-1138. Hou, K., & Moskowitz, T. J. (2005). Market frictions, price delay, and the cross-section of expected returns. Review of Financial Studies, 18(3), 981-1020. Islam, A., & Khaled, M. (2005). Tests of weak‐form efficiency of the Dhaka stock exchange. Journal of Business Finance & Accounting, 32(7‐8), 1613-1624. Ito, M., Noda, A., & Wada, T. (2014). International stock market efficiency: a non-bayesian time-varying model approach. Applied Economics, 46(23), 2744-2754. Ito, M., Noda, A., & Wada, T. (2016). The evolution of stock market efficiency in the US: a non-Bayesian time-varying model approach. Applied Economics, 48(7), 621-635. Johnson, S., Boone, P., Breach, A., & Friedman, E. (2000). Corporate governance in the Asian financial crisis. Journal of Financial Economics, 58(1), 141-186. Katusiime, L., Shamsuddin, A., & Agbola, F. W. (2015). Foreign exchange market efficiency and profitability of trading rules: Evidence from a developing country. International Review of Economics & Finance, 35, 315-332. Khediri, K. B., & Charfeddine, L. (2015). Evolving efficiency of spot and futures energy markets: A rolling sample approach. Journal of Behavioral and Experimental Finance, 6, 67-79. Kim, J. H. (2009). Automatic variance ratio test under conditional heteroskedasticity. Finance Research Letters, 6(3), 179-185. Kim, J. H., & Shamsuddin, A. (2008). Are Asian stock markets efficient? Evidence from new multiple variance ratio tests. Journal of Empirical Finance, 15(3), 518-532. Kim, J. H., & Shamsuddin, A. (2015). A closer look at return predictability of the US stock market: evidence from new panel variance ratio tests. Quantitative Finance, 15(9), 1501-1514. Kim, J. H., Shamsuddin, A., & Lim, K.-P. (2011). Stock return predictability and the adaptive 39

markets hypothesis: Evidence from century-long US data. Journal of Empirical Finance, 18(5), 868-879. Kim, E. H., & Singal, V. (2000). Stock Market Openings: Experience of Emerging Economies. Journal of Business, 73(1), 25-66. Kumar, N. (2012). Foreign capital flows and development: lessons from South Asian Experiences. In R. Jha (Eds.), Routledge Handbook of South Asian Economics (1st ed., pp. 238-257). Abingdon, Oxon: Routledge. Lagoarde-Segot, T., & Lucey, B. M. (2008). Efficiency in emerging markets—Evidence from the MENA region. Journal of International Financial Markets, Institutions and Money, 18(1), 94-105. Lai, S., Ng, L., & Zhang, B. (2014). Does PIN affect equity prices around the world? Journal of Financial Economics, 114(1), 178-195. Lim, K.-P., Brooks, R. D., & Kim, J. H. (2008). Financial crisis and stock market efficiency: Empirical evidence from Asian countries. International Review of Financial Analysis, 17(3), 571-591. Lim, K.-P., Luo, W., & Kim, J. H. (2013). Are US stock index returns predictable? Evidence from automatic autocorrelation-based tests. Applied Economics, 45(8), 953-962. Lo, A. W. (2004). The adaptive markets hypothesis: Market efficiency from an evolutionary perspective. Journal of Portfolio Management,30(5), 15-29. Lo, A. W. (2005). Reconciling efficient markets with behavioral finance: the adaptive markets hypothesis. Journal of Investment Consulting, 7(2), 21-44. Lo, A. W., & MacKinlay, A. C. (1988). Stock market prices do not follow random walks: Evidence from a simple specification test. Review of Financial Studies, 1(1), 41-66. Lo, A. W., & MacKinlay, A. C. (1990). An econometric analysis of nonsynchronous trading. Journal of Econometrics, 45(1), 181-211. Lo, A. W., & MacKinlay, A. C. (2011). A non-random walk down Wall Street. Princeton: Princeton University Press. Martens, M. (1998). Price discovery in high and low volatility periods: open outcry versus electronic trading. Journal of International Financial Markets, Institutions and Money, 8(3), 243-260. Mech, T. S. (1993). Portfolio return autocorrelation. Journal of Financial Economics, 34(3), 307-344. Mehta, P. S., & Yadav, N. (2013). Trade relations: Some predictions and lessons. In A. Najam, & M. Yusuf (Eds.), South Asia 2060: Envisioning regional futures (1st ed., pp. 153-159). London, England: Anthem Press. Mitton, T. (2002). A cross-firm analysis of the impact of corporate governance on the East Asian financial crisis. Journal of Financial Economics, 64(2), 215-241. Morck, R., Yeung, B., & Yu, W. (2000). The information content of stock markets: why do emerging markets have synchronous stock price movements? Journal of Financial Economics, 58(1), 215-260. Naidu, G., & Rozeff, M. S. (1994). Volume, volatility, liquidity and efficiency of the Singapore Stock Exchange before and after automation. Pacific-Basin Finance Journal, 2(1), 23-42. Neal, R. (1987). Potential competition and actual competition in equity options. The Journal of Finance, 511-531. Nelson, C. R., & Kim, M. J. (1993). Predictable stock returns: The role of small sample bias. The Journal of Finance, 48(2), 641-661. Noda, A. (2016). A test of the adaptive market hypothesis using a time-varying AR model in Japan. Finance Research Letters, 17, 66-71. Patro, D. K., & Wu, Y. (2004). Predictability of short-horizon returns in international equity 40

markets. Journal of Empirical Finance, 11(4), 553-584. Poshakwale, S. (2002). The random walk hypothesis in the emerging Indian stock market. Journal of Business Finance & Accounting, 29(9‐10), 1275-1299. Romero-Torres, J., Wells, S., & Selwyn-Khan, S. (2013). Development of Capital Markets in Member Countries of the South Asian Association for Regional Cooperation. Mandaluyong City, Philippines: Asian Development Bank. Sahoo, P., Nataraj, G., & Dash, R. K. (2006). Foreign direct investment in South Asia: Policy, trends, impact and determinants. New Delhi, India: Springer India. Shamsuddin, A., & Kim, J. H. (2010). Short‐Horizon Return Predictability in International Equity Markets. Financial Review, 45(2), 469-484. Shiller, R. J. (1983). Do Stock Prices Move Too Much to be Justified by Subsequent Changes in Dividends? Reply. The American Economic Review, 73(1), 236-237. Shleifer, A., & Vishny, R. W. (1997). A survey of corporate governance. The Journal of Finance, 52(2), 737-783. Simon, H. A. (1955). A behavioral model of rational choice. The Quarterly Journal of Economics, 69(1), 99-118. Urquhart, A., Gebka, B., & Hudson, R. (2015). How exactly do markets adapt? Evidence from the moving average rule in three developed markets. Journal of International Financial Markets, Institutions and Money, 38, 127-147. Urquhart, A., & Hudson, R. (2013). Efficient or adaptive markets? Evidence from major stock markets using very long run historic data. International Review of Financial Analysis, 28, 130-142. Urquhart, A., & McGroarty, F. (2014). Calendar effects, market conditions and the Adaptive Market Hypothesis: Evidence from long-run US data. International Review of Financial Analysis, 35, 154-166. Welch, I., & Goyal, A. (2008). A comprehensive look at the empirical performance of equity premium prediction. Review of Financial Studies, 21(4), 1455-1508. World Bank (2016). World development indicators. Retrieved August 15, 2016 from http://data.worldbank.org/products/wdi Wright, J. H. (2000). Alternative variance-ratio tests using ranks and signs. Journal of Business & Economic Statistics, 18(1), 1-9.

Fig 1 Wild Bootstrapped p-values for aggregate market and size-sorted portfolios from overlapping subsamples India Pakistan Panel A. Aggregate market portfolios

Sri Lanka

Bangladesh

1.0

1.0

1.0

1.0

0.8

0.8

0.8

0.8

0.6

0.6

0.6

0.6

0.4

0.4

0.4

0.4

0.2

0.2

0.2

0.2

0.0 95 96

97

98

99

00

01

Full sample

02

03

04

05

25% Filter

0.0 06 07 08 09 10 11 95 96 97 98 99 00 01

50% Filter

Full sample

02 03 04

0.0 05 069507960897099810991100

25% Filter

41

50% Filter

01

Full sample

02

03

04

05

25% Filter

0.0 0695 0796 0897 09 98 10 99 11 00 01 02 03 04 05 06 07 08 09 10 11

50% Filter

Full sample

25% Filter

50% Filter

Panel B. Size-sorted portfolios 1.0

1.0

1.0

1.0

0.8

0.8

0.8

0.8

0.6

0.6

0.6

0.6

0.4

0.4

0.4

0.4

0.2

0.2

0.2

0.2

0.0

0.0 95 96 97 98 99 00 01 02 03 04 05 06 950796089709 9810 9911 00

Large

Medium

Small

01

02

03

Large

04

05

0.0 069507960897099810991100

Medium

Small

01

02

Large

03

04

05

Medium

0.0 0695 0796 0897 09 98 10 99 11 00 01 02 03 04 05 06 07 08 09 10 11

Small

Large

Medium

Small

This figure represents p-values of the WBAVR test applied to overlapping subsamples of aggregate market and size-sorted portfolio returns. The total number of overlapping windows is 201. The window period is 24-months and it advances monthly. The shaded area in each graph indicates statistical significance at the 5% level. The test is applied to weekly returns. In the case of size-sorted portfolios, p-values are computed for the subsample obtained using the 25% price-change filter.

Fig 2 Local price delay for size-sorted portfolios for overlapping subsamples India

Pakistan

.36

.32

.32

.28

.28

Sri Lanka

Bangladesh

.20

1.0

.16

0.8

.12

0.6

.08

0.4

.04

0.2

.24

.24

.20

.20

.16 .16

.12

.12

.08

.08

.04

.04 .00

.00 .00 0.0 95 96 97 98 99 00 01 02 03 04 05 06 9507 96 08 97 09 98 10 99 11 00 01 02 03 04 05 06 07 95 08 96 97 09 98 10 99 11 00 01 02 03 04 05 06 09 99 10 00 11 01 02 03 04 05 06 07 08 09 10 11 95 07 96 08 97 98

Large

Medium

Small

Large

Medium

Small

Large

Medium

Small

Large

This figure reports local price delay for overlapping subsamples for size-sorted portfolios. Although the delay measure is calculated for all quintile portfolios, we present results only for the largest, medium and smallest quintile portfolios for brevity. We compute delay as ( / ) following Hou and Moskowitz (2005). The unrestricted model involves a regression of current size-sorted portfolio returns on contemporaneous and lagged market returns. The restricted model, on the other hand, involves a regression of current size-based portfolio returns on only contemporaneous market returns. R2 is obtained from rolling regression. The window length is 104 weeks (2 years) and windows advance on a 4-week basis.

Table 1 Mean returns on aggregate market and size-sorted portfolios Panel A: Mean returns on market portfolios (%) India Pakistan Sri Lanka Ful l sa

25 % Fi

50 % Filt

Ful l sa

25 % Fil

50 % Filt

Ful l sa

25 % Fi

42

Bangladesh 50 % Filt

Ful l sa

25 % Fil

50 % Fi

Medium

Small

mp le 0.0 9 0.4 2

Daily returns Weekly returns

lte r 0. 09 0. 44

er 0.0 9 0.4 3

mp le 0.1 0 0.4 8

ter

er

0. 15 0. 48

0.0 8 0.4 0

mp le 0.1 0 0.4 6

lte r 0. 16 0. 49

er 0.0 8 0.3 4

Panel B: Mean returns on size-sorted portfolios (%) In Pakista di n a La rge

mp le 0.0 8 0.3 5

ter 0. 08 0. 36

lte r 0. 09 0. 38

Sri Lanka

Banglades h

4

3

2

Sm all

La rg e

4

3

2

Sm all

La rge

4

3

2

S m all

La rg e

4

3

2

S m all

0. 1 1 0. 1 0 0. 1 3

0. 14

0. 05 0. 05

0. 21

0. 05

0. 0 7 0. 0 6 0. 0 6

0. 0 8 0. 0 7 0. 0 8

0. 08

0. 17

0. 0 7 0. 0 9 0. 1 0

0. 5 1 0. 4 5 0. 5 8

0. 64

0. 23 0. 22

0. 96

0. 21

0. 3 0 0. 2 5 0. 2 5

0. 3 5 0. 3 2 0. 3 6

0. 34

0. 77

0. 3 0 0. 3 9 0. 4 3

Daily returns Full sampl e 25% Filter

0.0 3

0.0 4

0. 07

0.1 1

0.1 6

0. 04

0.0 6

0.0 7

0. 10

0.1 6

0.0 5

0. 07

0. 09

0.0 4

0.0 4

0. 06

0.1 1

0.1 9

0. 04

0.0 6

0.0 7

0. 08

0.1 6

0.0 5

0. 08

0. 08

50% Filter

0.0 4

0.0 4

0. 05

0.1 0

0.1 7

0. 03

0.0 6

0.0 7

0. 09

0.1 1

0.0 7

0. 10

0. 10

0. 11 0. 11

Weekly returns Full sampl e 25% Filter

0.1 5

0.1 9

0. 33

0.5 1

0.7 5

0. 19

0.2 9

0.3 1

0. 45

0.7 4

0.2 2

0. 31

0. 41

0.1 8

0.1 9

0. 31

0.5 3

0.9 0

0. 18

0.2 7

0.3 5

0. 40

0.7 3

0.2 1

0. 36

0. 36

50% Filter

0.1 8

0.1 9

0. 26

0.4 6

0.8 4

0. 16

0.2 7

0.3 2

0. 42

0.5 2

0.3 2

0. 44

0. 46

0. 48 0. 49

Panel A and Panel B report mean returns (%) on aggregate market and size-sorted portfolios, respectively. The full sample includes all stocks, the 25% price-change filter includes the stocks that are traded on at least 25% of the total trading days in a year, and the 50% price-change filter includes the stocks that are traded on at least 50% of the total trading days in a year. Size-sorted portfolios are constructed on the basis of year-end market capitalization. Weekly returns are calculated by adding log daily returns from each Wednesday to Tuesday to avoid daily seasonality in returns.

Table 2 AVR statistics and wild bootstrapped p-values for returns on aggregate market portfolios India Full 25% sample filter

50% filter

Panel A: Daily returns AVR 6.0 6.11 6.08 statistic 8* *** *** ** p-value 0.0 0.00 0.00 0 Panel B: Weekly returns AVR 0.1 0.22 0.22 statistic 8 p-value 0.7 0.72 0.72

Pakistan Full 25% sampl filter e

50% filter

Sri Lanka Full 25% sampl filter e

50% filter

Bangladesh Full 25% 50% sampl filter filter e

7.11* **

6.93 ***

6.65 ***

10.35* **

9.53 ***

7.92 ***

6.35* **

6.35 ***

5.93 ***

0.00

0.00

0.00

0.00

0.00

0.00

0.00

0.00

0.00

3.51* ** 0.00

3.36 ** 0.01

3.35 ** 0.01

5.01** * 0.00

4.13 *** 0.00

3.14 ** 0.01

3.75* * 0.03

3.67 ** 0.01

3.36 ** 0.02

43

4 The table displays wild bootstrapped automatic variance ratio (AVR) statistics and respective pvalues for the full sample and the two subsamples obtained by implementing price-change filters. The full sample includes all stocks, the 25% price-change filter includes the stocks that are traded on at least 25% of the total trading days in a year and the 50% price-change filter includes the stocks that are traded on at least 50% of the total trading days in a year. Results for daily and weekly returns are on Panels A and B, respectively. While implementing the AVR test, the number of bootstrap iterations is set to 500 (Kim, 2009). ***, ** and * represent statistical significance at the 1%, 5% and 10% levels, respectively.

Table 3 AVR statistics and wild bootstrapped p-values for size-sorted portfolio returns Full sample Lar ge

4

25% pricechange filter

50% pricechange filter

3

2

Sma ll

Lar ge

4

3

2

Sma ll

Lar ge

4

3

2

Sma ll

Panel A: Daily returns India

Pakist an

Sri Lanka

Bangl adesh

AV R pvalu e AV R pvalu e AV R pvalu e AV R pvalu e

5.68 *** 0.00

17.2 2*** 0.00

24.3 8*** 0.00

32.7 0*** 0.00

51.7 8*** 0.00

5.60 *** 0.00

16.28 *** 0.00

21.99 *** 0.00

30.0 0*** 0.00

44.0 5*** 0.00

5.62 *** 0.00

15.15 *** 0.00

19.99 *** 0.00

26.7 4*** 0.00

36.5 1*** 0.00

6.94 *** 0.00

10.6 1*** 0.00

12.6 8*** 0.00

12.3 6*** 0.00

15.7 9*** 0.00

6.65 *** 0.00

9.22* ** 0.00

10.29 *** 0.00

10.5 3*** 0.00

12.2 5*** 0.00

6.14 *** 0.00

7.82* ** 0.00

7.68* ** 0.00

7.91 *** 0.00

9.28 ** 0.01

8.59 *** 0.00

8.99 *** 0.00

9.58 *** 0.00

8.75 *** 0.00

12.1 0*** 0.00

5.17 *** 0.00

6.38* ** 0.00

5.85* * 0.01

6.06 *** 0.00

6.66 *** 0.00

2.94 *** 0.00

3.09* * 0.04

3.70* * 0.02

3.19 ** 0.01

4.41 *** 0.00

4.20 ** 0.03

5.52 *** 0.00

7.70 *** 0.00

11.3 7*** 0.00

17.4 9*** 0.00

3.8* * 0.03

6.51* ** 0.00

7.53* ** 0.00

8.76 *** 0.00

13.1 3*** 0.00

4.08 ** 0.04

5.99* ** 0.00

7.10* ** 0.00

6.74 *** 0.00

8.86 *** 0.00

4.90 *** 0.00

6.19 *** 0.00

8.76 *** 0.00

15.1 9*** 0.00

0.17

4.76* ** 0.00

5.64* ** 0.00

7.77 *** 0.00

12.8 0*** 0.00

0.17 0.77

4.67* ** 0.00

5.27* ** 0.00

6.78 *** 0.00

10.0 *** 0.00

6.73 *** 0.00

8.37 *** 0.00

6.99 *** 0.00

2.44

4.55* ** 0.00

4.82 *** 0.00

7.48 *** 0.00

1.68

0.11

3.76* ** 0.00

1.28

0.17

5.5* ** 0.00

0.32

0.11

3.82* ** 0.00

4.17 *** 0.00

5.01 *** 0.00

8.59 *** 0.00

8.99 *** 0.00

9.58 *** 0.00

10.5 0*** 0.00

11.1 0*** 0.00

6.17 *** 0.00

6.38* ** 0.00

6.90* * 0.01

7.06 *** 0.00

7.66 *** 0.00

4.94 *** 0.00

5.09* * 0.04

6.70* * 0.02

6.98 ** 0.01

7.48 ** 0.03

4.20 ** 0.03

5.50 *** 0.00

7.70 *** 0.00

11.4 *** 0.00

17.4 9*** 0.00

3.8* * 0.03

6.51* ** 0.00

7.53* ** 0.00

8.76 *** 0.00

13.1 3*** 0.00

3.08 * 0.08

5.99* ** 0.00

7.10* ** 0.00

6.74 *** 0.00

8.86 *** 0.00

Panel B: Weekly returns India

Pakist an

Sri Lanka

Bangl adesh

AV R pvalu e AV R pvalu e AV R pvalu e AV R pvalu e

0.17 0.70

2.64

0.78

Panel A and Panel B present AVR statistics and p-values for daily and weekly returns of size-sorted portfolios. AVR statistics are calculated for the full sample and the subsamples obtained by implementing price-change filters. The full sample includes all stocks; the 25% price-change filter includes the stocks that are traded on at least 25% of the total trading days in a year. Furthermore, the 50% price-change filter includes the stocks that are traded on at least 50% of the total trading days in a year. Size-sorted portfolios are constructed on the basis

44

of year-end market capitalization. While implementing the AVR test, the number of bootstrap iterations is set to 500 (Kim, 2009). ***, ** and * represent statistical significance at the 1%, 5% and 10% levels, respectively.

Table 4 Proportion of overlapping subsamples with statistically significant AVR India Pakistan Full 25% 50% Full 25% sampl Filte Filte sampl Filte e r r e r Panel A: Aggregate market portfolios Daily 41.29 44.7 45.7 29.85 29.3 returns 8 3 5 Weekl 5.97 6.47 5.97 25.87 24.3 y 8 returns Panel B: Size-sorted portfolios Daily returns Large 23.38 23.3 22.8 22.89 23.8 8 9 8 Mediu 87.06 76.6 65.1 74.13 61.6 m 2 7 9 Small 100 100 99.5 83.58 76.1 2 Weekly returns Large 5.47 4.98 4.48 24.38 25.3 7 Mediu 54.73 50.2 40.3 42.79 33.3 m 5 3 Small 92.54 91.0 82.9 51.74 51.2 4 4

Sri Lanka Full 25% 50% sampl Filte Filte e r r

Bangladesh Full 25% 50% sampl Filte Filte e r r

25.8 7 20.9 0

76.12

19.4

50% Filte r

70.6 5 35.8 2

68.1 6 28.8 6

22.39

51.74

45

16.92

61.6 9 55.2 2

71.14

50

61.19

45.5

38.0 2 25.6 2 32.2 3

22.8 9 23.8 8 19.4 0

31.84

25.8 7 25.8 7 27.8 6

19.6 6 18.8 0 14.5 3

39.30

34.83 50.75

22.8 9 12.9 3

21.3 9 11.9 4

15.9 2 42.7 9 77.6 1

10.9 5 38.3 1 58.2 1

7.46

6.96

6.47

22.39

20.9 0 20.4 0

15.9 2 16.4 2

12.93

52.74 81.09

20.89

The table reports the percentages of total number of overlapping windows with significant AVR at the 5% level for the full sample and the two subsamples obtained through price-change filters. Results for aggregate market and size-sorted portfolios are on panel A and panel B, respectively. The total number of overlapping windows is 201. Window length is 24 months and it advances monthly. The WBAVR test is applied to each window for daily and weekly frequencies. For conducting the AVR test, the number of bootstrap iterations is set to 500 (Kim, 2009). The test is implemented on five size-sorted portfolios; however, results only for the largest, medium and smallest quintile portfolios are reported for simplicity.

45

Table 5 Price delay Portfo lio

India Local delay

Small 2 3 4 Large

0.1814 0.0720 0.0373 0.0181 0.0001

Global delay 0.1799 0.0719 0.0377 0.0201 0.0007

Small 2 3 4 Large

0.1312 0.0569 0.0307 0.0171 0.0001

0.1304 0.0563 0.0312 0.0186 0.0004

Small 2 3 4 Large

0.0911 0.0445 0.0263 0.0141 0.0002

0.0896 0.0441 0.0270 0.0156 0.0005

Pakistan Local delay

Sri Lanka Local Global delay delay

Global delay Full sample 0.1201 0.1244 0.0134 0.0973 0.0958 0.0121 0.0360 0.0377 0.0037 0.0113 0.0115 0.0066 0.0004 0.0004 0.0013 25% Price-change filter 0.0429 0.0480 0.0033 0.0290 0.0357 0.0058 0.0096 0.0101 0.0015 0.0087 0.0048 0.0017 0.0003 0.0004 0.0009 50% Price-change filter 0.0141 0.0273 0.0030 0.0072 0.0116 0.0030 0.0007 0.0012 0.0014 0.0067 0.0010 0.0013 0.0002 0.0003 0.0009

Bangladesh Local Global delay delay

0.0229 0.016 0.0066 0.0086 0.0009

0.0358 0.0101 0.0016 0.0023 0.0019

0.1265 0.0736 0.0061 0.0038 0.0013

0.0109 0.0093 0.0012 0.0067 0.0006

0.0161 0.0054 0.0007 0.0019 0.0014

0.0334 0.0173 0.0045 0.0018 0.0012

0.0097 0.0062 0.0039 0.0059 0.0008

0.0105 0.0047 0.0042 0.0011 0.0006

0.0198 0.0073 0.0006 0.0011 0.0007

Price delay is calculated from weekly returns for the full sample and the two subsamples obtained by implementing price-change filters. The full sample includes all stocks, the 25% price-change filter includes the stocks that are traded on at least 25% of the total trading days in a year, and the 50% price-change filter includes the stocks traded on at least 50% of the total trading days in a year. We compute delay as ( / ) following Hou and Moskowitz (2005). The unrestricted model involves a regression of current size-sorted portfolio returns on contemporaneous and lagged market returns. The restricted model, on the other hand, involves a regression of current size-based portfolio returns on only contemporaneous market returns. Local delay measures consider the speed of incorporation of local market information into prices, while global delay considers both local and global market information.

Table 6 Regression Results Dependent Variable: | India

|

Pakistan

Variable

Coeffi cient

tStatisti c

Constant

-1.243

(1.208)

Market development

-0.046

(0.602)

Liquidity

2.950

Volatility

-0.765

Eff ect siz e

Sri Lanka

Coeffi cient

tStatisti c

-7.364

(3.922)

-0.902

(4.801 ***)

0.1 4 1.1 9

(3.478* **) (6.172 ***)

(-

-

-3.416

(-

6.736

46

Eff ect siz e

Bangladesh

Coeffi cient

tStatisti c

Eff ect siz e

11.655

(5.218)

0.3 5 0.7 5

-0.459

(1.412)

243.5

(5.987 ***)

0.1 1 0.7 4

-

-4.415

(-

-

Coeffi cient

tStatis tic

-2.089

(1.346 ) (1.56 1)

0.2 5

(1.222 ) (-

0.1 2 -

0.814

-9.912

-1.216

Eff ect siz e

2.166* *) (2.360* *) (0.254) (4.618* **)

0.2 3

0.3 2

0.3 3

-0.441

Financial crisis dummy Market downturn dummy R2

-0.057

0.435

0.329

0.608

0.165

F-statistic

24.876 ***

15.864 ***

50.133 ***

6.121* **

0.953 0.644

(2.098 **) (1.451)

-0.534

6.600* **) (1.858* ) (5.076 ***) (0.754)

Automated trading dummy

-0.561

-0.583

5.076* **) (1.533)

1.038 -0.252

-0.860

0.383 0.252

1.906 *) (1.591 *) (1.82 6*) (1.30 3)

*** , ** and * indicate statistical significance at the 1% , 5 % and 10% levels, respectively. Figures in parentheses are Newey-West heteroskedasticity and autocorrelation consistent (HAC) t-statistics. The absolute value of statistics is taken as the dependent variable and market condition variables are used as independent variables. statistics are generated by applying the WBAVR test on 24-month overlapping windows, which advance monthly. statistics for the regression model are calculated from weekly returns on aggregate market portfolio with the imposition of the 25% price-change filter. Market development is the ratio of market capitalization to industrial production, liquidity is the ratio of trading volume to market capitalization, and volatility is calculated as log of the square root of the sum of squared daily returns over a month. These continuous variables are estimated by taking the moving average of the monthly observations for a respective overlapping window. Automated trading, financial crisis and market downturn are dummy variables, taking a value of 1 for a particular market condition, and 0 if otherwise. Since the dummy variables cannot be constructed for overlapping windows, if a specific market condition prevails for at least 4 consecutive months in a window then the window takes a value of 1, and 0 if otherwise. Automated trading dummy is equal to 1 after the automation of trading mechanism, and 0 before that. The financial crisis dummy takes a value of 1 for the Asian financial crisis and global financial crisis, and 0 if otherwise. Market downturn dummy takes value of 1 when cumulative monthly return is negative for a particular window, and 0 if otherwise. In order to calculate effect size, we first multiply regression coefficient with standard deviation of the |. corresponding variable, then the resulting estimate is expressed as a percentage of average |

47

0.1 7

Table 7 Results of Wright’s (2000) ranks- and signs-based variance ratio tests for returns on aggregate market portfolios India Pakistan Sri Lanka Bangladesh Full 25% 50% Full 25% 50% Full 25% 50% Full 25% sampl Filter Filter sample Filter Filter sample Filter Filter sample Filter e Panel A: Ranks-based test result L a R1 R1 R1 R1 R1 R1 R1 R1 R1 R1 R1 g 2 -14.29 -13.05 15.33 15.3 15.3 14.3 14.4 13.2 14.1 -12.94 13.1 4 6 2 0 3 0 4 4 -11.16 -10.22 11.52 11.5 11.5 11.1 11.1 10.2 10.5 -10.62 10.6 2 4 8 9 7 0 8 8 -8.51 -8.49 -8.50 -8.19 -8.19 -8.24 -7.72 -7.74 -7.82 -8.06 -8.10 1 -5.97 -5.96 -5.96 -5.92 -5.92 -5.94 -5.54 -5.53 -5.56 -5.56 -5.59 6 Panel B: Signs-based test result L S1 S1 S1 S1 S1 S1 S1 S1 S1 S1 S1 a g 2 -9.72 -9.98 -9.05 -8.92 -9.30 10.52 10.6 10.6 10.3 -8.38 -8.70 5 5 7 4 -8.50 -8.54 -8.54 -7.60 -7.74 -7.84 -6.77 -6.50 -6.43 -7.63 -7.67 8 -6.29 -6.34 -6.34 -5.71 -5.70 -5.70 -4.97 -4.80 -4.94 -5.95 -6.02 1 -4.28 -4.33 -4.33 -4.08 -4.18 -4.19 -3.59 -3.32 -3.68 -4.06 -4.10 6

50% Filter

R1 13.2 5 10.7 3 -8.12 -5.64 S1

-8.57 -7.46 -5.84 -4.19

Panel A and Panel B document ranks- and signs-based test results, respectively. The tests are implemented for four different lags (2, 4, 8 and16). We apply the test on the full sample and two subsamples obtained from the 25% and 50% price-change filters. All the test statistics are significant at the 1% level.

Table 8 Results of Wright’s (2000) ranks- and signs-based variance ratio tests for size-sorted portfolio returns India Pakistan Sri Lanka Larg Mediu Sma Larg Mediu Sma Larg Mediu Panel A:e Ranks-based m test ll resulte m ll e m La R1 R1 R1 R1 R1 R1 R1 R1 g Full sample 2 -15.5 -11.1 -13.2 -13.2 -13.7 -13.0 4 -11.5 -10.1 10.0 -9.5 -10.6 -11.0 14.2 -10.3 -10.5 11.4 8 -8.5 -8.2 -7.9 -8.1 -8.2 -8.4 -7.8 -7.9 16 -6.0 -5.9 -5.8 -6.0 -6.1 -6.1 -5.6 -5.9 50% Price-change filter 2 -15.4 -11.7 -14.8 -13.6 -12.6 -10.9 4 -11.5 -10.4 10.9 -9.9 -11.4 -11.0 14.4 -9.0 -8.4 11.4 48

Bangladesh Smal Larg Mediu l e m R1 R1 R1

Sma ll R1

-14.0 -10.7 -8.2 -6.0

13.3 10.9 -8.2 -5.7

-11.9 -10.1 -7.8 -5.5

11.7 10.6 -7.8 -5.6

-11.9 -9.0

13.7 -11

-12.1 -10.3

13.8 11.2

8 -8.5 -8.3 -8.0 -8.4 16 -6.0 -6.0 -5.8 -6.1 Panel B: Signs-based test result S1 S1 S1 S1 2 4 8 16

-11 -8.5 -6.4 -4.3

-8.9 -7.6 -6.1 -4.4

-7.5 -7 -5.4 -4

-9.6 -7.1 -5.4 -3.9

2 4 8 16

-11.3 -8.6 -6.3 -4.3

-8.9 -7.7 -6.1 -4.4

-7.8 -7.1 -5.6 -3.9

-10.1 -7.5 -5.5 -3.9

-8.4 -6.2

-8.4 -6.0

S1

-6.6 -4.7

S1 S1 Full sample -8.6 -9.6 10.7 -7.3 -8.3 -7.1 -5.8 -6.1 -5.3 -4.6 -4.4 -3.8 50% Price-change filter -9.4 -8.9 10.8 -7.6 -8.5 -6.4 -5.8 -6.2 -4.4 -4.4 -4.4 -3.3

-6.3 -4.6

-6.6 -4.9

-8.3 -5.8

-8.2 -5.8

-8.4 -6.1

S1

S1

S1

S1

S1

-9.8 -7.4 -5.4 -4

-9.5 -7.3 -5.6 -3.9

-8.2 -7.4 -6.0 -4.6

-7.2 -6.6 -5.7 -4.3

-8.5 -7.2 -5.4 -4.1

-7.5 -5.1 -3.8 -2.4

-9.2 -7 -5.1 -3.9

-8.5 -7.3 -5.6 -4.0

-8.6 -7.5 -6.1 -4.3

-8.2 -7.2 -5.7 -4.2

Ranks- and signs-based test results are on panel A and panel B, respectively. We apply the test on the full sample and two subsamples obtained from the 25% and 50% pricechange filters, but the results pertaining to the 25% price-change filter are not reported due to space constraints. The tests are conducted for five size-sorted portfolios; however, results are reported only for the largest, medium and smallest quintile portfolios to conserve space. Four alternative lags (2,4,8, and 16) are used to conduct the tests. All the test statistics are significant at the 1% level.

49