Nowcasting BRIC+M in real time

International Journal of Forecasting 33 (2017) 915–935

Contents lists available at ScienceDirect

International Journal of Forecasting journal homepage: www.elsevier.com/locate/ijforecast

Nowcasting BRIC+M in real time Tatjana Dahlhaus, Justin-Damien Guénette, Garima Vasishtha * Bank of Canada, International Economic Analysis Department, 234 Wellington Street, Ottawa, ON K1A 0G9, Canada

article

info

Keywords: Dynamic factor model Nowcasting Real-time data Emerging markets

a b s t r a c t Given the growing importance of emerging market economies (EMEs) in driving global GDP growth, timely and accurate assessments of current and future economic activity in EMEs are important for policy-makers not only in these countries, but also in advanced economies. This paper uses state-of-the-art dynamic factor models (DFMs) to nowcast real GDP growth for Brazil, Russia, India, China, and Mexico (‘‘BRIC+M’’). The DFM framework is particularly suitable for EMEs, as it enables the efficient handling of data series that are characterized by different publication lags, frequencies, and sample lengths. It also allows the extraction of model-based ‘‘news’’ from a data release and the assessment of the impact of such ‘‘news’’ on nowcast revisions. Overall, we find that the DFMs generally display a good directional accuracy and provide reliable nowcasts for GDP growth. Furthermore, the ‘‘news’’ pertaining to domestic indicators is the main driver of changes in nowcast revisions, while exogenous variables play a relatively minor role. © 2017 International Institute of Forecasters. Published by Elsevier B.V. All rights reserved.

1. Introduction The global economic landscape has changed considerably since the turn of the century, with the share of major emerging-market economies, namely Brazil, Russia, India, China and Mexico (henceforth ‘‘BRIC+M’’), growing from about 21% of global GDP in 2000 to nearly 32% in 2014, in terms of purchasing power parity (PPP) (Fig. 1). The trade and financial linkages between advanced and emerging-market economies (EMEs) have also become much stronger, and the notion that advanced economies have become more dependent on the demand from relatively fast-growing EMEs has gained ground. Thus, timely and accurate assessments of the economic activity of the BRIC+M group are of great importance for policymakers not only in these economies, but also in advanced economies. Although the business cycles of major emerging markets share some common characteristics, growth dynamics author. * Corresponding E-mail addresses: [email protected] (T. Dahlhaus), [email protected] (J. Guénette), [email protected] (G. Vasishtha).

also show strong idiosyncratic fluctuations (Fig. 2). This probably reflects not only cyclical factors, but also important differences in economic structure across countries, such as size, the degree of financial and trade openness, the dependence on commodity exports vs. imports, and the pace of structural transformation. These differences pose challenges to the use of a common approach for predicting GDP growth in these countries, and underline the need for a flexible modelling strategy that can be adapted to different country-specific characteristics. At the same time, the limitations of the data for emerging markets lead to further challenges for the nowcasting of GDP growth, defined here as the prediction of GDP growth in the very recent past, the present and the very near future.1 For the majority of EMEs, data on GDP and many of the related hard indicators are released with longer lags than is the case for most advanced economies. For example, the first flash estimates of GDP for the United Kingdom and the United States are available 4 weeks after 1 Our definition of nowcasting follows (Banbura, Giannone, Modugno, & Reichlin, 2013). The term is a contraction of ‘‘now’’ and ‘‘forecasting’’, and was first introduced in the context of economic predictions by Giannone, Reichlin, and Small (2008).

http://dx.doi.org/10.1016/j.ijforecast.2017.05.002 0169-2070/© 2017 International Institute of Forecasters. Published by Elsevier B.V. All rights reserved.

916

T. Dahlhaus et al. / International Journal of Forecasting 33 (2017) 915–935

Fig. 1. Shares of global GDP in terms of PPP: 2000 vs. 2014. (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article.) Source: International Monetary Fund World Economic Outlook database and Haver Analytics.

Fig. 2. Real GDP growth for the BRIC+M. Note: Aggregate GDP growth is the PPP-GDP weighted aggregate of the individual countries’ real GDP growth. (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article.) Source: International Monetary Fund World Economic Outlook, Haver Analytics and authors’ calculations.

the end of the quarter, compared with 13 weeks for Russia and more than 8 weeks for Brazil, India and Mexico. In addition to the delay in data releases, there are also a number of other data challenges that further underscore the need to develop nowcasting tools for EMEs that are designed to address these issues. These challenges include (1) data published at different frequencies (weekly, monthly or quarterly); (2) unbalanced data patterns at the end of the sample due to non-synchronous data releases (‘‘ragged edge’’)2 ; (3) missing data at the beginning of the sample and relatively small sample sizes, since many macro indicators for EMEs have become available only recently; (4) varying data formats (such as year-over-year versus quarter-over-quarter growth rates); and (5) substantial data revisions. While these peculiarities of the data are not unique to EMEs, they are more pronounced for these countries than for advanced economies. 2 ‘‘Ragged edges’’ generally arise in real-time applications when varying numbers of observations are missing at the end of the sample, since different series are released at different points in time and are subject to different publication lags.

The aim of this paper is to nowcast real GDP growth for the BRIC+M, and to exploit our cross-country analysis to uncover the common features of the emerging market data flow. We use the modelling framework of Giannone et al. (2008), based on dynamic factor models (hereafter DFMs) à la Forni, Hallin, Lippi, and Reichlin (2000) and Forni and Lippi (2001), but also rely on several extensions due to Banbura and Modugno (2014). This setup is flexible enough to capture some of the country-specific characteristics, and is also highly suitable for overcoming the data challenges highlighted above. We account for the differences in economic structures across countries by building separate DFMs for Brazil, Russia, India, China and Mexico that utilize a variety of different monthly indicators for each economy. We allow the model specifications to vary across countries. Furthermore, these country-specific DFMs allow us to provide a BRIC+M aggregate nowcast which could be of use to policy-makers in monitoring global economic developments. Although the DFM framework has become the workhorse model for forecasters at central banks and

T. Dahlhaus et al. / International Journal of Forecasting 33 (2017) 915–935

917

other institutions, the majority of applications have focused on advanced economies.3 Relatively few studies have focused on EMEs, with key recent ones being that of Bragoli, Metelli, and Modugno (2014) for Brazil, Giannone, Miranda-Agrippino, and Modugno (2013) for China, Caruso (2015) for Mexico, Luciani, Pundit, Ramayandi, and Veronese (2015) for Indonesia, Liu, Matheson, and Romeu (2012) for 10 Latin American countries, and Porshakov, Deryugina, Ponomarenko, and Sinyakov (2015) for Russia. Our paper contributes to this relatively nascent, yet growing, body of literature by using DFMs to nowcast GDP growth in EMEs along three main dimensions. First, a key contribution of our study is the creation of bimonthly data vintages for the BRIC+M group, which allows us to evaluate the pseudo real-time performances of the DFMs for each country. The creation of these bimonthly pseudo real-time datasets is very valuable, since real-time data are either not available for these countries or very scanty. Although Liu et al. (2012) also compiled pseudo real-time data sets for Brazil and Mexico (along with other Latin American countries), to the best of our knowledge this paper is the first to do so at the bimonthly frequency, thus mimicking real-time data more closely.4 This feature of our framework may be particularly appealing for policymakers, as it allows EMEs’ economic activity to be assessed in a more timely manner. We construct these pseudo real-time data vintages by using data release dates from previous years to create a standardized bimonthly release calendar for each variable in the data set. Together, this provides data release patterns for 131 variables across the five EMEs. These data sets mimic precisely the data that were available to the nowcaster at the middle and end of a given month (hence ‘‘real time’’), but do not incorporate the possibility of revisions (hence ‘‘pseudo’’). Second, applying the same framework to five major EMEs enables us to provide a cross-country comparison of our results, including the relative importance of various data releases in driving movements in the nowcasts. Another novel element of our paper is its analysis of DFMs in a real-time setting in order to test whether these models are robust to data revisions. We do this by using real-time data vintages for GDP growth and a small set of macroeconomic variables that are available from the OECD for Brazil and Mexico.5 Our results show that the DFM framework provides reliable nowcasts of GDP growth for the BRIC+M. We find that nowcasts based on DFMs generally outperform those based on univariate benchmark models. Broadly speaking, the models display good directional accuracies and perform

reasonably well in terms of capturing the GDP dynamics of the countries under consideration. These results are generally robust to alternative model specifications. Moreover, using real-time OECD data for Brazil and Mexico, we also find some evidence that DFMs are generally robust to data revisions, which is in line with the findings of Giannone et al. (2008). Finally, we compute and evaluate the impacts on nowcast revisions of both ‘‘news’’, which is the difference between the expected and realized data, and parameter reestimation. We find that re-estimation has a significantly smaller impact on nowcast revisions than do new data releases. Nonetheless, parameter re-estimation generally improves the accuracy of the nowcasts across the BRIC+M, implying that it is important to consider re-estimation when working with relatively volatile emerging-markets data. Furthermore, the contribution of ‘‘news’’ to the nowcast revisions varies considerably across both countries and indicator groups, although the results reveal a few common patterns across the BRIC+M. First, data releases of financial variables and soft indicators play a major role in explaining nowcast revisions early in the nowcasting period. Second, news pertaining to domestic variables is by far the most important contributor to nowcast revisions throughout the nowcasting period, while exogenous (i.e., global and US) variables seem to improve the forecast accuracy early on. Finally, data releases that occur late in the nowcasting period (i.e., when predicting GDP growth for the previous quarter) are of little relevance in improving the prediction accuracy. Our paper also has some interesting implications for both researchers and policy-makers. We provide evidence that the DFM approach, which was first implemented by Giannone et al. (2008) for the United States and has been applied successfully to several advanced economies, performs well in terms of nowcasting the BRIC+M. This is a striking finding, given the heterogeneity between the two groups of countries in terms of both size and economic structure, and underscores the flexibility and applicability of the DFM approach. The remainder of the paper is organized as follows. Section 2 describes the econometric framework in detail. Section 3 describes the data and the nowcasting exercise, while Section 4 presents the results. Section 5 presents the aggregate BRIC+M nowcast, the real-time analysis for Brazil and Mexico, and additional robustness checks. Finally, Section 6 provides a concluding discussion.

3 The framework has been applied to several advanced economies, such as the euro area (Angelini, Camba-Mendez, Giannone, Reichlin, and Runstler, 2011; Banbura and Runstler, 2011; and Runstler, Barhoumi, Cristadoro, Den Reijer, Jakaitiene, Jelonek, Rua, Ruth, Benk, and Van Nieuwenhuyze, 2009), France (Barhoumi, Darne, & Ferrara, 2010), Ireland (D’Agostino, McQuinn, & O’Brien, 2012), New Zealand (Matheson, 2010), Norway (Astveit & Trovik, 2012; Luciani & Ricci, 2014), and Switzerland (Siliverstovs & Kholodilin, 2010). Banbura et al. (2013) provide a review of the literature on economic nowcasting based on multivariate dynamic models. 4 Bhattacharya, Pandey, and Veronese (2011) also compiled a pseudo

2.1. The dynamic factor model

real-time dataset for India, but only at the monthly frequency. 5 We thank an anonymous referee for this suggestion.

where ft is a r × 1 vector of (unobserved) common factors, Λ is an n × r matrix of factor loadings, ϵt is a vector of

2. Econometric framework

We start by characterizing the dynamics for the monthly data. Let xt = (x1,t , x2,t , . . . , xn,t )′ , t = 1, . . . , T , denote the n-dimensional vector of (stationary) monthly variables that has the following factor model representation: xt = µ + Λft + ϵt ,

(1)

918

T. Dahlhaus et al. / International Journal of Forecasting 33 (2017) 915–935

idiosyncratic components, and µ is a vector of unconditional means. Regarding the dynamics of the idiosyncratic component of monthly variables, ϵt , we allow it either to be serially uncorrelated, or to follow an AR(1) process: ei,t ∼ i.i.d. N(0, σi2 ),

ϵi,t = αi ϵi,t −1 + ei,t ,

(2)

with E [ei,t ej,s ] = 0 for i ̸ = j. For each of the five EMEs under consideration, we choose between these two alternative cases based on the out-of-sample evaluation exercise.6 The common factors are modeled as a stationary vector autoregressive (VAR) process of order p, with A1 , . . . , Ap being r × r matrices of autoregressive coefficients: ft = A1 ft −1 + · · · + Ap ft −p + ut ,

ut ∼ i.i.d. N(0, Q ).

(3)

Large DFMs can be estimated in a few different ways, such as by using principal components (e.g., Stock and Watson, 2002) or by quasi maximum likelihood (Doz, Giannone, and Reichlin, 2012). We adopt the second approach; specifically, we use the EM algorithm of Banbura and Modugno (2014), which can handle both mixed frequencies and missing data. Doz et al. (2012) have established the consistency and robustness properties of the maximum likelihood approach when both the size of the sample and the cross-section are large. For technical details on the EM iterations, we refer the reader to Banbura and Modugno (2014). We now discuss how quarterly variables can be included within the DFM framework. Following Mariano and Murasawa (2003), we express each quarterly variable in terms of a partially-observed monthly counterpart. For Q example, the level of the quarterly GDP (GDPt ) can be expressed as the sum of the contributions from its unobserved monthly counterparts (GDPM t ): M M GDPt = GDPM t + GDPt −1 + GDPt −2 , Q

Q

t = 3, 6, 9, . . . . (4)

Q

Let Yt = 100 × log(GDPt ) and YtM = 100 × log(GDPM t ). We assume that the (unobserved) monthly growth rate of GDP, yt = 1YtM , follows the same factor model representation as the real monthly variables: yt = µQ + ΛQ ft + ϵt . Q

(5)

When working with emerging-market data, we need to take into account two alternative data formats, namely quarter-over-quarter and year-over-year growth rates. In general, most national statistical agencies publish their GDP series in levels, and it is common practice for researchers and policy-making organizations to nowcast the quarter-over-quarter GDP growth. This is indeed the case for Brazil, Russia, India and Mexico, but the National Bureau of Statistics of China officially reports its real GDP series in year-over-year terms, thus requiring GDP growth nowcasts in this format.7 The online appendix describes our approach to incorporating quarterly series, expressed in different data formats, into the monthly factor model. 6 Section 3.2 describes the out-of-sample evaluation exercise in detail. 7 Note that the National Bureau of Statistics of China has recently made available a quarter-over-quarter GDP growth rate (seasonally adjusted), starting in 2010.

3. Data and nowcasting framework 3.1. Data We use dynamic factor models to nowcast real GDP growth for five major EMEs: Brazil, Russia, India, China and Mexico. Our models include a large number of monthly indicators that can be grouped into nine categories, although the coverages within the categories differ across countries.8 These categories are purchasing managers’ indexes (PMIs), soft indicators (‘ind’), industrial production (‘IP’), vehicle production, sales and use (‘vehicle’), balance of payments (‘BoP’), financial indicators (‘financial’), labor market indicators (‘labor’), prices, and exogenous variables (‘exog’).9 Tables 1–5 provide detailed descriptions of the country-specific data panels, including the publication lags and applied transformations for each series. The non-synchronous nature of data releases (i.e., the ragged edge) both within each country and across the BRIC+M group makes these countries interesting case studies for applying the DFM, since the approach is adept at dealing with such data publication patterns. Further, focusing on five major EMEs enables us to tease out some common features of data releases among these countries, as well as the differences relative to advanced economies such as the US. For example, industrial production, which is a crucial business cycle indicator, is available only with a considerable lag of over six weeks for Brazil, Russia, India and Mexico, while the corresponding lag for the US (and China) is two weeks. Similarly, vehicle sales data for EMEs are generally available with a lag of about two weeks, while for the US they are available a few days after the end of the month (i.e., a lag of less than a week). Similarly, labor market indicators for the BRIC+M are published with longer lags than for the US. The final selection of the variables for each country that are used in this paper is determined as follows. First, we consider the series of indicators that are available for each economy from Bloomberg and Haver Analytics, with a focus on the ones that are followed most closely by market participants. We then refine this set of available indicators by using economic judgment to select the variables that are most relevant for forming expectations of the current GDP for the particular country (Banbura, Giannone, & Reichlin, 2010a, b). For example, we include worker remittances for Mexico and the trend of German industrial production for Russia. Following this approach, we end up with a relatively small number of variables compared with typical factor model applications. However, Banbura et al. (2010a, b) have shown that medium-sized models (i.e., 20 to 40 variables) perform just as well as large models (with 100 or more variables) in terms of forecast accuracy. Furthermore, Luciani (2014) shows that it is useful to capture 8 We convert variables that are available at a daily frequency, such as oil prices and most financial variables, into a monthly frequency by taking the average over the month. 9 We do not include PMI series for Mexico, as they are available only over the last five years or so. Soft indicators, such as the index of consumer expectations and indexes of manufacturing, business and consumer confidence, and data on vehicle production are not available for India. We also do not include labor market indicators for India and China.

Brazil PMI: Manufacturing Output (SA, 50+ = Expansion) Brazil PMI: Manufacturing Stocks of Finished Goods (SA, 50+ = Expansion) Brazil PMI: Manufacturing Input Prices(SA, 50+ = Expansion) Brazil PMI: Manufacturing (SA, 50+ = Expansion) Brazil PMI: Manufacturing New Orders (SA, 50+ = Expansion) Brazil PMI: Manufacturing Output Prices (SA, 50+ = Expansion) Brazil PMI: Manufacturing Stocks of Purchases (SA, 50+ = Expansion) Brazil: Manufacturing Confidence Index (SA, Points) Brazil: ICEI: Business Confidence Index (50+ = Growth) Brazil: Total Leading Indicator (NSA, Amplitude Adjusted) Brazil: Economic Activity Indicator (SA) Brazil: Consumer Expectations Index (SA, Points) Brazil: Consumer Confidence Index (SA, Points) Brazil: Industrial Production (SA, 2002 = 100) Brazil: Industrial Production: Metals (SA, 2002 = 100) Brazil: Total Production of Vehicles (NSA, Units) Brazil: Retail Trade: Volume of Sales (NSA, 2011 = 100) Brazil: Merchandise Exports (NSA, Mil.US$) Brazil: Merchandise Imports (NSA, Mil.US$) Brazil: Money Supply, M1 (EOP, NSA, Mil.Reais) Brazil: Money Supply, M2 (EOP, NSA, Mil.Reais) JPMorgan Broad REER Index: Brazil (2000 = 100) Brazil: US$ Exchange Rate: Commercial rate (Reais/US$) Brazil: Stock Price Index: Bovespa Brazil: Unemployment Rate (NSA, %) [HEADLINE] Brazil: Emp: Mining & Mfg &(NSA, Thous) Brazil: Nominal Salaries: Manufacturing (NSA, 2006 = 100) Brazil: Natl Consumer Price Index [INPC] (12/93 = 100) Brazil: National Core CPI [Extended, IPCA] (NSA, 12/95 = 100) Brazil: CPI: Food & Beverages [INPC] (NSA, Jul 2006 = 100) Industrial Production Index (SA, 2007 = 100) Petroleum, Brent Spot Price (US$/Barrel) World: Commodity Price Index: All Commodities (2005 = 100) Federal Open Market Committee: Fed Funds Target Rate: Upper Limit (%) ISM Composite Index (SA, >50 = Increasing) Stock Price Index: Standard & Poor’s 500 Composite (1941–43 = 10)

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36

PMI PMI PMI PMI PMI PMI PMI Ind Ind Ind Ind Ind Ind IP IP Vehicle Ind BoP BoP Financial Financial Financial Financial Financial Labour Labour Labour Prices Prices Prices Exog Exog Exog Exog Exog Exog

Group S223MG@PMI S223MSF@PMI S223MPI@PMI S223M@PMI S223MO@PMI S223MPO@PMI S223MSP@PMI S223VM@EMERGELA N223VBC@EMERGELA C223LIAT@OECDMEI S223GVI@EMERGELA S223VCEF@EMERGELA S223VCCF@EMERGELA S223D@EMERGELA S223DMA@EMERGELA N223OMVU@EMERGELA C223SRV@EMERGELA C223TXD@EMERGELA C223TMD@EMERGELA C223M1E@EMERGELA C223M2E@EMERGELA N223XJRB@EMERGELA (C223XLDO@EMERGELA)/2 C223KN@EMERGELA C223EURM@EMERGELA N223ETXC@EMERGELA C223EWC@EMERGELA C223PCN@EMERGELA C223PCX@EMERGELA N223CPFB@EMERGELA IP@USECON C112CLPE@IFS C001CXAP@IFS FFEDTAH@USECON ISMC@USECON SP500@USECON

Haver code 0 0 0 0 0 0 0 0 0 0 1 0 0 1 1 1 1 1 1 1 1 1 1 1 0 1 1 1 1 1 1 1 1 0 0 1

Trans (log) 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 0 1 1 1 1 1 1 1 1 0 1 1

Trans (diff) 1 1 1 1 1 1 1 1 1 2 3 1 1 2 2 2 3 1 1 2 2 1 1 1 2 2 2 1 1 1 2 2 1 1 1 1

M10 1 1 1 1 1 1 1 0 0 2 2 0 0 2 2 1 2 1 1 1 1 0 0 1 1 1 2 1 1 1 1 1 1 0 1 0

M11 1 1 1 1 1 1 1 1 1 2 3 1 1 2 2 2 3 1 1 2 2 1 1 1 2 2 2 1 1 1 2 2 1 1 1 1

M20 1 1 1 1 1 1 1 0 0 2 2 0 0 2 2 1 2 1 1 1 1 0 0 1 1 1 2 1 1 1 1 1 1 0 1 0

M21 1 1 1 1 1 1 1 1 1 2 3 1 1 2 2 2 3 1 1 2 2 1 1 1 2 2 2 1 1 1 2 2 1 1 1 1

M30

1 1 1 1 1 1 1 0 0 2 2 0 0 2 2 1 2 1 1 1 1 0 0 1 1 1 2 1 1 1 1 1 1 0 1 0

M31

Notes: 1. The columns ‘trans log’ and ‘trans diff’ show the transformations that have been made to the raw data series in order to achieve stationarity. An entry of ‘1’ in the ‘trans log’ column means that the series is expressed in logs, while an entry of ‘0’ means otherwise. Similarly, the column ‘trans diff’ takes a value of 1 if the series are expressed in first differences, and 0 otherwise. Series that have ‘1’ entries in both columns have been log-differenced. 2. ‘M10’ and ‘M11’ refer to the beginning and end of the first month of a quarter, respectively; ‘M20’ and ‘M21’ refer to the beginning and end of the second month of a quarter; and ‘M30’ and ‘M31’ refer to the beginning and end of the third month of a quarter. 1 corresponds to a one-month data release lag, 2 to a two-month lag, and so on.

Name

No

Table 1 Data and publication lags for Brazil.

T. Dahlhaus et al. / International Journal of Forecasting 33 (2017) 915–935 919

920

T. Dahlhaus et al. / International Journal of Forecasting 33 (2017) 915–935

Table 2 Data and publication lags for China. No 1 2 3

4 5 6 7 8 9

10

11 12

13

Name

Group

Haver code

Trans (log)

Trans (diff)

M10

M11

M20

M21

M30

M31

China: Purchasing Managers’ Index (SA, 50+ = Expansion) China PMI: Manufacturing (SA, 50+ = Expansion) China: Macroeconomic Climate Index: Leading Index (NSA, 1996 = 100) China: Real Gross Value Added (NSA, yr/yr % change) China: Production of Energy: YoY: Electricity (%) CN: Freight Ton Kilometer Carried (Ton km Mn) China Exports fob (Mil.US$) China Imports cif (Mil.US$) China: JPMorgan Broad Real Effective Exchange Rate Index (2000 = 100) China: Index: Shanghai Stock Exchange: Composite (Dec-19-91 = 100) CN: Consumer Price Index (PY = 100) OECD Total: Industrial Production ex Construction (SA, 2010 = 100) European Free Market Price: Brent Crude Oil ($/Barrel)

PMI

CBAWLX@CHINA

0

1

1

1

1

1

1

1

PMI

S924M@MKTPMI

0

1

1

0

1

0

1

0

Ind

N924VLD@EMERGEPR

0

1

2

1

2

1

2

1

IP

n924gdxy@emergepr

0

0

1

1

1

1

1

1

IP

CRBADJS@CHINA

0

0

1

1

1

1

1

1

Vehicle

CTCB@CHINA

1

1

2

1

2

1

2

1

BoP BoP Financial

CJAA@CHINA CJAC@CHINA N924XJRB@EMERGEPR

1 1 1

1 1 1

2 2 1

1 1 0

2 2 1

1 1 0

2 2 1

1 1 0

Financial

aggany(CDIAA@CHINA)

1

1

1

0

1

0

1

0

Prices

CIEA@CHINA

0

0

1

1

1

1

1

1

Exog

C003IZ@OECDMEI

1

1

4

4

4

4

4

4

Exog

aggany(PZBRT@DAILY)

1

1

1

0

1

0

1

0

Notes: See notes to Table 1.

the dynamics of sectoral/disaggregated data when the goal is to forecast a disaggregated variable, but that aggregate data are sufficient for forecasting an aggregate variable, which is the objective of this paper. Finally, we fine-tune the selection of variables by comparing the out-of-sample performances of the DFMs based on different combinations of variables.10 Table 6 shows the final number of series in each category, as well as the total number of series for each country. The number of indicators used for the BRIC+M ranges from 13 to 41. We use 13 series for China and 14 for Russia, since these smaller variable sets are better than larger data sets in terms of the out-of-sample performances of the respective DFMs. This finding for the Chinese data is in line with that of Giannone et al. (2013), who also suggest that a smaller set of indicators performs well for nowcasting Chinese GDP growth. All indicators are differenced or log-differenced where appropriate, in order to achieve stationarity. In the cases of Brazil, India, Mexico and Russia, we transform all level series into sequential month-over-month growth rates. For China, we transform the level series to year-over-year growth rates, since many series, including GDP growth, are available only in the form of year-over-year growth rates.11 Finally, it is important to note that the DFM can be applied in both quarter-over-quarter and year-over-year growth

10 Section 3.2 describes the out-of-sample exercise in detail. 11 All growth rates are approximated using log differences.

formats in all cases, as long as the underlying indicator and GDP growth data are available. 3.2. Pseudo out-of-sample design Table 7 summarizes the design of the nowcasting exercise. The model is estimated over the period from 1996Q1 to 2014Q1 for Brazil and Mexico, while for India the estimation sample begins in 1996Q2.12 In the case of China, although the official year-over-year GDP growth series is available from 1992 onward, the sample starts in 2000Q1 in order to avoid data distortions from the Asian crisis. For Russia, the estimation period starts in 2001Q1 in order to avoid the 1998 crisis. The out-of-sample evaluation period is 2008Q1 to 2014Q1. Over this period, we evaluate the nowcasts at a bimonthly frequency – at the middle and end of each month – for a period of three quarters, i.e., up to 18 bimonthly periods surrounding the quarter of interest. For example, nowcasts for 2014Q1 are computed at a bimonthly frequency from mid-October 2013 until 30 June 2014 or until the 2014Q1 GDP is released, whichever occurs earlier. At each bimonthly point in time, more history is added to the unbalanced data panel, the model is re-estimated and the nowcasts are obtained. We therefore produce a sequence of nowcasts for each quarter of interest, i.e., projections in the preceding, current and following quarters (with respect to the quarter 12 The reason why the sample for India begins a quarter later is that the official quarterly GDP series for India is available only after that quarter.

IMEF Business Climate Index: Manufacturing (SA, 50+ = Econ Expand) IMEF Business Climate Index: Mfg: Inventory (NSA, 50+ = Econ Expand) IMEF Business Climate Index: Mfg: New Orders (SA, 50+ = Econ Expand) Indicator of Global Economic Activity (SA, 2008 = 100) Consumer Expectations (NSA, 01/03 = 100) Consumer Confidence (NSA, 01/03 = 100) Industrial Production (SA, 2008 = 100) Industrial Production: Primary Metal Manufacturing (SA, 2008 = 100) Industrial Production: Utilities (SA, 2008 = 100) IP: Electric Power Generation, Transmiss & Distribution (SA, 2008 = 100) Total Vehicle Production (NSA, Units) Total Vehicle to Sales to Distributors (NSA, Units) Retail Sales (SA, 2003 = 100) Exports, fob (SA, USD) Imports, fob (SA, USD) Non-petroleum Exports (SA, Mil.US$) Worker’s Remittances: Total (NSA, Mil.US$) Money Supply, M1 (NSA, Loc.Cur) Money Supply, M2 (NSA, Loc.Cur) Total Performing Commercial Bank Loans (EOP, NSA, Loc.Cur) JPMorgan Broad REER Index: Mexico (2000 = 100) Exchange Rate (NewPeso/US$) Stock Price Index: IPC (Avg, 11/78 = 0.78) 91-Day Treasury Certificates (%) 364-Day Treasury Certificates (%) 10 Year Government Bond Yield Unemployment Rate (SA, %) Employed: Manufacturing (NSA, Persons) Mfg Remunerations (NSA, Thous.Pesos) Coincident Index: IMSS Insured Workers (NSA, Points) Consumer Price Index (NSA, Dec 16–31, 2010 = 100) Consumer Price Index: Core (NSA, Dec 16–31, 2010 = 100) CPI: Food, Beverages & Tobacco (NSA, Dec 16–31, 2010 = 100) PPI: Total Production Incl oil & Services (NSA, Jun 2012 = 100) PPI: Final Merchandise Domestic Demand (NSA, Jun 2012 = 100) Industrial Production Index (SA, 2007 = 100) Petroleum, Brent Spot Price (US$/Barrel) World: Commodity Price Index: All Commodities (2005 = 100) Federal Open Market Committee: Fed Funds Target Rate: Upper Limit (%) ISM Composite Index (SA, >50 = Increasing) Stock Price Index: Standard & Poor’s 500 Composite (1941–43 = 10)

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41

Notes: See notes to Table 1.

Name

No

Table 3 Data and publication lags for Mexico.

Ind Ind Ind Ind Ind Ind IP IP IP IP Vehicle Vehicle Ind BoP BoP BoP BoP Financial Financial Financial Financial Financial Financial Financial Financial Financial Labour Labour Labour Labour Prices Prices Prices Prices Prices Exog Exog Exog Exog Exog Exog

Group S273VM@EMERGELA N273VMI@EMERGELA S273VMO@EMERGELA S273AE@EMERGELA C273VCE@EMERGELA C273VCC@EMERGELA S273DW@EMERGELA S273DMAB@EMERGELA S273DV@EMERGELA S273DVE@EMERGELA N273OMV@EMERGELA N273TSV@EMERGELA S273TRS@EMERGELA S273TXD@EMERGELA S273TMD@EMERGELA S273TXND@EMERGELA N273BW@EMERGELA C273M1@EMERGELA C273M2@EMERGELA N273UCH@EMERGELA N273XJRB@EMERGELA C273XLDV@EMERGELA C273KNV@EMERGELA C273RC91@EMERGELA C273RC1Y@EMERGELA N273RG10@EMERGELA S273EUR@EMERGELA N273ET@EMERGELA N273EE@EMERGELA N273VLC4@EMERGELA N273PJ@EMERGELA N273PJXQ@EMERGELA N273PJF@EMERGELA N273PP@EMERGELA N273PDF@EMERGELA IP@USECON C112CLPE@IFS C001CXAP@IFS FFEDTAH@USECON ISMC@USECON SP500@USECON

Haver code 0 0 0 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 0 0 0 0 1 1 0 1 1 1 1 1 1 1 1 0 0 1

Trans (log) 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 0 0 0 0 1 1 0 1 1 1 1 1 1 1 1 0 1 1

Trans (diff) 1 1 1 3 1 1 2 2 2 2 1 1 2 2 2 2 2 2 2 2 1 1 1 1 1 1 2 3 3 2 1 1 1 1 1 2 2 1 1 1 1

M10 1 1 1 2 1 1 2 2 2 2 1 1 1 1 1 1 2 1 1 1 0 1 1 0 0 0 1 2 2 2 1 1 1 1 1 1 1 1 0 1 0

M11 1 1 1 3 1 1 2 2 2 2 1 1 2 2 2 2 2 2 2 2 1 1 1 1 1 1 2 3 3 2 1 1 1 1 1 2 2 1 1 1 1

M20 1 1 1 2 1 1 2 2 2 2 1 1 1 1 1 1 2 1 1 1 0 1 1 0 0 0 1 2 2 2 1 1 1 1 1 1 1 1 0 1 0

M21 1 1 1 3 1 1 2 2 2 2 1 1 2 2 2 2 2 2 2 2 1 1 1 1 1 1 2 3 3 2 1 1 1 1 1 2 2 1 1 1 1

M30

1 1 1 2 1 1 2 2 2 2 1 1 1 1 1 1 2 1 1 1 0 1 1 0 0 0 1 2 2 2 1 1 1 1 1 1 1 1 0 1 0

M31

T. Dahlhaus et al. / International Journal of Forecasting 33 (2017) 915–935 921

922

T. Dahlhaus et al. / International Journal of Forecasting 33 (2017) 915–935

Table 4 Data and publication lags for India. No 1 2 3 4 5 6 7 8 9

10 11 12

13

14 15

16 17 18 19 20

21 22

Name

Group

Haver code

Trans (log)

Trans (diff)

M10

M11

M20

M21

M30

M31

India PMI: Manufacturing Output (SA, 50+ = Expansion) India: Industrial Production (SA, FY2004 = 100) India: IP: Basic metals (NSA, FY2004 = 100) India: IP: Electricity (NSA, FY2004 = 100) India: Exports, fob (NSA, USD) India: Imports, fob (NSA, USD) India: Money Supply: M1 (NSA, Bil.Rupees) India: Money Supply: M2 (NSA, Bil.Rupees) India: JPMorgan Broad Real Effective Exchange Rate Index (2000 = 100) India: Rupee/US$ Exchange Rate (AVG) India: Stock Price Index: NSE 500 (AVG, 1994 = 1000) India: 91-Day Treasury Bill Implicit Cut-Off Yield (% per annum) India: 364-Day Treasury Bill Implicit Cut-Off Yield (% per annum) India: 10-Year Government Bond Yield (% per annum) India: Consumer Price Index: Food, Beverages, Tobacco (NSA, 2010 = 100) India: Wholesale Price Index: All Items (NSA, FY04 = 100) Industrial Production Index (SA, 2007 = 100) Petroleum, Brent Spot Price (US$/Barrel) World: Commodity Price Index: All Commodities (2005 = 100) Federal Open Market Committee: Fed Funds Target Rate: Upper Limit (%) ISM Composite Index (SA, >50 = Increasing) Stock Price Index: Standard & Poor’s 500 Composite (1941–43 = 10)

PMI

S534MG@PMI

0

1

1

1

1

1

1

1

IP

N534D@EMERGEPR

1

1

2

2

2

2

2

2

IP

N534DMAB@EMERGEPR 1

1

2

2

2

2

2

2

IP

N534DVE@EMERGEPR

1

1

2

2

2

2

2

2

BoP BoP Financial

N534IXD@EMERGEPR N534IMD@EMERGEPR N534FM1@EMERGEPR

1 1 1

1 1 1

2 2 1

1 1 1

2 2 1

1 1 1

2 2 1

1 1 1

Financial

N534FM2@EMERGEPR

1

1

1

1

1

1

1

1

Financial

N534XJRB@EMERGEPR

1

1

1

0

1

0

1

0

Financial

N534XUSV@EMERGEPR

1

1

1

0

1

0

1

0

Financial

N534SK5@EMERGEPR

1

1

1

0

1

0

1

0

Financial

N534RJ3M@EMERGEPR

0

0

1

1

1

1

1

1

Financial

N534RJ1Y@EMERGEPR

0

0

1

1

1

1

1

1

Financial

N534RG10@EMERGEPR

0

0

2

1

2

1

2

1

Prices

N534PCF@EMERGEPR

1

1

1

1

1

1

1

1

Prices

N534PW@EMERGEPR

1

1

1

1

1

1

1

1

Exog

IP@USECON

1

1

2

1

2

1

2

1

Exog

C112CLPE@IFS

1

1

2

1

2

1

2

1

Exog

C001CXAP@IFS

1

1

1

1

1

1

1

1

Exog

FFEDTAH@USECON

0

0

1

0

1

0

1

0

Exog

ISMC@USECON

0

1

1

1

1

1

1

1

Exog

SP500@USECON

1

1

1

0

1

0

1

0

Notes: See notes to Table 1.

of interest), which we denote Q(−1), Q(0) and Q(+1), respectively.13 Nowcasts made in the first, second, and third month of a quarter are referred to as M1, M2, and M3, while D15 and D30 stand for the middle and end of a month, respectively. For example, Q(+1) M1 D30 would denote the nowcast performed at the end (D30) of the first month (M1) in the quarter following the quarter of interest (Q(+1)). We generate the pseudo real-time data that are required for the out-of-sample evaluation by constructing a grid that tracks the publication lag of each indicator for each of the five EMEs. Specifically, we use information

13 Note that some applications label projections made in the preceding, current, and following quarters as ‘‘forecasts’’, ‘‘nowcasts’’ and ‘‘backcasts’’, respectively.

from J.P. Morgan’s data release calendars and/or countryspecific sources, such as the central bank, to construct a data grid for each country. Let us use the grid for Mexico (Table 3) as an example. Let M10 and M11 refer to the beginning and end of the first month of a quarter. Looking at rows 14 and 15, we see that the exports and imports series are available with lags of one to two months and are released in the middle of the month.14 Thus, only the November figures would be available in early January, but the December exports and imports data would be released by the end of January. Thus, the pseudo real-time data sets allow us to assess the contribution of data releases to 14 A ‘2’ in the table corresponds to a two-month lag, whereas a ‘1’ corresponds to a one-month lag.

T. Dahlhaus et al. / International Journal of Forecasting 33 (2017) 915–935

923

Table 5 Data and publication lags for Russia. No 1 2 3 4

5

6 7 8 9 10 11 12 13 14

Name

Group

Haver code

Trans (log)

Trans (diff)

M10

M11

M20

M21

M30

M31

Russia PMI: Manufacturing (SA, 50+ = Expansion) Russia PMI: Manufacturing New Orders (SA, 50+ = Expansion) Russia: Business Confidence: Manufacturing (NSA, %) Russia: Output: Industrial Production Index: Total (SA, 2011 = 100) Russia: Passenger Car Sales Imported Plus Domestic (NSA, Units) Russia: Money Supply: M2 (NSA, Loc.Cur) Russia: Stock Price Index: RTS (avg, Sep 1995 = 100) Russia: Gov’t Securities: Up to 90 Days AVG (%) Russia: Zero Coupon Yield Curve: 10-Year (AVG, %) Russia: Producer Price Index: Total (NSA, 2000 = 100) Industrial Production Index (SA, 2007 = 100) World: Commodity Price Index: All Commodities (2005 = 100) ISM Composite Index (SA, >50 = Increasing) German IP ex construction (Trend)

PMI

S922M@MKTPMI

0

1

1

1

1

1

1

1

PMI

S922MO@MKTPMI

0

1

1

1

1

1

1

1

Ind

N922VBM@EMERGECW 0

1

1

0

1

0

1

0

IP

S922D@EMERGECW

1

1

2

1

2

1

2

1

Vehicle

N922CVLF@EMERGECW 1

1

1

1

1

1

1

1

Financial

N922FM2@EMERGECW

1

1

2

1

2

1

2

1

Financial

N922FKRV@EMERGECW 1

1

1

0

1

0

1

0

Financial

N922RGS@EMERGECW

0

0

1

1

1

1

1

1

Financial

N922G10@EMERGECW

0

0

1

1

1

1

1

1

Prices

N922PP@EMERGECW

1

1

2

1

2

1

2

1

Exog

IP@USECON

1

1

2

1

2

1

2

1

Exog

C001CXAP@IFS

1

1

2

1

2

1

2

1

Exog

ISMC@USECON

0

1

1

1

1

1

1

1

Exog

T134D@G10

1

1

2

2

2

2

2

2

Notes: See notes to Table 1.

Table 6 Number of indicators by type for BRIC+M.

Brazil China Mexico India Russia

PMI

Soft indicators

IP

Vehicle

BoP

Financial

Labour

Prices

Exogenous

Total

7 2 0 1 2

7 1 7 0 1

2 2 4 3 1

1 1 2 0 1

2 2 4 2 0

5 2 9 8 4

3 0 4 0 0

3 1 5 2 1

6 2 6 6 4

36 13 41 22 14

Notes: PMI stands for purchasing managers’ index, IP stands for industrial production and BoP stands for balance of payments.

nowcasts at the bimonthly frequency. We then overlay this completed grid onto our data set in order to reproduce the characteristic ‘‘ragged edge’’ or unbalanced panel at each point in time. We construct the unbalanced data panels for the other four countries in a similar fashion. These data sets mimic precisely the data sets that were available to the econometrician at the middle and end of each month (hence ‘‘real time’’), but do not account for the possibility of revisions (hence ‘‘pseudo’’).

3.3. Model selection We select the optimal model for each economy by comparing the out-of-sample performances for all possible combinations of model parameters, including the number of factors (r = 1, 2, . . . , 5), the number of lags in the VAR (p = 1, 2, . . . , 5), and the idiosyncratic component modeled as an autoregressive process of order 1 (i.e., AR(1)). We use the root mean squared forecast error (RMSE) as the

measure of forecast accuracy. The size of the state space increases significantly with each additional parameter, such as an extra factor, thus limiting the set of admissible factor model specifications. Table 7 shows the chosen model parameters for each economy. The fact that the BRIC+M countries differ in their underlying economic features can be seen from the number of common factors that are chosen for obtaining the best nowcasting performances. We find that simple model specifications, with one or two factors and one lag, often yield the best out-of-sample performances. However, adding more factors helps to reduce the RMSEs in the cases of Mexico and Brazil. One obvious reason for this is the larger set of indicators that is included relative to other countries in our sample, which reflects our variable selection criteria (as was discussed in Section 3.1). Modeling the idiosyncratic component as an AR(1) process also improves the out-of-sample performance of the model in the cases of Russia, China and Brazil, but not for India and Mexico. This suggests that the differences in economic structures

924

T. Dahlhaus et al. / International Journal of Forecasting 33 (2017) 915–935

Table 7 Model set-up.

Data format Estimation period Specification Idiosyncratic AR?

Brazil

China

Mexico

India

Russia

Q/Q 96Q1–14Q1 r = 4, p = 1 Yes

Y/Y 00Q1–14Q1 r = 1, p = 1 Yes

Q/Q 96Q1–14Q1 r = 3, p = 4 No

Q/Q 96Q2–14Q1 r = 2, p = 1 No

Q/Q 2001Q1–14Q1 r = 2, p = 1 Yes

Fig. 3. DFM vs. AR(2) benchmark: Pseudo out-of-sample RMSEs. (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article.)

across the BRIC+M may also imply distinct behavioral idiosyncrasies. For example, non-common dynamics could be stronger in India than in Russia. Section 5.3 estimates a DFM with one factor and two lags for each country and shows that the country-specific choice of parameter values based on our out-of-sample evaluation exercise does better than a uniform parsimonious specification for all countries. 4. Results This section reports the out-of-sample performances of our dynamic factor models and the impact of data releases or ‘‘news’’ on the model nowcast for each of the five EMEs. 4.1. Forecast accuracy: pseudo out-of-sample evaluation We begin by assessing the forecast accuracy of the DFMs, as measured by the RMSE. We present the models’ performances relative to two standard benchmarks: AR(2) and MA(4). Fig. 3 shows the ratio of the RMSEs from the dynamic factor model to those from an AR(2) model on the y-axis. The x-axis shows each bimonthly nowcast evaluation period, from the middle of the first month of the quarter immediately preceding the quarter of interest to the end of the last month of the quarter immediately following the quarter of interest. This adds to up to 17 different nowcast origins, depending on the timeliness of the GDP releases across countries. For example, in the case of China, the RMSE ratios can be obtained up to Q(+1) M1 D15, since GDP growth is published three weeks after the close of the quarter. A number below one indicates that

the dynamic factor model forecasts are more accurate than those produced by the AR(2) benchmark. Two key results emerge from Fig. 3. First, the RMSE of the DFM decreases as the actual GDP release date approaches for all of the countries under consideration. The finding that the nowcast accuracy of the DFM tends to improve as more data become available is standard in the literature (see for example Giannone et al., 2008, and Liu et al., 2012). Second, the RMSE ratios are below one across all nowcast origins for Brazil, China, India and Mexico (except for Brazil at Q(−1) M1 D15). Thus, the DFMs beat the naive AR(2) benchmark at very early nowcast origins, which suggests that they are good at nowcasting GDP growth not only in the present and past, but also in the future. In the case of Russia, the DFM nowcast improves upon the benchmark only in the quarter of interest (i.e., Q(0)). This is likely to be due to the long publication lags for Russian data. Fig. 4 shows the forecast accuracy of the DFMs relative to an MA(4) benchmark. Qualitatively, the results are very similar to those discussed above. However, the DFM for China beats the MA(4) benchmark one forecast origin later than the AR(2) benchmark. 4.2. Pseudo out-of-sample nowcasts Figs. 5–9 show the out-of-sample nowcasts for the five EMEs over the evaluation period. The blue bars represent the realized quarterly GDP growth (non-annualized). Let Q(0) refer to the quarter of interest, M1 to the first month of that quarter, and D30 to the end of the month. The red line shows the results of a nowcast performed early in the quarter of interest, i.e., at the end of the first month of the

T. Dahlhaus et al. / International Journal of Forecasting 33 (2017) 915–935

925

Fig. 4. DFM vs. MA(4) benchmark: Pseudo out-of-sample RMSEs. (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article.)

Fig. 5. Out-of-sample nowcasts for Brazil. Notes: Q(0) refers to the quarter of interest, M1 refers to the first month of that quarter, and D30 refers to the end of the month. Q(+1) refers to the quarter following the quarter of interest. (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article.)

quarter being nowcasted. For example, if we are nowcasting 2013Q4, the red line would show the nowcast results as of 30 October 2013. At this point in the quarter of interest, there is usually very little data available beyond the highfrequency financial indicators. The green line shows the results of a nowcast performed later in the monitoring period. Specifically, it represents the nowcasts performed at the end (D30) of the first month (M1) in the quarter following the quarter of interest, i.e., Q(+1). Again, if we were nowcasting 2013Q4, the green line would represent the nowcast of 2013Q4 GDP growth at the end of January 2014. This nowcast would be based on a lot of data for the quarter of interest, including multiple releases of hard indicators, such as industrial production, and complete quarterly data for higher-frequency financial variables. Finally, we also provide a comparison of our nowcasts with those obtained from an AR(2) model for the later nowcast origin, which is represented by the dotted green line.

Considering the nowcast of Brazilian GDP growth as of Q(0) M1 D30 (Fig. 5), we can see that the model tracks the observed quarterly GDP growth relatively well, despite the lack of data at that point in time. The directional accuracy improves for the nowcast made later in the monitoring period (i.e., Q(+1) M1 D30), as more data are available by then. In particular, the earlier nowcast seems to predict the trough during the global financial crisis of 2008–09 with a lag of one quarter, while the nowcast made later tracks it well. In contrast, the AR(2) nowcast appears to be relatively flat, with little directional accuracy. A very similar picture arises in the case of China (Fig. 6). Note that in this case we consider nowcasts as of Q(−1) M3 D30 (early nowcast) and Q(0) M3 D30 (late nowcast), since the publication lag of the Chinese GDP is shorter than those for the other countries. The nowcast as of Q(0) M3 D30 generally performs better than that made earlier in the monitoring period. In particular, the green line captures the downturn and subsequent pickup in Chinese GDP growth

926

T. Dahlhaus et al. / International Journal of Forecasting 33 (2017) 915–935

Fig. 6. Out-of-sample nowcasts for China. Notes: Q(0) refers to the quarter of interest, M3 refers to the third month of that quarter, and D30 refers to the end of the month. Q(−1) refers to the quarter preceding the quarter of interest. (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article.)

Fig. 7. Out-of-sample nowcasts for Mexico. See the notes to Fig. 5. (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article.)

during the global financial crisis more accurately and in a more timely manner than the red line. The nowcast as of Q(−1) M3 D30 generally appears to track the direction of GDP growth with a lag, as does the AR(2) nowcast. The DFM performs similarly well (and generally better than the AR(2) model) in nowcasting quarterly GDP growth for Mexico, India and Russia (Figs. 7–9). Once again, a comparison of the red and green lines for these countries shows that incorporating more information typically improves the performance of the nowcast. In general, the DFMs display good directional accuracies and perform well for capturing GDP dynamics in these countries during the global financial crisis, with the exception of India.

4.3. Impact of news and re-estimation on nowcasts An important feature of the methodology used here is its ability to decompose the sources of changes in the nowcasts following new data releases. Since the DFM produces forecasts for all variables in the data set, the methodology allows us to decompose the changes in the model outcomes precisely into the contributions of individual indicators and/or groups of indicators, or what we refer to as the impact of ‘‘news’’, and the impact of parameter updates, i.e., re-estimation.15 We assess the relative importance of different ‘‘news’’ groups and re-estimation on the GDP 15 See the online appendix for details.

T. Dahlhaus et al. / International Journal of Forecasting 33 (2017) 915–935

927

Fig. 8. Out-of-sample nowcasts for India. See the notes to Fig. 5. (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article.)

Fig. 9. Out-of-sample nowcasts for Russia. See the notes to Fig. 5. (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article.)

Fig. 10. Brazil: average absolute impacts of news and re-estimation. (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article.)

928

T. Dahlhaus et al. / International Journal of Forecasting 33 (2017) 915–935

Fig. 11. China: average absolute impacts of news and re-estimation. (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article.)

Fig. 12. Mexico: average absolute impacts of news and re-estimation. (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article.)

Fig. 13. India: average absolute impacts of news and re-estimation. (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article.)

T. Dahlhaus et al. / International Journal of Forecasting 33 (2017) 915–935

929

Fig. 14. Russia: average absolute impacts of news and re-estimation. (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article.)

nowcasts by focusing on the average absolute impacts of data releases and re-estimation, averaged over the period 2008Q2 to 2014Q1 (Figs. 10–14). In addition, Figures A1– A5 of the online appendix present the evolution of the RMSE across data releases for different nowcast origins. For ease of exposition, we group the ‘‘news’’ from individual releases into the groups discussed in Section 3.1. In the case of Brazil, Fig. 10 shows the breakdown of the contributions across indicator groups, as well as the impact of re-estimation, in average absolute terms. In general, we find that the impact of re-estimation on the GDP nowcasts is relatively small compared with that of ‘‘news’’. Focusing first on news, the results show that the GDP nowcast responds to several indicator groups, with financial variables such as the stock market index, PMIs, and soft indicators shifting the GDP nowcast with every release. Furthermore, most of the changes in the GDP nowcast early in the monitoring period can be attributed to the release of soft indicators and financial variables, but the importance of these indicator groups seems to diminish as we move through the nowcast period, whereas that of the IP and BoP groups increases. The release of exogenous indicators is relevant early in the nowcasting period. We also find that the contribution of GDP releases per se to the nowcast revision is non-trivial. Furthermore, Figure A1 of the online appendix shows the evolution of the RMSE across data releases for different nowcast origins. It is interesting to note that although reestimation has only a small relative impact on the GDP nowcasts, it does generally play a role in reducing the RMSE, except in the cases of Q(−1) M1, Q(−1) M2 and Q(0) M2. Also, the RMSE remains essentially flat after Q(0) M3, implying that new data releases have little impact on the prediction accuracy late in the nowcasting period. A slightly different picture emerges with regard to China when we look at the average absolute contributions of the various groups of indicators. Fig. 11 shows that the actual GDP growth releases are the biggest contributors to nowcast revisions. Both the year-over-year format and the relatively short time lag in the release of Chinese

GDP data probably help explain why it contains important information for the nowcast of interest. Apart from the importance of the news from GDP releases, we also find the release of financial variables and industrial production to have significant effects on the GDP nowcast throughout the monitoring period. On average, the PMI releases seem to be of little importance for the nowcast of Chinese GDP growth, which is somewhat surprising, since this group of indicators is generally regarded as providing early signals of broad macroeconomic developments. Furthermore, reestimation has only a small average absolute impact on the GDP nowcasts relative to that of the various indicator groups taken together, although it still plays a role in explaining nowcast revisions, especially in the month of the GDP release. Finally, Figure A2 in the online appendix shows the evolution of the RMSE as new data are released. Clearly, the prediction of current-quarter GDP does not improve further with new data releases once the previous quarter’s GDP has been released in Q(0) M1. Fig. 12 shows the corresponding results for Mexico. ‘‘News’’ related to BoP, financial indicators, exogenous variables and prices are the biggest contributors to nowcast revisions. However, the importance of financial indicators diminishes over the nowcast period. Once again, the overall contribution of data releases to changes in the GDP nowcasts dominates that of re-estimation, in average absolute terms. Nevertheless, re-estimation still matters, since it leads to a decrease in RMSEs over the nowcasting period (see Figure A3 in the online appendix). The previous quarter’s GDP release in Q(0) M2 is also found to play a role. The prediction accuracy does not improve after Q(0) M3, as the RMSE does not change much with further data releases or re-estimation. In the case of India, financial indicators and exogenous variables are the only data categories for which data releases shift the nowcast revisions early in the monitoring period (Fig. 13). The release of the previous quarter GDP growth is also an important contributor to the nowcast revision. While the average absolute impact of re-estimation is negligible, it can help to improve the prediction accuracy

930

T. Dahlhaus et al. / International Journal of Forecasting 33 (2017) 915–935

in Q(−1) M1, as can be seen in Figure A4 of the online appendix. Interestingly, re-estimating the parameters following new data releases can also worsen the nowcast accuracy (for instance, in Q(−1) M2). Finally, data releases after Q(0) M3 have very little impact on either the nowcast or the RMSE. Fig. 14 shows the results for Russia. Looking at the average absolute contributions of different data releases, we find that the news from soft indicators, financial indicators and PMIs are robust contributors to GDP nowcast revisions. Further, GDP releases early in the nowcast period do not change the predictions much, and nor do data releases in the couple of months prior to the actual GDP release. Finally, re-estimation moves the GDP nowcast much more for Russia than in the cases of the other four EMEs, in average absolute terms. It also plays an important role in improving the nowcast accuracy (see Figure A5 in the online appendix). This may reflect the relatively volatile nature of Russian data. Finally, Table 8 summarizes the results discussed above, comparing indicators’ relevance across countries. Data releases are ranked according to the associated reduction (or lack thereof) in RMSEs, as shown in Figures A1–A5. We obtain the ranking by simply ordering the indicator group whose release led to the largest decrease in the RMSE first, down to the one whose release led to the smallest decrease (or largest increase) last. For the sake of brevity, we only present the three most important data release groups in each case. One common feature of data releases across the BRIC+M is the importance of the financial, exogenous and soft indicator groups early in the nowcasting period. The previous quarter’s GDP release is also highly relevant for all countries in terms of improving the nowcast accuracy. Apart from these commonalities, the importance of data releases seems to vary considerably across the five EMEs. For example, the release of PMI data seems to be among the most important throughout the nowcast period for Russia, while it never figures in the top three releases in the case of Brazil. Also, BoP releases only play a large role in improving the nowcast accuracy in the cases of Mexico and India. Furthermore, IP is one of the most relevant indicator groups throughout the nowcasting period for China, where it accounts for almost 50% of GDP. In contrast, it is important only later in the nowcasting period for Brazil, and is even less important in the cases of Russia, India, and Mexico. News pertaining to the release of vehicle production is found to improve the nowcasts in the cases of the Latin American countries, while price data releases matter in the cases of Russia, India, and Mexico.16 In summary, there seems to be a considerable degree of heterogeneity as to the effects of different indicator releases on GDP nowcasts across the BRIC+M. However, that being said, financial indicators play a dominant role for all of the EMEs under consideration early in the nowcasting period. We also find data releases for exogenous variables, such as the oil price and the S&P 500, to be important in explaining nowcast revisions (especially early on). Nevertheless, on average, news pertaining to domestic indicators 16 Note that PMI, BoP, and vehicle indicators are not included in the data set in the cases of Mexico, Russia and India, respectively.

Table 8 Ranking data releases (top three). Brazil

Russia

India

China

Mexico

Q(−1) M1

Financial Exog Ind

Exog Financial Ind

Financial Exog Prices

Financial GDP IP

BoP Financial Prices

Q(−1) M2

Financial Ind Vehicle

Exog Ind Prices

Financial BoP PMI

Financial PMI Exog

Ind Financial Vehicle

Q(−1) M3

Financial Ind Exog

PMI Exog IP

Financial BoP Prices

Financial IP Exog

Exog Ind Vehicle

Q(0) M1

Exog Ind Financial

PMI Exog Prices

Exog Prices BoP

GDP Financial IP

Exog BoP Financial

Q(0) M2

BoP Ind Exog

PMI Prices Ind

Financial Prices PMI

IP Exog Financial

Ind Prices GDP

Q(0) M3

IP Vehicle BoP

GDP PMI IP

Exog Financial IP

Financial IP Ind

BoP Ind Exog

Q(+1) M1

IP Vehicle Financial

Ind PMI Prices

Exog BoP Prices

GDP IP PMI

BoP Labour Vehicle

Q(+1) M2

GDP IP Ind

PMI Ind Financial

GDP BoP PMI

Q(+1) M3

GDP IP Prices

GDP Vehicle Exog

Notes: Data releases are ranked according to their associated RMSE reduction (or lack thereof; see Figures A1–A5 of the online Appendix). We list the indicator group whose release led to the greatest decrease in RMSE first, and the one whose release led to the smallest decrease (or biggest increase) last. For the sake of brevity, we only present the three most important data releases here, but the complete ranking for all indicator groups is available in Table A1 of the online appendix.

is by far the biggest contributor to nowcast revisions of GDP growth in our sample. 4.4. A few caveats While our approach provides reliable predictions for the BRIC+M, it is subject to certain caveats. First, there are potential issues in relation to the relatively small time series dimension of our data set. As was shown by Doz et al. (2012), the effects of misspecification, such as omitted cross-sectional and serial correlation in the idiosyncratic component, are negligible when the sample size is large. However, misspecification could cause issues in small samples such as ours where the time series dimension of the data is constrained by the data availability for emerging markets. Furthermore, including too many variables when the time series dimension of the data is relatively small is likely to introduce estimation uncertainty, which may also affect the nowcast performance. We try to mitigate these potential issues in two ways: first, we work with small to medium-sized DFMs when predicting GDP growth in the BRIC+M; second, we allow for serial correlation of the idiosyncratic component, which improves the nowcast accuracy for China, Brazil and Russia (see Section 3.3).

T. Dahlhaus et al. / International Journal of Forecasting 33 (2017) 915–935

The second caveat relates to the assumption embedded in our DFMs that the joint dynamics of GDP and its predictors can be captured by a few unobservable factors. In general, data for emerging markets are noisier and of a poorer quality than data for advanced economies, which makes it challenging for any model to deliver reliable predictions of the GDP growth. Finally, the DFM used in our empirical analysis is linear, with constant parameters. This could be an important limitation, given the ongoing structural changes that are observed in several emerging markets.17 Taking potential parameter instabilities into account is beyond the scope of this paper, but would be an interesting avenue for further research. 5. Additional analyses 5.1. Aggregate BRIC+M nowcast One useful feature of our approach is that the individual country nowcasts can be combined successfully to produce a BRIC+M aggregate nowcast, which gives a glimpse of macroeconomic developments across emerging markets. This aggregate is constructed using purchasing power parity (PPP) weights provided by the International Monetary Fund. All of the individual nowcast results are converted into a year-over-year format to allow for aggregation. The aggregate BRIC+M results display many of the salient features of the individual country results: the aggregate nowcast tracks the dynamics of the aggregate BRIC+M GDP well over the evaluation period, and the nowcasts outperform both the statistical and private-sector (Bloomberg) benchmarks (Fig. 15).18 5.2. Evidence from real-time data The data revisions for emerging markets are often larger than those for advanced economies, such as the United States. Table 9 illustrates this by comparing the sizes of the revisions for selected macroeconomic variables for the United States with those for Brazil and Mexico. All data are from the OECD Original Release Data and Revisions Database.19 We calculate the revision for each variable as the difference between the first and last releases of that variable, where the last release is that available in 2016Q2. The sample period is from September 2001 to March 2014. As can be seen from the table, the average absolute revisions for Brazil and Mexico are at least twice as large as those for the United States for most of the variables considered. Further, Brazilian data are revised more heavily than Mexican data, with the Brazilian IP and retail trade growth being revised by as much 5.4% and 17 However, Stock and Watson (2008) showed that factor models can be robust to the presence of structural instabilities in some cases. 18 When comparing private-sector and model-based nowcasts, it should be recalled that private-sector nowcasts are made in real time, while our exercise abstracts from data revisions. 19 See http://stats.oecd.org/mei/default.asp?rev=1.

931

Table 9 Summary statistics for real-time data revisions: US vs. Brazil and Mexico. Source: OCED Original Release Data and Revisions Database and authors’ calculations. US Brazil Mexico GDP

IP

Retail trade

Exports

Imports

Average Average (absolute) Standard deviation

−0.10 0.27 0.33

0.24 0.61 0.70

0.01 1.80 2.52

Average Average (absolute) Standard deviation

−0.04

−0.01

−0.02

0.31 0.40

5.42 6.34

0.72 0.93

Average Average (absolute) Standard deviation

−0.08

0.01 3.48 7.64

0.03 1.40 1.99

Average Average (absolute) Standard deviation

−0.18 1.01 1.49

0.13 3.29 4.54

0.66 2.46 3.27

Average Average (absolute) Standard deviation

0.00 1.02 1.55

0.54 4.06 5.32

0.38 2.34 3.31

0.86 1.66

Notes: Revisions are defined as the difference between the first release for a particular variable and its last release as of 2016Q2. All series are expressed in terms of growth rates, and all numbers are in percentage points. The sample is from September 2001 to March 2014.

3.5% points, respectively, in average absolute terms. The standard deviations of data revisions for Brazil and Mexico are also generally at least twice as large as those for US variables. Thus, given the importance of data revisions for emerging markets, we now test whether the DFMs are robust to data revisions. However, data availability limitations place significant constraints on such an empirical exercise. Realtime vintages are not available for all of the series in our BRIC+M data panels, and thus we cannot perform pure real-time evaluations for these countries. For Brazil, Russia, India and Mexico, the OECD reports real time information for a small number of series, namely GDP, industrial production, retail trade (Brazil only), exports and imports. However, the quality of the real GDP data that are available for Russia and India poses significant challenges for our empirical exercise.20 Thus, for the DFM setup, we use the above set of real-time indicators for Brazil and Mexico only. We augment the real-time indicator data with a set of variables that normally are not revised, e.g., financial variables, such as the exchange rate and the stock market indexes. For Brazil, the estimation sample is December 2004 to March 2014, while for Mexico it is July 2006 to March 2014. The out-of-sample evaluation period is March 2011 to March 2014. Finally, we compare the performances of the smallerscale real-time models to that of the AR(2) benchmark and the Bloomberg forecast. Fig. 16 compares the nowcasting performance of the real-time DFM with that of the AR(2) model. For Brazil, the ratio of the RMSEs from the DFM to those from the AR(2) benchmark is almost always lower than one. In the 20 In the cases of Russia and India, poor data quality, such as inconsistencies in the seasonal adjustment or missing GDP revisions, restricts the sample size such that no meaningful empirical analysis is possible. In a recent contribution, Bragoli and Fosten (2016) constructed a new series for real-time Indian GDP since May 2009 using press releases from the Central Statistics Office.

932

T. Dahlhaus et al. / International Journal of Forecasting 33 (2017) 915–935

Fig. 15. Out-of-sample nowcasts for aggregated BRIC+M GDP growth. Notes: In the cases of Brazil, Russia, India and Mexico, the late DFM nowcast refers to the nowcast as of Q(+1) M1 D30. For China, the late nowcast is made at Q(0) M3 D30. (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article.)

Fig. 16. Real-time DFM vs. AR(2) nowcast: pseudo out-of-sample RMSE ratios. (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article.)

Fig. 17. DFM nowcast performance vs. Bloomberg mean forecast: Brazil. (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article.)

case of Mexico, though, the DFM does not outperform the AR(2) model consistently, but mainly at the later nowcast

origins. A similar picture emerges when comparing the DFM nowcasts with the Bloomberg forecasts when both

T. Dahlhaus et al. / International Journal of Forecasting 33 (2017) 915–935

933

Fig. 18. DFM nowcast performance vs. Bloomberg mean forecast: Mexico. (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article.)

Fig. 19. Simple DFM vs. AR (2): pseudo out-of-sample RMSE ratios. (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article.)

are made at the end of the quarter of interest, i.e, Q(0) M3 (Figs. 17 and 18).21 For Brazil, the real-time DFM provides a more accurate nowcast than the corresponding Bloomberg forecasts over much of the evaluation period, which is quite a high bar to meet. The model does not perform so well for Mexico, but still gives more precise nowcasts than Bloomberg for about half of the predictions made in the evaluation sample. Further, the DFMs for Brazil and Mexico track the realized GDP growth fairly well. This is not surprising, since the literature on factor models has shown that these models are robust to data revisions, as revision errors are idiosyncratic by nature and do not affect the factor estimation (see Giannone et al., 2008). 5.3. Robustness checks This section tests the robustness of our DFMs by comparing the results for our preferred models with those 21 The Bloomberg forecasts for Brazil and Mexico are only available in the year-over-year format at the quarterly frequency. We compare these forecasts with those from our DFMs by converting the quarterover-quarter nowcasts into the year-over-year format. Specifically, we first obtain a nowcast of the real-time output level using our growth rate predictions, then calculate the year-over-year growth rates for this projected output level for each quarter.

obtained from simpler DFM specifications. For each country, we estimated a ‘simple’ DFM with one factor and two lags. We kept things simple by assuming the idiosyncratic components of the monthly variables to be serially uncorrelated. Fig. 19 shows the ratio of the RMSE from the ‘simple’ DFM to that obtained from the naive AR(2) benchmark model, while Fig. 20 shows the corresponding ratios for our preferred DFM specifications vs. the ‘simple’ DFM for each country. Overall, we find that the simple models beat the AR(2) benchmarks, except for Russia, and for China later on in the nowcasting period (Fig. 19). In particular, the simple DFM model beats the AR(2) benchmark across all nowcast origins in the case of Brazil. However, our preferred DFM specification is superior to the simple DFM benchmark for nowcasting and backcasting (Fig. 20). In the case of Mexico, the simple DFM performs better than the AR(2) but not better than our preferred DFM. For China, we find that adding an idiosyncratic AR(1) component (as in our preferred DFM specification) improves the nowcast performance relative to the simple DFM specification. Finally, in the case of Russia, the performance of the simple DFM is inferior to both the AR(2) benchmark and our preferred DFM specification. Thus, taken together, the evidence presented in Figs. 19 and 20 supports our claim

934

T. Dahlhaus et al. / International Journal of Forecasting 33 (2017) 915–935

Fig. 20. Preferred DFM vs. simple DFM: pseudo out-of-sample RMSE ratios. (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article.)

that a more nuanced, country-specific choice of parameter values based on our out-of-sample evaluation exercise does better than a uniform parsimonious specification for all countries. The model specifications reported in Table 7 have been chosen in such a way as to provide the most reasonable GDP nowcasts across multiple forecast origins. At the same time, though, these results show that even a parsimonious and simple DFM provides accurate forecasts that are generally better than those of simple benchmark models across the BRIC+M. 6. Conclusion This paper provides nowcasts of real GDP growth for the BRIC+M group based on state-of-the-art dynamic factor models. A number of key findings emerge from this study. First, the DFMs generally display good directional accuracies, and perform well for the BRIC+M in terms of capturing GDP dynamics. Second, the accuracy of the nowcasts obtained from the DFMs tends to improve as more data become available. However, data releases that occur late in the nowcasting period are generally of little importance. Third, our results show a considerable degree of heterogeneity in the way in which news pertaining to different indicator groups affects the nowcasts of GDP growth across the five EMEs in our sample. One common finding is that releases of domestic indicators are the main drivers of nowcast revisions, while the role of exogenous variables is only relevant early in the nowcasting period. This may reflect the relatively closed nature of these economies. Furthermore, news pertaining to financial variables and soft indicators is highly relevant early in the nowcasting period. Finally, parameter re-estimation generally improves the accuracy of the GDP nowcast across the BRIC+M, demonstrating the model’s usefulness when working with the more volatile data that tends to arise from emerging markets. Our paper also highlights some interesting implications for researchers. We find that the nowcasting approach that was developed originally for advanced economies is also

applicable to EMEs, despite their differences in terms of both size and economic structure. One potential interpretation of this result is that all economies share similar features regarding their business cycle fluctuations, which in turn allows a simple model to capture the salient features of the data. Acknowledgments We thank Marta Banbura for sharing her Matlab code and Domenico Giannone for his insightful comments. We would also like to thank Michael McCracken and three anonymous referees, as well as participants at the 25th Annual Meeting of the Midwest Econometrics Group and the 2015 CIRANO Real-Time Workshop, and seminar participants at the Bank of Canada for helpful comments and suggestions. Bryce Shelton provided excellent research assistance. Appendix A. Supplementary data Supplementary material related to this article can be found online at http://dx.doi.org/10.1016/j.ijforecast.2017. 05.002. References Angelini, E., Camba-Mendez, G., Giannone, D., Reichlin, L., & Runstler, G. (2011). Short-term forecasts of euro area GDP growth. The Econometrics Journal, 14, C25–C44. Astveit, K., & Trovik, T. (2012). Nowcasting Norwegian GDP: The role of asset prices in a small open economy. Empirical Economics, 42, 95–119. Banbura, M., Giannone, D., Modugno, M., & Reichlin, L. (2013). Handbook of economic forecasting: Vol. 2. Now-casting and the real-time data flow (pp. 195–237). Elsevier. Banbura, M., Giannone, D., & Reichlin, L. (2010a). Large Bayesian vector auto regressions. Journal of Applied Econometrics, 25(1), 71–92. Banbura, M., Giannone, D., & Reichlin, L. (2010b). Nowcasting, European Central Bank working paper series no. 1275, European Central Bank. Banbura, M., & Modugno, M. (2014). Maximum likelihood estimation of factor models on data sets with arbitrary pattern of missing data. Journal of Applied Econometrics, 29(1), 133–160.

T. Dahlhaus et al. / International Journal of Forecasting 33 (2017) 915–935 Banbura, M., & Runstler, G. (2011). A look into the factor model black box: Publication lags and the role of hard and soft data in forecasting GDP. International Journal of Forecasting, 27(2), 333–346. Barhoumi, K., Darne, O., & Ferrara, L. (2010). Are disaggregate data useful for factor analysis in forecasting French GDP? Journal of Forecasting, 29(1–2), 132–144. Bhattacharya, R., Pandey, R., & Veronese, G. (2011). Tracking India growth in real time, Working Papers 11/90, National Institute of Public Finance and Policy. URL: https://ideas.repec.org/p/npf/wpaper/11-90. html. Bragoli, D. & Fosten, J. (2016). Nowcasting Indian GDP, University of East Anglia School of Economics Working Paper Series 2016-06, School of Economics, University of East Anglia, Norwich, UK. Bragoli, D., Metelli, L., & Modugno, M. (2014). The importance of updating: Evidence from a Brazilian nowcasting model, Feds working paper no. 2014-94, Federal Reserve Board, Washington, D.C. Caruso, A. (2015). Nowcasting Mexican GDP, ECARES working paper 201540. D’Agostino, A., McQuinn, K., & O’Brien, D. (2012). Now-casting Irish GDP. Journal of Business Cycle Measurement and Analysis, 2, 21–31. Doz, C., Giannone, D., & Reichlin, L. (2012). A quasi-maximum likelihood approach for large, approximate dynamic factor models. The Review of Economics and Statistics, 94(4), 1014–1024. Forni, M., Hallin, M., Lippi, M., & Reichlin, L. (2000). The generalized dynamic factor model: Identification and estimation. The Review of Economics and Statistics, 82(4), 540–554. Forni, M., & Lippi, M. (2001). The generalized dynamic factor model: representation theory. Econometric Theory, 17(06), 1113–1141. Giannone, D., Miranda-Agrippino, S., & Modugno, M. (2013). Nowcasting China real GDP, Technical report. Giannone, D., Reichlin, L., & Small, D. (2008). Nowcasting: The real-time informational content of macroeconomic data. Journal of Monetary Economics, 55(4), 665–676. Liu, P., Matheson, T., & Romeu, R. (2012). Real-time forecasts of economic activity for Latin American economies. Economic Modelling, 29, 1090–1098. Luciani, M. (2014). Forecasting with approximate dynamic factor models: The role of non-pervasive shocks. International Journal of Forecasting, 30(1), 20–29. Luciani, M., Pundit, M., Ramayandi, A., & Veronese, G. (2015). Nowcasting Indonesia, Finance and economics discussion series 2015-100, Washington: Board of Governors of the Federal Re- serve System. Luciani, M., & Ricci, L. (2014). Nowcasting Norway. International Journal of Central Banking, 10(4), 215–248. Mariano, R. S., & Murasawa, Y. (2003). A new coincident index of business cycles based on monthly and quarterly series. Journal of Applied Econometrics, 18, 427–443.

935

Matheson, T. D. (2010). An analysis of the informational content of New Zealand data releases: The importance of business opinion surveys. Economic Modelling, 27, 304–314. Porshakov, A., Deryugina, E., Ponomarenko, A., & Sinyakov, A. (2015). Nowcasting and short-term forecasting of Russian GDP with a dynamic factor model, BOFIT discussion papers 2015–19. Runstler, G., Barhoumi, K., Cristadoro, R., Den Reijer, A., Jakaitiene, A., Jelonek, P., Rua, A., Ruth, K., Benk, S., & Van Nieuwenhuyze, C. (2009). Short-term forecasting of GDP using large monthly data sets: a pseudo real-time forecast evaluation exercise. Journal of Forecasting, 28(7), 595–611. Siliverstovs, B. & Kholodilin, K. (2010). Assessing the real-time informational content of macroe- conomic data releases for now-/forecasting GDP: Evidence for Switzerland, Discussion papers of DIW Berlin, DIW Berlin. Stock, J., & Watson, M. (2002). Forecasting using principal components from a large number of predictors. Journal of the American Statistical Association, 97, 1167–1179. Stock, J., & Watson, M. (2008). Forecasting in dynamic factor models subject to structural instability. In J. Castle, & N. Shephard (Eds.), The methodology and practice of econometrics, a festschrift in honour of professor David F. Hendry. Oxford University Press.

Tatjana Dahlhaus is a Principal Researcher in the Emerging Markets Division of the International Economic Analysis Department at the Bank of Canada. Her research interests lie in the fields of applied macroeconomics and econometrics. Currently, she focuses on studying the transmission of monetary policy in linear as well as non-linear environments. Tatjana received her Ph.D. in Economics from Universitat Autonoma de Barcelona, Spain.

Justin-Damien Guénette is a Principal Economist in the United States Division of the International Economic Analysis Department at the Bank of Canada. His primary interests include emerging markets and energy commodities. He obtained his Masters in Global Governance and Economics from the University of Waterloo.

Garima Vasishtha is a Senior Policy Advisor in the International Economic Analysis Department at the Bank of Canada. Her research interests lie in the broad fields of international finance and open economy macroeconomics. She has conducted research on issues related to the transmission of international shocks, sovereign spreads in emerging markets, crossborder financial flows, and determinants of movements in commodity prices. She holds a Ph.D. in International Economics from the University of California, Santa Cruz.