Nowcasting Indonesia

Finance and Economics Discussion Series Divisions of Research & Statistics and Monetary Affairs Federal Reserve Board, Washington, D.C. Nowcasting Indonesia Matteo Luciani, Madhavi Pundit, Arief Ramayandi, and Giovanni Veronese 2015-100 Please cite this paper as: Luciani, Matteo, Madhavi Pundit, Arief Ramayandi, and Giovanni Veronese (2015). “NowcastingIndonesia,”FinanceandEconomicsDiscussionSeries2015-100. Washington: Board of Governors of the Federal Reserve System, http://dx.doi.org/10.17016/FEDS.2015.100. NOTE: Staff working papers in the Finance and Economics Discussion Series (FEDS) are preliminary materials circulated to stimulate discussion and critical comment. The analysis and conclusions set forth are those of the authors and do not indicate concurrence by other members of the research staff or the Board of Governors. References in publications to the Finance and Economics Discussion Series (other than acknowledgement) should be cleared with the author(s) to protect the tentative character of these papers.

∗ Nowcasting Indonesia Matteo Luciani Madhavi Pundit Federal Reserve Board Asian Development Bank matteo.luciani@frb.gov mpundit@adb.org Arief Ramayandi Giovanni Veronese Asian Development Bank Banca d’Italia aramyandi@adb.org giovanni.veronese@bancaditalia.it September 2015 Abstract We produce predictions of the current state of the Indonesian economy by estimating a dynamic factor model on a dataset of eleven indicators (also followed closely by market operators) over the time period 2002 to 2014. Besides the standard difficulties associated with constructing timely indicators of current economic conditions, Indonesia presents additional challenges typical to emerging market economies where data are often scant and unreliable. By means of a pseudo-real-time forecasting exercise we show that our model outperforms univariate benchmarks, and it does comparably with predictions of market operators. Finally, we show that when quality of data is low, acarefulselectionofindicatorsiscrucialforbetterforecastperformance. JEL codes: C32, C53, E37, O53 Keywords: Nowcasting,DynamicFactorModels,EmergingMarketEconomies ∗ We would like to thank Dennis Sorino for excellent research assistance and participants of EconomicsandResearchandRegionalCooperationDepartment,ADBseminarseriesforvaluable comments. The project was prepared under RDTA8951: Macroeconomic modeling for improved economic assessment. This paper was written while Matteo Luciani was charg´e de recherches F.R.S.-F.N.R.S., and gratefully acknowledges their financial support. Of course, any errors are our responsibility. Disclaimer: theviewsexpressedinthispaperarethoseoftheauthorsanddonotnecessarilyreflect the views and policies of the Asian Development Bank, of the Banca d’Italia or the Eurosystem, and of the Board of Governors or the Federal Reserve System.

1 Introduction It is well known that macroeconomic data are released with a substantial delay. Additionally, in emerging market economies, low frequency data, i.e. annual national accounts, rely on a smaller array of surveys and indicators than in advanced economies, and provide a partial picture of the economy. However, complete and up to date information on the current state of the economy is crucial for policy makers, market participants and public institutions. Indeed, agents periodically update their forecasts, and monitoring economic conditions in real-time helps them to assess whether the forecasts are on track or need to be revised. Similarly, the process of policymaking often requires long term projections of the economy that heavily rely on accurate initial conditions and forecasts. Therefore, constructing timely “predictions” of current economic conditions, namely nowcasts, is of fundamental importance for decision making. A lot of information is contained in economic indicators that are available on a quarterly, monthly, weekly and even daily basis, and in principle it is possible to use this information to build “predictions” of the current state of the economy. However, high frequency data in emerging market economies are often scant, noisy, releasedwithalag, andcanhavemissinginformation. Thiscomplicatesthedifficult task of real-time monitoring and decision making, particularly in an environment wheregrowthvolatilityistypicallyhigh, andwherethereisconsiderableuncertainty surrounding trend growth as it may undergo changes due to rapid catching-up phases or persistent slowdowns. In other words, in addition to standard problems, constructing timely indicators on current economic conditions for emerging market economies presents some extra challenges. In this paper we focus on Indonesia, the largest economy in Southeast Asia which is rapidly gaining influence in the world economy. With a number of high frequency data indicators available and yet facing problems that commonly plague emerging economy datasets, Indonesia provides an interesting training case for developing a nowcasting framework that can be applied to monitor other similar economies in the region. Two main issues emerge with regard to monitoring in real-time: how many and whichindicatorstoselect,andwhateconometricmodeltousetoextractinformation from the data. In this paper, we produce “predictions” of the current state of the Indonesian economy by estimating a dynamic factor model on a dataset of eleven indicators (also followed closely by market operators) over the time period 2002 to 2014. Our choice of the model is based on the fact that it is parsimonious and is able to cope with missing data and mixed frequency indicators; and can potentially be estimated on a large number of variables. Further, since the seminal paper of Giannone et al. (2008) this model has become a standard tool for monitoring economic activity, as it has proved to be successful in nowcasting several economies, including emerging ones such as China (Giannone et al., 2014) and Brazil (Bragoli et al., 2014). 2

The rest of the paper proceeds as follows: Section 2 presents Indonesia’s GDP data, and discusses the problems of having several GDP series with different base years, and no official seasonally adjusted data. Section 3 discusses our nowcasting procedures. This section is divided in two parts: in the first part we describe the process of choosing a set of indicators that contains useful information on economic activity, and in the second we present the application of a dynamic factor model to Indonesia’s data. The evaluation of our model is presented in Section 4. Several results emerge. First, incorporating high frequency data in a rigorous framework leads to an improvement in the forecast accuracy of Indonesia’s economy compared to simple univariate benchmarks. Second, too many variables are not always optimal for the purpose of monitoring as they can be noisy or uninformative (see also Ban´bura et al., 2013; Luciani, 2014b), particularly so when the target variable, namely Indonesia’s GDP growth, has limited number of observations. A careful selection of meaningfulvariablesimprovestheforecastperformance. Third, ourmodeldoeswell in predicting quarterly GDP growth when compared to private forecasters such as Bloomberg, and also does well in predicting annual GDP growth when compared to institutional forecasts of the International Monetary Fund and the Asian Development Bank. Finally, Section 5 concludes. 2 Indonesia’s GDP data: Patterns and issues In this Section we present Indonesia’s GDP data, and discuss a number of issues in the data that need to be carefully tackled even before starting any monitoring process. The first is that there is no single long series available from official statistical sources. The left hand side chart in Figure 1 plots the level of GDP at constant prices in trillion rupiah from 1993 to 2014, where the three lines refer to GDP across different base years, 1993, 2000 and 2010 respectively. The aggregation methodology was common between the 1993 and 2000 base years, but changed from SNA 1993 to SNA 2008 for the 2010 base year.1 Base changes are common for GDP data as they can incorporate changes in the economy’s structural composition. However the strikingly different slopes among the lines despite a few years of overlaps between series suggests that substantial revisions in data releases affect not only the level of GDP but also its growth rate.2 This raises questions on the composition of the aggregate series and methodologies used in the construction, which is exacerbated by a lack of publicly available information on procedures used. The right plot in Figure 1 shows the year-on-year (y-o-y) growth rate of quarterly 1 A detailed explanation of the System of National Accounts (SNA) can be found in http://unstats.un.org/unsd/nationalaccount/sna.asp. 2 In fact even nominal GDP data for the overlapping time periods are not comparable between the different bases for GDP series. Data are available in Table VII.1 at http://www.bi.go.id/en/statistik/seki/terkini/riil/Contents/Default.aspx. 3

Figure 1 Gross Domestic Product series for Indonesia haipuR noillirT Real GDP 9000 8000 7000 6000 5000 4000 3000 2000 1000 2010p 2000p 1993p 0 1993 1995 1997 1999 2001 2003 2005 2007 2009 2011 2013 2015 Real GDP y−o−y Growth 15 10 5 0 −5 −10 −15 2010p 2000p 1993p −2109 94 1996 1998 2000 2002 2004 2006 2008 2010 2012 2014 Notes:ThefigureontheleftplotsannualGDPatconstantpricesinlevelsstartingin1993,andthethreelinesrepresentthedifferent baseyears. TherewasalsoanaccountingchangefromSNA1993toSNA2008. Thefigureontherightplotstheyear-on-yeargrowth rateofquarterlyGDPatconstantprices,againcorrespondingtothethreebaseyears. GDP. As we can see, GDP growth is characterized by large fluctuations due to different crises hitting the economy. In particular, the Asian financial crisis in the late 1990s stands out as a unique episode for Indonesia’s growth path with GDP falling by more than 15% in 1998. Then, since 2002 the growth rate stabilized somewhat at a yearly rate of around 5%, with low volatility. As a consequence of this pattern, we exclude the Asian financial crisis years from our sample and use the series only from 2002. This is unavoidable because the evolution of GDP growth over the financial crisis would dominate our estimates if not excluded. It would capture features of the data (i.e. co-movements) that are potentially spurious and misleading for the nowcasting exercise since the relationships that held during the financial crisis could be different from those that hold during expansions and more moderate recessions. The second big issue with Indonesia’s GDP data is that the growth series exhibits a marked seasonal pattern (Figure 2), and yet, there is no seasonally adjusted GDP series available from official sources. So this leaves us with the problem of having to deal with seasonality in the data, particularly when trying to combine series with different base years to obtain coherent and long time series. Indeed, as shown in the right-hand side panel in Figure 2 which plots the quarter-on-quarter (q-o-q) growth rates, the seasonality of GDP data with 2010 base year exhibits a clear departure from the seasonal pattern in the series with 2000 base year. This adds to the complication of splicing the different series together. Suppose we start with the latest available GDP series which is of 2010 base, and extend it backwards using the q-o-q growth rates of the previous base series. As seen in the left hand side panel of Figure 3, it results in inconsistent seasonal patterns within the spliced series, i.e., between the actual data and the extended data. On the contrary, the reconstructed series based on y-o-y growth rates does not seem to have the same defect. In order to deal with the two issues highlighted, we construct a long time series for GDP by using y-o-y growth rates as shown in the right hand side panel of Figure 3. Furthermore, in our analysis going forward we continue to use y-o-y growth rates. We are well aware that using y-o-y growth makes the series smoother as it effectively tackles the issue of seasonality in the data, but it also lags quarter-on-quarter 4

Figure 2 Quarter-on-quarter growth of real GDP Seasonal fluctuations: Real GDP growth Real GDP growth 10 5 8 4 6 3 4 2 2 1 0 0 −2 −1 −4 −2 −6 −8 R M e e a a l n G s D by P q q u − a o r − te q r −3 2 2 0 0 0 1 0 0 p p −10 Q1 Q2 Q3 Q4 − 2 4 0 07 2008 2009 2010 2011 2012 2013 2014 2015 Note: Uses real GDP data reconstructed from y−o−y Notes: The figure on the left side shows the seasonal fluctuations in q-o-q GDP growth, where real GDP data is reconstructed usingy-o-ygrowthrates. Thefigureontherightsideplotsq-o-qgrowthratesofrealGDPseriesavailable,basedontwodifferent accountingmethodologies,SNA1993andSNA2008andtwodifferentbaseyears,2000and2010. growth. However, the alternative of using some standard procedure to seasonally adjust the series brings about non-trivial problems: first, any technique to eliminate seasonal fluctuations in the data would introduce an estimation bias associated with the specific procedure being utilized; second, it would not be possible to compare our “predictions” with any credible benchmark. Conversely, using y-o-y growth provides us with a comparable platform with many forecasters such as IMF, ADB and others who look at annual growth rates. Figure 3 Reconstructing GDP series using y-o-y growth Real GDP q−o−q growth Reconstructed GDP Series 5 8 4 7.5 3 7 2 1 6.5 0 6 −1 5.5 −2 5 −3 −4 Reconstructed from y−o−y 4.5 R Pr e e c v o i n o s u t s r u b c a t s e e d from y−o−y Reconstructed from q−o−q Reconstructed from q−o−q −5 4 2007 2008 2009 2010 2011 2012 2013 2014 2015 2007 2008 2009 2010 2011 2012 2013 2014 2015 Notes: Sincethelatestavailabledatastartsin2010,wehavetoreconstructalongtimeseriesandtheleftsidegraphdisplaysthe growthpatternsusingq-o-qandy-o-ygrowthrates. Thefigureontherightsideplotsthereconstructedseriesusingdifferentgrowth rates. 3 Nowcasting Typically GDP data provide the most comprehensive picture of the economy, by aggregating activity of different sectors. Unfortunately, the data come with long delays and are not available on a high frequency basis. In most cases, it is published quarterly and in some developing countries, even annually. For Indonesia, GDP data is reported quarterly with a delay of about five weeks. This means that the growth rate of the economy in the first quarter that ends in 5

March is not known until the second week of May. While this is a reasonable delay compared to most emerging and some advanced economies, it is still insufficient for the purpose of real-time monitoring as it is not possible to make an assessment of the strength of economic activity for almost five months into the year. Toassesshighfrequencymovementsineconomicactivityinemergingmarketeconomies, analysts typically use the industrial production index as a proxy for output as it is available on a monthly basis.3 The use of this index relies on the assumption that movements in the industrial sector are a good approximation of aggregate economic activity. While the assumption may hold well for some economies, in others, the cyclical component of the index is not found to be sufficiently synchronized with GDP (Fulop and Gyomai, 2012). A close examination on the relation between Indonesia’s GDP growth and its y-o-y growth rate of industrial production shows that the two are only weakly linked (Figure 4). The correlation coefficient between the two series is just 0.35, which suggests that industrial production in Indonesia may not be a reliable proxy for GDP. Figure 4 Gross Domestic Product & Industrial Production 2 1.5 1 0.5 0 −0.5 −1 −1.5 −2 −2.5 −3 GDP IP 2002 2003 2004 2005 2006 2007 2008 2009 2010 2011 2012 2013 2014 2015 Notes: Theblacklineisy-o-yGDPgrowthrate,whilethegreylineisy-o-ygrowthrateofIndustrialProductionIndex. Thedata forIndustrialProductionwereconvertedtoquarterlyfrequencybyaveragingmonthlyobservations. To gauge the current state of the Indonesian economy in real-time, we need to construct a prediction of GDP growth before the official data is released. This means that at each point in time we want to predict not only the current and next quarter estimates for GDP growth (henceforth nowcast and forecasts), but also, wherever the official data has not been published yet, the past quarter GDP growth (backcast). Ideally, if the true data sources and compilation methods for GDP were known, we could simply attempt to reverse engineer the process performed quarterly by the Indonesian statistical office, but at a higher frequency. Unfortunately though, as discussed in Section 2, very little information on these methods is available from the 3 See, for example, Ma´ckowiak (2007) and Raghavan and Dungey (2015) for applications to a set of emerging economies, and Kasri and Kassim (2009) and Kubo (2009) for Indonesia specifically. 6

statistical agency. This forces us (1) to build an information set that is informative for describing the GDP growth process, possibly by including variables which may be outside the scope of the statistical office, but that may still contain useful leading information, and (2) to choose an econometric model to build our prediction. In the next two subsections we address these issues. In particular, in Section 3.1 we explain how we construct the database, while in Section 3.2 we introduce Dynamic Factor models (DFMs). We choose to use a DFM since it is a parsimonious model that is able to cope with missing data, mixed frequency, and potentially can be estimated on a large number of variables. Furthermore, this model proved to be very successful in nowcasting several economies, including emerging ones.4 3.1 What variables to select? There is a wide set of potentially useful monthly and quarterly series which could help to extract information on the state of the Indonesian economy. However, the limited number of time series observations available for our target, quarterly yearon-year GDP growth (see Section 2), constrains our choice of both the variables and the model. In principle, Dynamic Factor models are consistently estimated when the number of variables is diverging to infinity, and in practice they are usually estimated on relatively large datasets. However, when using a small number of time series observations, if the sample size is severely limited, then including too many variables is likely to introduce a lot of estimation uncertainty, ultimately worsening the prediction performance. This is a particular source of concern when dealing with Indonesian data, as only few series display a marked comovement with the GDP quarterly dynamics and hence the risk is to introduce excessive noise in the model estimation.5 The questions then are: on the basis of which criteria should we select variables? And, how many variables should we select? A possible strategy to draw from a large pool of variables could be to rely on a purely mechanical statistical selection procedure. For example, Bai and Ng (2008) suggest selecting with the LARS algorithm only those variables that are really informative for forecasting the target variable, while Camacho and Perez-Quiros (2010) suggest first selecting a core group of variables, and then evaluating if other possible predictors are useful. An alternative strategy, pioneered by Ban´bura et al. (2013) and followed by Luciani and Ricci (2014), Giannone et al. (2014), and Bragoli et al. (2014), is to exploit 4 A non exhaustive list of countries and papers is: the US (Ban´bura et al., 2011, 2013; Giannone et al., 2008), the Euro Area (Angelini et al., 2011; Ban´bura and Ru¨nstler, 2011), Germany (Marcellino and Schumacher, 2010), France (Barhoumi et al., 2010), Ireland (D’Agostino et al., 2012), Norway (Aastveit and Trovik, 2012; Luciani and Ricci, 2014), China (Giannone et al., 2014), Brazil (Bragoli et al., 2014), New Zealand (Matheson, 2010), the Global Economy (Matheson, 2013), and Latin America (Liu et al., 2012) 5 Furthermore, the literature on nowcasting with large-dimensional DFMs has reached the conclusionthat,unlessoneneedtomonitorthedataflowwithmanydatareleases,thereisnoneedof a large database when forecasting with Dynamic Factor Model as long as the variables on which the model is estimated are appropriately selected (see Ban´bura et al., 2013; Luciani, 2014a,b). 7

the “revealed preferences” of professional forecasters who follow the Indonesian economy on the Bloomberg platform. These analysts subscribe to the Bloomberg news alert for specific data releases of the variables that they monitor, and use them to form their expectations on current and future fundamentals of Indonesia. Since Bloomberg constantly ranks the analysts’ demand for these alerts by constructing a relevance index for each macroeconomic indicator, we can select variables based on this relevance index.6 We adopt the latter approach, since the automatic selection approach risks leading toanunstablechoiceofvariablesinareal-timescenario.7 Thisinstabilitywouldnot only be difficult to justify from an economic standpoint, it would also complicate the interpretation of the forecasts’ revisions. Moreover, we also tried the automatic selection approach and the performance of our model is worse in this case than when we used the revealed preference approach (see the Appendix). It turns out that for Indonesia only a relatively small number of macroeconomic series are tracked in real-time by the markets (see Table 1). On the one hand, there are indicators describing macroeconomic developments (e.g. the GDP itself, car sales, exports, imports and manufacturing PMI). On the other hand, given their direct impact on the foreign exchange and fixed income markets, analysts also monitor indicators that directly describe the monetary policy stance. These are the central bank reference interest rate as well as key monetary aggregates. Starting from the set of indicators in Table 1 followed by business analysts we constructed our database as follows: 1) We excluded all those indicators that either had too few observations, or we did not manage to retrieve. This is the case of PMI for which data are available only startingfromJune2012,andofDanareksaConsumerConfidenceandMotorcycle Sales for which we were not able to retrieve data.8 2) We screened each of the remaining indicators in order to understand whether they are followed by analysts because they convey information on the state of the real economy, or they are directly related to the stance of the Central Bank and its balance sheet. Therefore, since Bank of Indonesia has an inflation target, we discarded CPI, and furthermore, we also removed Foreign Reserves, and Net Foreign Assets as they are mainly related to the foreign exchange policy. 3) We then excluded those variables that are the sum of other variables in the database or are too similar to other series. So we kept Imports and Exports, but we excluded Current Account; and we kept M1, but discarded M2. 6 The implicit assumption here is that since (also) based on their expectations on future fundamentals analysts allocate their investments, they (better) know what are the relevant series to monitor in order to form appropriate expectations on GDP growth. 7 As shown by De Mol et al. (2008) since there is a lot of comovement among macroeconomic data, the set of indicators selected with statistical criteria is extremely unstable. 8 As the index compiled by Danareksa was not available to us, we experimented with the household consumer confidence index compiled by the Bank of Indonesia. The latter however displays trending pattern which appears difficult to reconcile with the state of the economy, and we hence discarded it. 8

Table 1 Bloomberg Calendar: follow the market revealed preference Variable Reference Release Freq Rel. period date Bank Indonesia Reference Rate 17-Mar Mar-17 D 95 GDP YoY 4Q Feb-2 Q 64 PMI Mfg Markit Mar Apr-4 M 82 CPI YoY Jan Mar-6 M 86 Foreign Reserves Dec Mar-3 M 86 Trade Balance Feb Mar-15 M 23 CPI NSA MoM Feb Mar-6 M 27 CPI Core YoY Feb Mar-6 M 55 Exports YoY Feb Mar-15 M 27 GDP QoQ 4Q Feb-2 Q 59 Consumer Confidence Index Feb Mar-4 M 64 Local Auto Sales Feb Mar-16 M 50 Motorcycle Sales Feb Mar-16 M 32 Net Foreign Assets IDR Feb Mar-28 M 45 Imports YoY Feb Mar-15 M 50 Money Supply: M2 YoY Feb Mar-28 M 32 Danareksa Consumer Confidence Feb Mar-5 M 36 Money Supply M1 YoY Jan Mar-28 M 32 BoP Current Account Balance 4Q Mar-15 Q 14 Notes: Fromlefttoright: Variablereportsthenameofthevariable;Referenceperiodreportstheperiodtowhichthedatathatwill bereleasedrefersto,whileReleaseDatereportswhenthedatawillbereleased. ForexampleonFebruary2,2015,thestatistical officereleasedthedataforGDPQ42014. “Freq.”reportsatwhichfrequencythevariableispublished. “Rel.”istherelevanceindex inBloombergwhichcountsthenumberofsubscriberstothenewsalertalertingthereleaseofthevariable. 4) Finally, we use our “expert judgement” and added a few indicators that we think provide some extra information about the Indonesian economy. To this end, to capture information regarding the increasing role of construction activity for the Indonesian economy we include domestic cement consumption. To account for spillovers from the foreign sector into the domestic economy we included a variable from the very timely Markit PMI manufacturing survey. In particular we included the aggregate for emerging economies, which is dominated by developments in China as well as countries in the Asian region. Finally, we also included asectoralbreakdownoftheimportsseriestobettercapturethepossiblydifferent lead/lag characteristics of each of these series with GDP growth. By following this strategy, we end up with a data set of ten macroeconomic indicators plus GDP (see Table 2). While GDP and Business Tendency Index are quarterly series, the remaining are at monthly frequency. The column “Delay” reports the publication delay expressed in number of days in our stylized calendar, and as we can see there are substantial differences between series in terms of their publication delay. For example, the PMI for developing economies is published just four days after the reference month, while data on imports are released a month after the reference month.9 9 For the policy rate we adopted the assumption that it is observed the first day of the month followingthereferencemonth. Forexample,thepolicyrateforJanuaryitisobservedonFebruary 1. OfcoursethisisanapproximationbecauseweknowwhatthepolicyrateiseverydayinJanuary. In principle, we could have accounted for daily observations in the interest rate since DFMs allow us to do so (Modugno, 2014). However, Ban´bura et al. (2013) have shown that including data at 9

Table 2 Data Description and Data Treatment Variable Freq. Source Start Delay Trans. Central Bank policy rate M Bank Indonesia Jan-93 1 PMI developing economies M JP Morgan Apr-04 4 Cement, domestic consumption M Statistics Indonesia Jan-94 10 yoy Exports M Statistics Indonesia Jan-93 15 yoy Car sales M PT Astra Jan-93 16 yoy Imports: Consumption Goods M Statistics Indonesia Mar-01 34 yoy Imports: Capital Goods M Statistics Indonesia Mar-01 34 yoy Imports: Raw materials M Statistics Indonesia Mar-01 34 yoy Gross Domestic Product Q Statistics Indonesia Q1 1993 36 yoy Business Tendency Index Q Bank Indonesia Q2 2000 38 M2 M Bank Indonesia Jan-93 59 yoy Notes: Fromlefttoright: Variablereportsthenameofthevariable;Freq. specifieswhetheravariableismonthly(M)orquarterly (Q); Source reports the original source of the data; Start specifies since when a variable is available; Delay reports the release delayexpressedinnumberofdaysinourstylizedcalendar;and,finally,Transspecifieswhetheravariablehasbeentransformedto year-on-yeargrowthratesoritisconsideredinlevels. Datafor“PMIdevelopingeconomies”areproducedbyJPMorganandwere downloadedfromThomsonReuters,Datastream. Datafor“Carsales”areproducedbyPTAstradataandweredownloadedfrom CEICData/ISIEmergingMarkets. 3.2 Dynamic Factor models Factor models are based on the idea that macroeconomic fluctuations are the result of few macroeconomic shocks, which affect the whole economy, and a number of sectoral/regional shocks that affect a part of the economy. Therefore, each variable in the dataset can be decomposed into a common part and an idiosyncratic part, where the common part is assumed to be characterized by a small number of common factors (f ) which are time series processes meant to capture the comovement t in the data, i.e. the business cycle. Formally, let x be the i-th stationary variable observed at month t, then it x = λ f +ξ i = 1,...,n (1) it i t it where f is an r ×1 vector (with r (cid:28) n) containing the common factors, and ξ is t it the i-th idiosyncratic component. The vector of common factors evolve over time as a VAR(p) process driven by the common shocks u ∼ N(0,I ), while each idiosynt r cratic component follows an independent AR(1) model driven by the idiosyncratic shocks e : it p (cid:88) f = A f +u (2) t s t−s t s=1 ξ = ρ ξ +e . (3) it i it−1 it Equations (1)-(3) define the Dynamic Factor model used in this paper.10 The model can be estimated by Principal Components (Stock and Watson, 2002a; Bai, 2003), the daily frequency is not particularly useful for nowcasting GDP, so we adopted the convention that the interest rate is monthly and is observed on the first day after the reference month. 10 This is the model studied in Doz et al. (2011, 2012), which is a special case of the model studied in Forni et al. (2009). In this model, the common shocks and the idiosyncratic shocks are assumed to be uncorrelated at all leads and lags, while the idiosyncratic shocks are allowed to 10

by using the Kalman Filter (Doz et al., 2011), or by maximum likelihood techniques through the EM algorithm (Doz et al., 2012). In this paper we will use maximum likelihood, and in particular we will use the EM algorithm proposed by Ban´bura and Modugno (2014) which can handle both mixed frequencies and missing data. In the next section we will use the model to produce real-time predictions of Indonesian GDP growth. DFM proved very successful in real-time forecasting, and when used for this task they work as follows: suppose that we are at day d, and that at date d it is available a given vintage of data: Xd. Further suppose that on the basis of Xd we have constructed our prediction: xˆd = λ(cid:98) ˆ fd +eˆ . Now, suppose it i t t that at day d+1 a new data is released (eg. Exports). Based on this new piece of information we can check if our stand about the business cycle is still correct or if we need to revised it, which is what the DFM does automatically. More specifically, at day d + 1 we have now a new vintage of data: Xd+1. Given this new vintage we can update our estimate of the factors, ˆ fd+1, and hence update our prediction: t xˆd+1 = λ(cid:98) ˆ fd+1 +eˆ it i t t 4 Empirics 4.1 The forecasting exercise To evaluate the performance of our model, we perform a pseudo real-time out-ofsample exercise. Predictions of Indonesian GDP growth are produced according to a recursive scheme, where the first sample starts in June 2002 and ends in December 2007, while the last sample starts in June 2002 and ends in December 2014. The model is estimated at the beginning of each quarter using only information available as of the first day of the quarter, and then the parameters are held fixed until the next quarter. To perform our exercise we construct real-time vintages by replicating the pattern of data availability implied by the stylized calendar (Table 2), and every time new data are released, we update the prediction based only on information actually available at that time. We call this exercise pseudo real-time since we are not able to track the full set of data revisions, an issue that we will discuss further later. For the estimation, we include two factors (r = 2) and two lags (p = 2) in the VAR model governing the evolution over time of the factors. The choice of including two factors deserves a comment. First the literature on factor models has shown that for forecasting it suffices to include a small number of factors (eg. Stock and Watson, 2002b; Forni et al., 2003). Furthermore, recent literature on small-medium Dynamic Factor models (Ban´bura et al., 2013; Luciani and Ricci, 2014; Giannone et al., 2014; Bragoli et al., 2014) often include one factor only. Therefore, a natural be cross-sectionally correlated, albeit by a limited amount (approximate factor structure). For a more comprehensive treatment of the DFM we refer the reader to the aforementioned references and to the survey by Luciani (2014b). 11

choice would be to follow the literature and to set r = 1. However, this literature estimates models for q-o-q growth rates, while we are estimating a model for y-oy growth rates, and if the model for q-o-q growth rates has one factor, then the corresponding model for y-o-y growth rates has four factors.11 Hence, we should set r = 4, and, indeed, by looking at the eigenvalues of the covariance matrix we can see clearly three/four diverging eigenvalues. However, among these three/four eigenvaluesthefirsttwoclearlydominatesuggestingthattheothertwocarrymainly noise, which motivates our choice to set r = 2.12 4.2 Comparison against Statistical Benchmark To judge the performance of our model and to evaluate the information contained in our dataset we start by comparing our model with three benchmark models. Our first benchmark is the naive forecast, obtained from the random walk model on GDP growth: yQ = yQ +ε . The second is a forecast from an autoregressive model t t−1 t of order two on GDP growth: yQ = ρ yQ + ρ yQ + ε . Given the high persist 1 t−1 2 t−2 t tence in our target series introduced by the y-o-y transformation, these univariate benchmarks are inherently tough competitors to match in our real-time exercise. Our last benchmark is a bridge model (Parigi and Schlitzer, 1995). Bridge models predict GDP growth by using its own past plus one or more monthly indicators.13 Formally, let y be the y-o-y GDP growth observed quarterly, i.e., at month t = t 3,6,9,... and let x be a monthly variable, then the Bridge model is defined as t follows: y = µ+αy +βx˜ +ε (4) t t−1 t t where x˜ = (cid:80)3 1x is the monthly indicator aggregated at the quarterly fret j=1 3 t−j quency by a simple average. In this paper, we estimated equation (4) by OLS, and when we have a missing observations in x we filled it by using an AR model. Furt thermore, the predictions from the bridge model are obtained by first estimating a model for each monthly indicator in the database (except for the PMI series), and then by averaging the prediction.14 Table 3 shows the Root Mean Squared Error (RMSE) at the end of each month of the DFM, an AR(2) model, a Random Walk, and the Bridge model. The Table is 11 Let X be a non-stationary variable in log-levels, and let xy =X −X be the y-o-y growth t t t t−4 rates and xq =X −X be the q-o-q growth rates, so that xy =xq+xq +xq +xq . Then, t t t−1 t t t−1 t−2 t−3 if thetrue model is xq =λf +e , we have xy =λ(1+L+L2+L3)f +(1+L+L2+L3)e , which t t t t t t can be rewritten as xy =λF +(1+L+L2+L3)e where F is a 4×1 singular vector. t t t t 12 Thefirsteigenvalueaccountfor70%ofthetotalvariance,thesecondfor20%,thethirdfor5%, and the fourth for 3%. In the appendix we show robustness results when the model is estimated by setting either r =1 or r =4. 13 AspointedoutbyBaffigietal.(2004),differentlyfromDFMs,bridgemodelsarenotconcerned with particular assumption underlying the DGP of the data, but rather, the inclusion of specific explanatory indicators is based on the simple statistical fact that they embody timely updated information about the target GDP growth series. 14 PMIdevelopingcountrieswasexcludedbecausetherearetoofewobservationsforthisindicator and the prediction is volatile. 12

divided in three parts: the first part, labelled as “Forecast”, reports the RMSE of the prediction of the next quarter; the second, labelled as “Nowcast”, reports the RMSE of the prediction of the current quarter; finally, the last section, labelled as “Backcast”, reports the MSE of the prediction of the previous quarter. As we can see from the fact that the RMSE in Table 3 are decreasing with each month,theDFMisabletocorrectlyreviseitsGDPpredictionasmoredatabecomes available. Furthermore, compared to the univariate benchmarks the RMSE of the DFM is consistently lower, up to a maximum reduction at the end of the first month afterthereferencemonthof39%comparedtotheAutoregressivemodel, andof15% compared to the Bridge model. This is an important finding since it tells us that there is valuable additional information in the Indonesian high frequency data that can be used to predict GDP growth. Table 3 Root Mean Squared Error: End of month Month DFM AR RW Bridge Forecast 1 0.595 0.703 0.847 0.627 2 0.525 0.619 0.661 0.567 3 0.467 0.619 0.661 0.536 Nowcast 1 0.441 0.609 0.666 0.494 2 0.342 0.455 0.430 0.391 3 0.298 0.455 0.430 0.361 Backcast 1 0.279 0.456 0.459 0.331 Notes: This table reports Root Mean Squared Error (RMSE) at the end of each month for the Dynamic Factor model (DFM), anAR(2)model,aRandomWalk(RW),andtheBridgemodel. Theupperpanellabelledas“Forecast”,reportstheRMSEofthe predictionofthenextquarter;themidpanel,labelledas“Nowcast”,reportstheRMSEofthepredictionofthecurrentquarter;the bottompanellabelledas“Backcast”,reportstheMSEofthepredictionofthepreviousquarter. In Table 4 we investigate which data release carries more information for the prediction with the DFM. Identifying such variables is particularly important since it can help policymakers understand what series to track while monitoring the Indonesian economy. More precisely, Table 4 shows the RMSE associated with each data release. We can see that some variables are particularly relevant for correctly updating the prediction of y-o-y GDP growth. Among these, the most important one is Exports, which accounts for the largest reduction in RMSE when the data is released. Other relevant ones are GDP of the previous quarter, Imports and Cement. Notice also that upon the release of some variables the “average” forecasting performance reported in Table 4 appears to deteriorate. Thus it would be tempting to drop these variables in order to “improve” the overall forecasting performance. However, each of them may also improve the estimation of the model by exploiting the commonality in the data, and hence make our forecasts more robust to one-off changes in a particular variable. We conclude this section with a caveat that will apply to most of the empirical exercise. As we argued in Section 2 we had to use data only from 2002 onwards, and this has limited us in two ways. First, as discussed in Section 3, we restrict the number of variables to include in the model. Second, since we are able to produce predictions for just 28 quarters, we are averaging over only 28 prediction errors to produce the RMSE. 13

Table 4 Root Mean Squared Error: Data Flow Day Release Forecast Nowcast Backcast Month 1 1 Policy rate 0.624 0.485 0.315 3 Imports∗ 0.606 0.455 0.297 4 PMI 0.612 0.460 0.300 10 Cement 0.611 0.456 0.296 15 Exports 0.595 0.442 0.280 16 Car sales 0.594 0.440 0.277 28 M2 0.595 0.441 0.279 Month 2 1 Policy rate 0.594 0.443 0.278 3 Imports∗ 0.592 0.458 0.285 4 PMI 0.598 0.467 0.285 5 GDP 0.589 0.436 7 BTI∗∗ 0.588 0.437 10 Cement 0.582 0.426 15 Exports 0.524 0.348 16 Car sales 0.525 0.342 28 M2 0.525 0.342 Month 3 1 Policy rate 0.518 0.344 3 Imports∗ 0.509 0.327 4 PMI 0.514 0.337 10 Cement 0.509 0.333 15 Exports 0.459 0.298 16 Car sales 0.464 0.295 28 M2 0.467 0.298 Notes: ThistablereportsRootMeanSquaredError(RMSE)incorrespondenceofeachdatareleases. Column“Forecast”,reports the RMSE of the prediction of the next quarter; column “Nowcast”, reports the RMSE of the prediction of the current quarter; column“Backcast”,reportstheMSEofthepredictionofthepreviousquarter. ∗Inthisday3differentseriesarereleased: Imports: ConsumptionGoods,Imports: CapitalGoods,andImports: Rawmaterialsbutresultsaregroupedinonevariable. ∗∗ BTIstands forBusinessTendencyIndex. 4.3 Comparison against Market Benchmark Another way to evaluate the forecasting performance of our model, is to compare the prediction obtained with the DFM with the prediction of market operators. In Figure 5 we compare our predictions with those of the Bloomberg Survey (BS). The BS consists of the median GDP prediction provided independently by a number of specialistsafewdaysbeforeGDPisreleased.15 Therefore, sincetheBloombergSurvey is released few days before previous quarter GDP, according to our terminology the BS prediction is a backcast, and we will compare it with our last prediction before GDP is released. When comparing our prediction with Bloomberg we have to be careful with respect todatarevisions. ItiswellknownthatGDPdataarerevisedoften,andasalludedin Section 4.1 we are not able to track data revisions. The literature on factor models has shown that these models are robust to data revisions (Giannone et al., 2008) as revision errors are by nature idiosyncratic and do not affect factor estimation. However the case of Indonesia appears rather exceptional since the statistical office has recently revised the GDP series substantially. The black line of the left plot in Figure 5 represents the last vintage available of the GDP series, while the black 15 Notice that both the number of specialists that provide the prediction as well as the survey release day vary from quarter to quarter. 14

line in the right plot of Figure 5 represents the first release of GDP growth by the statistical office. Clearly, these two series are quite different, and in particular from 2011 onwards the statistical office has systematically revised down its GDP growth estimates. This fact is crucial when we attempt to make comparisons with truly historical forecasts. In particular, the Bloomberg survey is targeting the first release of year-on-year quarterly GDP growth (right plot), while the DFM we have estimated is designed to target the final release (left plot). As we can see from Figure 5, the Bloomberg Survey is tracking the first release well, with a RMSE of 0.249. Similarly, we can also see that our prediction is good, and indeed our RMSE is 0.286 which is just 15% worse than that of the Bloomberg Survey. What is more striking though, is the magnitude of the revision error of the statistical office. Indeed, if we think of the first release as an estimate of the final release, we can than compute a RMSE which in this case is 0.334, higher than what we obtain with the DFM (albeit using only final figures). This result suggests that the process of GDP revisions is not entirely random, as it can be “improved upon” by exploiting the information available in our relatively small dataset. More generally, it suggests the difficulty in interpreting the reliability of official national accounts estimates. Figure 5 Prediction of Previous Quarter GDP Last Available Vintage and Dynamic Factor Model Prediction First GDP Figure Released and Bloomberg Survey 6.5 6.5 6 6 5.5 5.5 5 5 4.5 4.5 4 4 Dec07 Sep08 Jun09 Mar10 Dec10 Sep11 Jun12 Mar13 Dec13 Sep14 Dec07 Sep08 Jun09 Mar10 Dec10 Sep11 Jun12 Mar13 Dec13 Sep14 Notes: TheblacklineoftheleftplotrepresentsthelastvintageavailableoftheGDPseries,whilethegreylineistheprediction obtainedwiththeDFMthelastdaybeforeGDPisreleased. Theblacklineintherightplotrepresentsthefirstfigurereleasedby thestatisticalofficeofGDPgrowth,whilethegreylineisthemedianpredictionfromtheBloombergSurvey. 4.4 Comparison against Institutional benchmarks Finally, by following Luciani and Ricci (2014) we can construct predictions for the current annual growth rate and compare them with those published by policy institutions.16 In details, we compare our predictions with those published by the 16 LetXy =100×log(GDPy)beGDPoftheq-thquarterofyeary,andletZy =100×log(GDPy) q q be GDP of year y. Then, by definition xy = Xy −Xy−1 is the y-o-y growth rate, while zy = q q q Zy −Zy−1 is the annual growth rate. Following Mariano and Murasawa (2003), we make use of theapproximation Zy ≈(Xy+Xy+Xy+Xy)/4, whichallow usto writetheannual growthrate 1 2 3 4 asafunctionofy-o-ygrowthrates: zy =Zy−Zy−1 ≈(Xy+Xy+Xy+Xy)/4−(Xy−1+Xy−1+ 1 2 3 4 1 2 Xy−1+Xy−1)/4=(xy+xy+xy+xy)/4. 3 4 4 3 2 1 15

Asian Development Bank (ADB) in the Asian Development Outlook, those published by the International Monetary Fund (IMF) in the World Economic Outlook, and the prediction by Consensus Forecast (CF).17 The ADB publishes its prediction of current annual GDP growth twice a year, approximately in April and in late September; also the IMF publishes twice a year its prediction but these are released on April and October. Predictions by CF are available each month. The north west (NW) panel in Figure 6 shows annual GDP growth together with the prediction of current annual GDP growth obtained at the end of each month with the DFM. The other panels of Figure 6 show predictions from ADB, IMF, and CF together with annual GDP growth. Note however, that the annual GDP growth reported in the NW panel is different from that reported in the other panels. In the NW panel we are reporting annual growth computed on the basis of the last vintage of available data (reconstructed series using base year 2010 as described in Section 2), while the other panels report annual growth computed on the basis of the last vintage of the old GDP series (2000 basis). We do the latter because the ADB, the IMF, and CF predicted the series with the old base in real-time and not the new base. Figure 6 Prediction of Annual GDP Growth Rate Dynamic Factor Model Consensus Forecast 7 6.5 6 6.5 5.5 6 5 5.5 4.5 4 5 3.5 4.5 3 4 2.5 2008 2009 2010 2011 2012 2013 2014 2015 2008 2009 2010 2011 2012 2013 2014 2015 Asian Development Bank International Monetary Fund 7 7 6.5 6.5 6 6 5.5 5.5 5 5 4.5 4.5 4 4 2008 2009 2010 2011 2012 2013 2014 2015 2008 2009 2010 2011 2012 2013 2014 2015 Notes: Theblacklineinthenorth-westplotisannualGDPgrowth,whilethegreydiamondsarethepredictionobtainedatthe endofeachmonthwiththeDFM.InalltheotherpanelstheblacklineisannualGDPgrowthcomputedbyusingthelastvintageof theoldGDPseries(2000basis),whilethegreydiamondsarethepredictionofADB,IMF,andCF.DataforCFarefromConsensus EconomicsInc. From Figure 6 and the RMSE values in Table 5, we can see that the DFM is 17 Consensus Economics Inc. forecasts comprise quantitative predictions of private sector forecasters. Eachmonthsurveyparticipantsareaskedfortheirforecastsofarangeofmacroeconomic and financial variables for the major economies. 16

predicting annual GDP growth quite well, and in particular it correctly revises its prediction as more data becomes available during the calendar year. Furthermore, the prediction of the DFM is comparable to that of CF, and slightly superior to that of the ADB and the IMF. Of course here two caveats apply: the first is that using annual data we have only 7 observations, and the second is that our exercise is pseudo real-time, and therefore we have a better information set than the one available in real-time to forecasters. Table 5 Root Mean Squared Error: Annual GDP Growth Month DFM CF ADB IMF January 0.484 0.408 February 0.352 0.408 March 0.295 0.577 April 0.248 0.722 0.603 0.824 May 0.195 0.655 June 0.152 0.431 July 0.167 0.348 August 0.080 0.237 September 0.124 0.158 0.262 October 0.128 0.080 0.306 November 0.118 0.097 December 0.093 0.096 Notes: ThistablereportsRootMeanSquaredError(RMSE)ofthepredictionofannualGDPgrowthattheendofeachmonth. The RMSE of the DFM is computed with reference to the last vintage available for the GDP series, while the RMSE for ADB, CF,andIMF,iscomputedwithreferencetothelastvintageoftheoldGDPseries(2000basis). DataforCFarefromConsensus EconomicsInc. 5 Conclusions In this paper we have applied state of the art techniques for nowcasting Indonesia’s GDP growth. Our approach is based on a Dynamic Factor model, to efficiently exploit monthly and quarterly variables and to properly account for the sequence of macroeconomic data releases. We find that relying on market “revealed preferences” for certain indicators on the Indonesian economy is an effective guide to choosing what variables to include in our information set. To this end we have relied on the Bloomberg platform, which tracks the relevance of each series for its subscribers, and also on our “expertise judgment” by including a few indicators that we think provide extra information on the Indonesian economy. Based on this, despite using a relatively narrow set of variables, when focusing on the year-on-year growth rate, the Dynamic Factor model nowcast error falls by 35% compared to the benchmark AR, and by almost 40% for the backcast. Lacking a full time series of GDP revisions as well as for the information set used in the Dynamic Factor model we cannot assess how well our model predictions perform compared to those of experts’ forecast surveyed by Bloomberg. Still, our “pseudo-real-time” forecasting performance is comparable to the one achieved in a truly “real-time” setting by the median Bloomberg survey. 17

Furthermore, since our model can be used to forecast further ahead, we compute also calendar year annual growth rates. This exercise allows us to compare the tracking of our Dynamic Factor model forecasts on a smoother and medium-term indicator of Indonesia’s growth, arguably more important for policy decisions than the(potentiallyerratic)quarterlygrowthrate. Inthiscaseourmodelcompareswell withtheforecastsproducedbytheaverageofprivatesectorexpectations(Consensus Forecasts) as well as by the IMF-World Economic Outlook. Finally, ourexplorationintoIndonesia’sdatashedslightontoalackofvaluablehigh frequency statistics on economic growth, as well as on deficiencies in the statistical framework underlying national accounts. References Aastveit, K. A. and T. Trovik (2012). Nowcasting norwegian GDP: the role of asset prices in a small open economy. Empirical Economics 42(1), 95–119. Angelini, E., G. Camba-Mendez, D. Giannone, L. Reichlin, and G. Ru¨nstler (2011). Short-term forecasts of euro area gdp growth. The Econometrics Journal 14(1), C25–C44. Baffigi, A., R. Golinelli, and G. Parigi (2004). Bridge models to forecast the euro area GDP. International Journal of Forecasting 20(3), 447–460. Bai, J. (2003). Inferential theory for factor models of large dimensions. Econometrica 71, 135–171. Bai, J.andS.Ng(2008). Forecastingeconomictimeseriesusingtargetedpredictors. Journal of Econometrics 146(2), 304–317. Ban´bura, M., D. Giannone, M. Modugno, and L. Reichlin (2013). Now-casting and the real-time data-flow. In G. Elliott and A. Timmermann (Eds.), Handbook of Economic Forecasting, Volume 2, pp. 195–237. Amsterdam: Elsevier-North Holland. Ban´bura, M., D. Giannone, and L. Reichlin (2011). Nowcasting. In M. P. Clements and D. F. Hendry (Eds.), Oxford Handbook on Economic Forecasting. New York: Oxford University Press. Ban´bura, M. and M. Modugno (2014). Maximum likelihood estimation of factor models on data sets with arbitrary pattern of missing data. Journal of Applied Econometrics 29(1), 133–160. Ban´bura, M. and G. Ru¨nstler (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., O. Darn´e, and L. Ferrara (2010). Are disaggregate data useful for factor analysis in forecasting French GDP? Journal of Forecasting 29(1-2), 132– 144. 18

Bragoli, D., L. Metelli, and M. Modugno (2014). The importance of updating: Evidence from a brazilian nowcasting model. Finance and Economics Discussion Series 2014-94, Federal Reserve Board. Camacho, M. and G. Perez-Quiros (2010). Introducing the euro-sting: Short-term indicator of euro area growth. Journal of Applied Econometrics 25(4), 663–694. D’Agostino, A., K. McQuinn, and D. O’Brien (2012). Now-casting Irish GDP. OECD Journal: Journal of Business Cycle Measurement and Analysis 2012(2), 21–31. De Mol, C., D. Giannone, and L. Reichlin (2008). Forecasting using a large number of predictors: Is bayesian shrinkage a valid alternative to principal components? Journal of Econometrics 146, 318–328. Doz, C., D. Giannone, and L. Reichlin (2011). A two-step estimator for large approximate dynamic factor models based on kalman filtering. Journal of Econometrics 164(1), 188–205. Doz, C., D. Giannone, and L. Reichlin (2012). A quasi maximum likelihood approach for large approximate dynamic factor models. Review of Economics and Statistics 94(4), 1014–1024. Forni, M., D. Giannone, M. Lippi, and L. Reichlin (2009). Opening the Black Box: Structural Factor Models versus Structural VARs. Econometric Theory 25(5), 1319–1347. Forni, M., M. Hallin, M. Lippi, and L. Reichlin (2003). Do financial variables help forecasting inflation and real activity in the Euro Area? Journal of Monetary Economics 50(6), 1243–1255. Fulop, G. and G. Gyomai (2012). Transition of the OECD CLI system to a GDPbased business cycle target. Giannone, D., S. Miranda Agrippino, M. Modugno, and L. Reichlin (2014). Nowcasting China. mimeo. Giannone, D., L. Reichlin, and D. Small (2008). Nowcasting: The real-time informational content of macroeconomic data. Journal of Monetary Economics 55(4), 665–676. Kasri, R. and S. H. Kassim (2009). Empirical determinants of saving in the Islamic banks: Evidence from Indonesia. JKAU: Islamic Econ. 22(2), 181–201. Kubo, A. (2009). Monetary targeting and inflation: Evidence from Indonesia’s post-crisis experience. Economics Bulletin 29(3), 1805–1813. Liu,P.,T.Matheson,andR.Romeu(2012). Real-timeforecastsofeconomicactivity for Latin American economies. Economic Modelling 29(4), 1090–1098. Luciani, M. (2014a). Forecasting with approximate dynamic factor models: the role of non-pervasive shocks. International Journal of Forecasting 30(1), 20–29. 19

Luciani, M. (2014b). Large-dimensional dynamic factor models in real-time: A survey. Universit´e libre de Bruxelles. Luciani, M. and L. Ricci (2014). Nowcasting Norway. International Journal of Central Banking 10, 215–248. Ma´ckowiak, B. (2007). External shocks, US monetary policy and macroeconomic fluctuations in emerging markets. Journal of Monetary Economics 54(8), 2512– 2520. Marcellino, M. and C. Schumacher (2010). Factor MIDAS for nowcasting and forecastingwithragged-edgedata: AmodelcomparisonforGermanGDP. Oxford Bulletin of Economics and Statistics 72(4), 518–550. Mariano, R. S. and Y. Murasawa (2003). A new coincident index of business cycles based on monthly and quarterly series. Journal of Applied Econometrics 18(4), 427–443. Matheson, T. D. (2010). An analysis of the informational content of New Zealand data releases: The importance of business opinion surveys. Economic Modelling 27(1), 304–314. Matheson, T. D. (2013). New indicators for tracking growth in real time. Journal of Business Cycle Measurement and Analysis 2013(1), 51–71. Modugno,M.(2014). Now-castinginflationusinghighfrequencydata. International Journal of Forecasting 29(4), 664–675. Parigi, G. and G. Schlitzer (1995). Quarterly forecasts of the italian business cycle by means of monthly economic indicators. Journal of Forecasting 14(2), 117–141. Raghavan, M. and M. Dungey (2015). Should ASEAN-5 monetary policy-makers act pre-emptively against stock market bubbles? Applied Economics 47(11), 1086–1105. Stock, J. H. and M. W. Watson (2002a). Forecasting using principal components from a large number of predictors. Journal of the American Statistical Association 97, 1167–1179. Stock, J. H. and M. W. Watson (2002b). Macroeconomic forecasting using diffusion indexes. Journal of Business and Economic Statistics 20(2), 147–162. 20

A Appendix A.1 Robustness In this appendix we show robustness checks with respect to the number of factors and to the composition of the dataset. As we discuss in Section 3.1 we constructed the database by selecting variables from thesetofindicatorsthatmarketanalystsaremonitoring. InTable6weshowRMSE for a DFM parameterized as described in Section 4.1 but estimated on different datasets. In particular we considered the option of using all the indicators in Table 1 (Bloomberg Selection), and the option of selecting indicators automatically with the LARS algorithm (Automatic Selection) as in Bai and Ng (2008). As we can see our database clearly delivers the best performance, though at least in forecasting the performance of the DFM estimated over the indicators followed by Bloomberg is comparable. It is particularly disappointing the performance of the DFM when the indicators are selected with LARS. We believe that this is a consequence of having too few observations, and some missing values here and there in the time series of our indicators. Then, in Section 4.1 we motivated our choice of including two factors in the model, but in doing so we explained that two possible meaningful options were to estimate a model with one factor, or a model with four factors. In Table 6 we also show the RMSE for a DFM estimated by including different number of factors, and as we can see the choice of including two factors proved to be optimal. Table 6 Root Mean Squared Error: Different model specifications and datasets Month Benchmark One Four Automatic Bloomberg Model Factor Factors Selection Selection Forecast 1 0.595 0.614 0.716 0.921 0.588 2 0.525 0.549 0.666 0.810 0.526 3 0.467 0.517 0.564 0.751 0.533 Nowcast 1 0.441 0.499 0.499 0.726 0.542 2 0.342 0.389 0.408 0.554 0.423 3 0.298 0.362 0.383 0.507 0.433 Backcast 1 0.279 0.331 0.333 0.489 0.408 Notes: ThisTablereportsRootMeanSquaredErrorsfortheDFMestimatedunderdifferentconfigurationsoroverdatabasein whichtheselectionprocessisdifferentthantheoneexplainedinSection3.1. Theupperpanellabelledas“Forecast”,reportsthe RMSEofthepredictionofthenextquarter;themidpanel,labelledas“Nowcast”,reportstheRMSEofthepredictionofthecurrent quarter;thebottompanellabelledas“Backcast”,reportstheMSEofthepredictionofthepreviousquarter. 21