Real-Time Real Economic Activity Entering the Pandemic Recession
RReal-Time Real Economic Activity:Exiting the Great Recession andEntering the Pandemic Recession
Francis X. DieboldUniversity of PennsylvaniaFirst Version: June 26, 2020This Version, June 30, 2020
Abstract : We study the real-time signals provided by the Aruoba-Diebold-Scotti Index ofBusiness conditions (ADS) for tracking economic activity at high frequency. We start withexit from the Great Recession, comparing the evolution of real-time vintage beliefs to a“final” late-vintage chronology. We then consider entry into the Pandemic Recession, againtracking the evolution of real-time vintage beliefs. ADS swings widely as its underlyingeconomic indicators swing widely, but the emerging ADS path as of this writing (late June)indicates a return to growth in May. The trajectory of the nascent recovery, however, ishighly uncertain – particularly as COVID-19 spreads in the South and West – and could berevised or eliminated as new data arrive.
Acknowledgments : For helpful discussion I thank Boragan Aruoba, Glenn Rudebusch,Chiara Scotti, Minchul Shin, Keith Sill, and Tom Stark. For outstanding research assistanceand related discussion I thank Philippe Goulet Coulombe, Tony Liu, and Boyan Zhang. Theusual disclaimer applies.
Key words : Aruboba-Dieold-Scotti index, ADS index, nowcasting, business cycle, recession,expansion, coincident indicator, real economic activity, forecasting, Big Data
JEL codes : E32, E66
Contact : [email protected] a r X i v : . [ ec on . E M ] J un Introduction
Accurate assessment of of current real economic activity (“business conditions”) is key forsuccessful decision making in business, finance, and policy. It is difficult, however, to trackbusiness conditions in real time, both because no single observed economic indicator is “business conditions”, and because different indicators are available at different observa-tional frequencies, and with different release delays. Nevertheless there exists the tantalizingpossibility of accurate real-time business conditions assessment (“nowcasting”), and recentdecades have witnessed great interest in nowcasting methods and applications (e.g., BanBuraet al. (2011)).The workhorse nowcasting approaches involve dynamic factor models, which relate a setof observed real activity indicators to a single underlying latent real activity factor. Both“small data” approaches (e.g., based on 5 indicators) and “Big Data” approaches (e.g.,based on 500 indicators) are available. Small data approaches came first, and they typicallyinvolve maximum likelihood estimation (e.g., Stock and Watson (1989)). Subsequent BigData approaches, in contrast, typically involve two-step estimation based on a first-stepextraction of principal components (e.g., Stock and Watson (2002), McCracken and Ng(2016)).Both introspection and experience reveal that Big Data nowcasting approaches are notnecessarily better. First, they are more tedious to manage, and less transparent. Second,they may not deliver much improvement in factor extraction accuracy, which increases andstabilizes quickly as the number of indicators increases (Doz et al., 2012). Third, casualinclusion of many indicators can be problematic because a poorly-balanced set of indicatorscan create distortions in the extracted factor (Boivin and Ng, 2006), whereas small dataapproaches promote and facilitate hard thinking about a well-balanced set of indicators (Baiand Ng (2008)).Against this background, in this paper we assess the performance of a leading small-datanowcast, the Aruoba-Diebold-Scotti (ADS) Index of Business Conditions (Aruoba et al.,2009). ADS is designed to track real business conditions at high frequency, and it has beenmaintained and released in real time by the Federal Reserve Bank of Philadelphia contin-uously since 2008. Its modeling style and underlying economic indicators build on classic The production version used by FRB Philadelphia differs in some ways (e.g., included indicators andtreatment of trend) from the prototypes provided by Aruoba et al. (2009) and Aruoba and Diebold (2010),which themselves differ slightly. All discussion in this paper refers to the FRB Philadelphia version. Allmaterials, including the full set of vintage nowcasts, are available at . arly work in the tradition of Burns and Mitchell (1946), Sargent and Sims (1977), and Stockand Watson (1989). The underlying indicators span high- and low-frequency information onreal economic flows: weekly initial jobless claims; monthly payroll employment growth, in-dustrial production growth, personal income less transfer payments growth, manufacturingand trade sales growth; and quarterly real GDP growth.Crucially, we assess ADS performance using only information actually available in realtime. This is required for truly credible real-time evaluation, and it can only be achievedby using nowcasts produced and permanently recorded in real time, which is very differentfrom simply removing final-revised data and inserting vintage data into an otherwise expost analysis. Unfortunately, such evaluations are rare, because there simply are not manyinstances of long series of nowcasts produced and recorded in real time. ADS, however, hasbeen produced and recorded in real time roughly twice weekly since late 2008, so we canprovide real-time performance assessments both exiting the Great Recession and enteringthe Pandemic Recession.We proceed as follows. In section 2 we provide background on aspects of ADS construc-tion, updating, ex post characteristics, and performance evaluation. In section 3 we evaluateADS performance exiting the Great Recession. In section 4 we evaluate ADS performanceentering the Pandemic Recession. We conclude in section 5. Here we provide background on index construction (section 2.1), ex post historical char-acteristics (section 2.2), and general issues of relevance to assessing ex ante nowcastingperformance (section 2.3).
ADS is a dynamic factor model with multiple indicators and a single latent real activityfactor. The ADS index is an estimate of that latent real activity factor. Importantly,the model is specified such that the real activity factor tracks the de-meaned growth rate of real activity. Hence zero ADS does not indicate zero growth; rather, zero ADS indicates“normal” growth. Progressively more negative or positive values indicate progressively worse-or better-than-average real growth, respectively.ADS is specified at daily frequency, allowing as necessary for missing data for the less-2igure 1: ADS Index: Ex Post Path 03/01/1960 - 12/31/2013 (Vintage 6/26/2020)Notes: The shaded regions are NBER-designated recessions.frequently observed variables. Importantly, despite complications from missing data, time-varying system matrices, aggregation across frequencies, etc., the Kalman filter and associ-ated Gaussian pseudo likelihood evaluation via prediction-error decomposition remain valid,subject to some well-known modifications. Model estimation is therefore straightforward,after which the Kalman smoother produces an optimal extraction of the underlying realactivity factor. That is, the Kalman smoother produces the ADS index: The extractedsequence at any time t ∗ is the vintage- t ∗ ADS sequence, { ADS , ADS , ...ADS t ∗ } .The first ADS vintage was released 12/5/2008, covering 3/1/1960 through 11/30/2008.Since then, ADS has been continuously updated whenever new data are released. TheKalman smoother is re-run, generally within two hours of the release, and the newly-extracted index is re-written to the web from 3/1/1960 to “the present”. ADS has beenupdated approximately eight times per month on average since inception.3 .2 Ex Post Characteristics In Figure 1 we show the ADS index from 03/01/1960 through 12/31/2006, as assessed in the6/26/2020 vintage. The sample range is well before the vintage pull date, so the chronologydisplayed is (intentionally) ex post. We do this because it is instructive to examine the expost chronology before passing to real time assessment, which can only be done after ADSwent live in late 2008. Several features are noteworthy. For example, the ADS chronology coheres strongly withthe NBER chronology, plunging during NBER recessions. In addition, several often-discussedfeatures of the business-cycle are evident in ADS, such as the pronounced moderation involatility during the Greenspan era.The ADS value added relative to the NBER chronology stems from the facts that (1) itis a cardinal measure, allowing one to assess not only recession durations, but also depthsand patterns (see Table 1), and (2) its updates arrive in timely fashion, whereas NBERrecession start and end dates are typically not announced until long after the fact (againsee Table 1). Of course, if ADS is to be a useful guide for business and policy decisions, itsfrequently-arriving updates must provide reliable signals in real time, not just ex post as inFigure 1. We now turn to that issue.
Truly credible nowcasting performance assessment requires using vintage information , whichemerges as the limit of a sequence of progressively more realistic and credible nowcast/forecastevaluation approaches:
1. Use full-sample estimation, and use final revised data2. Use expanding-sample estimation, and use final revised data3. Use expanding-sample estimation, and use vintage data (“Pseudo Real Time”) The model must be specified at daily frequency, despite the fact that the highest-frequency indicatoris is weekly initial jobless claims, to account for the varying number of days/weeks per month, which alsoproduces time-varying system parameter matrices. See, for example, Durbin and Koopman (2001) on missing data, and Harvey (1991) on aggregation offlow variables. The sample period intentionally excludes the Pandemic Recession, which we will subsequently examinein detail. Note that nowcasts are effectively just h -step-ahead forecasts with horizon h =0. Notes: Recession dates and durations in months are from the NBER chronology; see . When available, the announcement dates appear in parentheses. The NBER troughmonth for the Pandemic Recession has not yet been announced. Recession depth is the minimum absolutedaily ADS value during the recession; more precisely, the depth D of recession R is D = | min i ( ADS i ) | , i ∈ R , where i denotes days. Recession severity S is the product of depth and duration. Both D and S usea late-vintage ADS chronology and the NBER recession chronology.
4. Use expanding-sample estimation, and use vintage information (“Real Time”).Approaches 1 and 2 are clearly unsatisfactory: Approach 1 uses time periods and data valuesnot available in real time, and approach 2 is an improvement but still uses data not availablein real time. Approach 3, involving vintage data , is typically viewed as the gold standard.It is implemented comparatively infrequently, however, due to the tedium involved and thefact that vintage data are often unavailable. Approach 4, use of vintage information , limitsthe information set to that available and actually used in real time, which is more restrictivethan merely limiting the data to that available in real time. It is, however, almost neverimplemented.To appreciate why fully-credible assessment requires vintage information rather than justvintage data, consider the following:1. Econometric/statistical theory and experience evolve, prompting changes to the esti-mation procedure; the frequency and timing of re-estimation and its interaction with The two key sources of U.S. vintage data are the Real-Time Dataset for Macroeconomists atthe Federal Reserve Bank of Philadelphia ( ), and ALFRED at the Federal Reserve Bank of St. Louis ( https://alfred.stlouisfed.org/ ). We refer to a real-time ADS extraction as a path, and a graph of a sequence of paths asa path plot. By following the path plot rightward, moving through time, we track theevolution of ADS beliefs about the chronology of business conditions. We cannot examinereal-time ADS performance when entering the Great Recession, because ADS did not startuntil December 2008, well after the great recession began. But we can examine real-timeADS performance when exiting the great recession. Sometimes we call path plots “tentacle plots”, because the paths resemble the tentacles of a jellyfish. At the time it was easy toread the cards as saying that the recession was ending, and ADS was a bit too optimistic,moving upward toward recovery.Now consider the remaining panels of Figure 2. In the second panel we show the next,and contrasting, 3/6/2008 ADS path. In the interim ADS has quickly learned the situation,the double dip in particular, and is very much on track, capturing the second dip in January2009. ADS continues to climb steadily through the third and fourth panels (6/5/2009 and9/3/3009, respectively), and by the time of the bottom panel (12/4/2009) it is clear that theGreat Recession ended in June or July, with ADS basically fluctuating around 0 after that.(Recall that ADS=0 means average growth, not zero growth.)All told, the five quarterly real-time ADS paths generally match the ex post path closely,and they correctly identify the recession’s end, well before the end of 2009 and indeed roughly1.5 years before the official NBER announcement in September 2010.To emphasize ADS timeliness, we plot the late vintage ADS in Figure 2 all the waythrough 2010, which allows inclusion of the NBER’s end-of-recession announcement on9/20/2010, long after the fact and not helpful for real-time decision making. ADS fillsthe gap left by the late-arriving NBER chronology, and it also provides a numerical measurethat allows one to track the recession’s pattern, depth, overall severity, etc., in addition to In particular, according to the Federal Reserves G.17 Industrial Production (IP) release of October 16,2008, September IP was severely affected by a highlyunusual and largely exogenous triple shock (HurricanesGustav and Ike, and a strike at a major aircraft manufacturer), which caused an annualized September IPdrop of nearly fifty percent. A similar pattern exists for Manufacturing and Trade Sales (MTS). IP and MTSalso rebounded unusually sharply in October indeed IP appears to overshoot presumably in an attempt bymanufacturers to make up for Septembers loss. Of course the NBER is not seeking to be helpful for real-time decision making; rather, they seek tometiculously construct the U.S. business cycle chronology of record, quite reasonably using all relevantinformation – even very late-arriving information. The black-dot sequence ofreal-time smooths is naturally less variable than the later-vintage (December 2010) smoothshown in red, because the latter has more information on which to condition, and thereforecaptures more variation. They are also filtered, because smoothed and filtered values coincide for the last observation in a sample.
We focus in this section on the Pandemic Recession that started in March 2020. It isinstructive to begin by comparing it to the Great Recession of 2007-2009. To that endwe show the ADS path in Figure 5, from late 2007 through the date of this writing. Theso-called “Great Recession” appears minor by comparison.
Figure 5 reveals the jaw-dropping ADS drop in the Pandemic Recession, more than five timesthat of any other recession since 1960. The ADS drop is entirely appropriate, due to similarlyjaw-dropping and historically unprecedented movements in its underlying indicators. We can interpret ADS depth by calibrating the sum of the worst two quarters of GDP See Appendices A and B for chronologies of data releases and ADS movements, respectively. (cid:92)
GDP R = − .
33 + . ADS R , (1)where GDP R is the sum of the worst two quarters of GDP growth in recession R , and ADS R is the worst ADS in recession R . The slope is quite precisely estimated ( s.e. = . R = .
61. As a quick approximation we can use (cid:92)
GDP R = . ADS R . Figure 5 shows an ADSminimum of roughly -25, corresponding (via equation (1)) to a two-quarter annualized GDPgrowth of roughly -12%, or roughly -6% for each of two quarters. As of this writing, the official trough month for the Pandemic Recession has not been an-nounced. It could be as early as May 2020, in which case the Pandemic Recession would bethe shortest in history. (A May trough turns out to be likely, if highly uncertain, conditionalon information available through June, as we will soon discuss.) The trough could of coursealso be later, perhaps with additional dips due to continued spread and potential resurgenceof COVID-19. 11igure 6: Worst Two GDP Quarters vs. Worst ADSNotes: We show the sum of the worst two quarters of GDP vs. the worst ADS value, foreach recession except the Pandemic Recession (both the GDP drop and the ADS drop areas-yet unknown for the Pandemic Recession), together with the fitted regression line.
In Figure 7 we show the latest-vintage Pandemic Recession path. The overall extractedpath is smooth and convex, with a minimum in early April, and a return to average growthby mid-May. We emphasize again, however, that ADS measures real activity growth, notlevel. Hence positive ADS does not necessarily mean “good times”; rather, it means “goodgrowth”, which may be from a very bad initial condition. That was evidently the situationin late May, as the battered U.S. economy evidently resumed growth.
In Figure 8 we show several end-of-month paths in black, starting with February 2020. Forcomparison, in each panel we also show the latest-vintage path in red. One should not thinkof the later-vintage path as “truth”, because it may be revised as new data arrive. Ideally wewould like to use a vintage from a year or two after the recession’s end, as with our use of the12igure 7: ADS Index: Ex Post Path 1/1/2020 - 6/26/2020 (Vintage 6/26/2020)December 2010 vintage as a reference path when assessing ADS during the Great Recessionin section 4, but we do not have that luxury at present. Meanwhile it is still informative tocompare the last-available path to earlier paths. Moving through the five panels of Figure 8:1. In the top panel we show the 2/28/2020 path. ADS has not moved.2. In the second panel we show the 3/27/2020 path, which looks very different. ADS hasbecome acutely aware of the disastrous situation; indeed most of the 3/27 path is wellbelow the previous all-time (post-1960) ADS low during the 1970s oil-shock recession.
3. In the third panel we show the 4/30/2020 path. The April initial claims news is bad,but less bad than March, which is good, and ADS shows a minimum in late Marchfollowed by a rise toward normalcy by the end of April.4. In the fourth panel we show the 5/29/2020 path. The May news is very bad, dominatedby the shockingly bad May 8 payroll employment number (for April), and the late-Maypath is massively down-shifted relative to the late-April path. The new minimum isin mid-April rather than late March, and the 5/29 ADS value is thoroughly dismal,nowhere near normalcy.5. In the fifth panel we show the 6/26/2020 path. Thanks to the strong May payrollemployment number (released June 5), ADS moved into normal territory, and stayed It is also apparent that the Kalman smoother may be smoothing “too much”, producing low ADS valueswell before mid-March, going back into February and even January. Its smoothing is optimal relative to thepatterns in historical data, but the March initial jobless claims movements were unprecedentedly sharp.
In Figure 9 we show the complete path plot during the Pandemic Recession through 6/26/2020,with the final path in red for comparison. In Appendix B we provide a corresponding anno-tated path chronology.There are wide real-time divergences between individual early paths and the latest-vintage red path. There are interesting patterns, however, with several real-time “metapaths” evident:1. The first extends through the 3/19/2020 ADS announcement. ADS does not move.Initial claims rise from 0.2m to 0.3m, a large move by historical standards, confirm-ing what everyone already knew: the pandemic would have important real economic15igure 10: Exiting the Pandemic Recession: Real-Time ADS Dot PlotNotes: We show the final values of all real-times 2020 ADS paths in black. For comparisonwe show the complete final-vintage path (6/26/2020) in red.consequences, but the Kalman smoother optimally but erroneously ascribes it to mea-surement error.2. The second meta-path begins with the 3/26/2020 and 4/2/2020 initial claims explo-sions. ADS plunges, but then recovers steadily despite a steady stream of bad news(it is bad, but getting less bad), almost back to 0 by the 5/7/2020 initial claims an-nouncement.3. The third meta-path begins with the horrific 5/8/2020 April payroll employment re-lease, with ADS again plunging. It then again begins mean reverting, and does socompletely when the strong May payroll employment number is released on 6/5/2020.In Figure 10 we show the corresponding dot plot, with the 6/26/2020 path again superim-posed. The dot plot is highly volatile and emphasizes the various meta-paths.16
Concluding Remarks
Our approach was part methodological and part substantive. On the methodological side,we clarified the meaning of truly honest real-time nowcast/forecast evaluation and illustratedit using the ADS Business Conditions Index, which has now been in operation over a longspan that includes emergence from the Great Recession and entry into the Pandemic Reces-sion. On the substantive side, we explored how views formed using best-practice nowcastsevolved when exiting the Great Recession, which is now a settled historical episode, andwhen entering the Pandemic Recession, which continues to swirl around us.We started with exit from the Great Recession, comparing the evolution of real-time vin-tage beliefs to a “final” late-vintage chronology. We then moved to entry into the PandemicRecession, which arrived abruptly and was caused by non-economic factors, again trackingthe evolution of real-time vintage beliefs. During March-June 2020, ADS swung widely asits underlying components swung widely, but as of the date of this writing (late June 2020)ADS indicates a very deep but also very short recession, with a business cycle trough in May2020. The trajectory of the nascent recovery, however, is highly uncertain – particularly asCOVID-19 spreads in the South and West – and could be revised or eliminated as new dataarrive.Models generally cannot be expected to perform well in out-of-sample environmentswildly different from the training environment. This is all the more so for linear mod-els like ADS, which in reality are just locally-linear approximations. From this vantagepoint the credible out-of-sample performance of ADS is rather remarkable. Nevertheless,nonlinear/non-Gaussian ADS extensions could perhaps be usefully entertained, for exampleusing the framework of Gunsilius and Schennach (2019).17 ppendices
A Pandemic Recession Entry: Data Releases
B Pandemic Recession Entry: Annotated Chronology
March 17 (February industrial production). The day of the last ADS update before theMarch 19 initial claims release. ADS continued its more-or-less random vibration aroundzero, sending the same signal that it had sent since the end of the Great Recession in 2009:the economy is growing normally. ADS=0.1.March 19 (initial jobless claims). IJC took a large move upward, confirming what everyonealready knew: the pandemic would have important real economic consequences. The Kalmansmoother optimally but erroneously ascribed this first-time IJC jump almost entirely tomeasurement error, and ADS basically did not move. ADS=-0.2.March 26 (initial jobless claims). IJC spiked in jaw-dropping off-the-chart fashion. Twohuge IJC moves in a row are not optimally ascribed to measurement error by the Kalmansmoother; rather, they are naturally ascribed to the underlying serially-correlated real ac-tivity factor – and ADS drops to approximately -15 in similarly (and literally) off-the-chartfashion. By way of comparison, the all-time ADS lows since 1960 were in the recessions of1973-1975 and 2007-2009, in both cases between -4 and -5. Note that the ADS path nowbegins its drop earlier in the year, a result of the serial correlation in IJC interacting withthe Kalman smoother. It is interesting to speculate as to whether real activity really was lower in February (say), due for example to the virus-induced January-February collapse ofa major trading partner (China). ADS=-14.5.March 26 (GDP Q4 release 3). Irrelevant. ADS=-14.5.March 27 (February personal income, January real manufacturing and trade sales). Irrele-vant. ADS=-14.3.April 2 (Initial jobless claims). IJC doubles off-the-charts, and ADS similarly doubles (down-ward) to -31. The Kalman smoother now has ADS beginning its decline in early January,again presumably an artifact of the serial correlation in IJC interacting with the Kalmansmoother. Or, again, perhaps it’s real. ADS=-31.0.19pril 3 (March payroll employment). PE drops but ADS rises . ADS evidently views thePE drop as good news, because it’s not such a big drop compared to the off-the-chartsADS=-21.2.April 9 (Initial jobless claims). Another massive IJC increase, but ADS largely unchanged.ADS=-20.6.April 15 (March industrial production). IP plunges, but it’s for the last month, and ADSactually continues its gradual upward mean reversion as initial claims improve. ADS=-17.1.April 16 (Initial jobless claims). IJC drops some, and ADS improves. ADS=-11.1.April 23 (Initial jobless claims). IJC and ADS again improve. ADS=-7.2.April 29 (Q3 GDP, first release). -4.8 % annualized. 2008Q4 was worse (-7.5 %) but the2020Q1 number was driven only by (part of) March. Had January, February, and earlyMarch 2020 been as bad as late March, 2020Q1 GDP growth would have been much worse.ADS is essentially unchanged. In the absence of the GDP news, ADS would presumablyhave risen, but the bad GDP news provided an offset. ADS=-6.9.April 30 (Initial jobless claims; March real personal income less transfer payments; Februaryreal manufacturing and trade sales). February RMTS is irrelevant. IJC continues slowlyimproving. PILTP for the previous month down sharply. The news is all bad, yet not so badas it was, and ADS improves, now it approximately equals its worst value during the GreatRecession. ADS=-3.9.May 7 (Initial jobless claims). IJC again improving. The IJC numbers continue to bebad, but they are getting less bad, and ADS seems driven by that. By this time the pathplot makes clear that new data are causing sizable revisions in entire paths. For example,the huge ADS trough was estimated to be approximately -32 in the 4/2 vintage, but itwas progressively moved upward in subsequent vintages, and now in the 5/7 vintage it isapproximately -12. ADS=-0.6.May 8 (Payroll employment). Plunges downward, and ADS plunges similarly to an all-timelow. ADS=-36.2.May 14 (Initial jobless claims) IJC almost unchanged; ADS improves slightly to ADS=-32.6.20ay 15 (April industrial production). Plunges but ADS nevertheless ADS improves. ADS=-26.9.May 21 (Initial jobless claims) IJC improves, and ADS improves to ADS=-22.9.May 28 (Initial jobless claims; GDP Q1 second release) IJC improves slightly, Q1 GDPrevised down slightly. ADS improves to ADS=-19.7.May 29 (Real manufacturing and trade sales, March; Real personal income less transferpayments, April) Both down sharply. ADS=-19.5.June 4 (Initial jobless claims) IJC continues its ever-so-slow reversion to normalcy. ADS=-16.8.June 5 (May payroll employment). PE increases sharply. Finally some surprising good news.ADS improves and goes positive. ADS=2.6.June 11 (Initial jobless claims). IJC improves slightly. ADS=3.6.June 16 (May Industrial production). IP up, ADS unchanged. ADS=3.6.June 18 (Initial jobless claims). IJC basically unchanged. ADS=3.1.June 25 (Initial jobless claims). IJC basically unchanged. ADS=2.8.June 16 (Real manufacturing and Trade Sales). MTS *****. ADS=*****.21 eferences
Aruoba, S.B. and F.X. Diebold (2010), “Real-Time Macroeconomic Monitoring: Real Ac-tivity, Inflation, and Interactions,”
American Economic Review , 100, 20–24.Aruoba, S.B., F.X. Diebold, and C. Scotti (2009), “Real-time Measurement of BusinessConditions,”
Journal of Business and Economic Statistics , 27, 417–427.Bai, J. and S. Ng (2008), “Forecasting Economic Time Series Using Targeted Predictors,”
Journal of Econometrics , 146, 304–317.BanBura, M., D. Giannone, and L. Reichlin (2011), “Nowcasting,” In M. Clemens and D.Hendry, eds.,
Oxford Handbook of Economic Forecasting , 193-224.Boivin, J. and S. Ng (2006), “Are More Data Always Better for Factor Analysis?”
Journalof Econometrics , 132, 169–194.Burns, A. and W.C. Mitchell (1946),
Measuring Business Cycles , Columbia University Pressfor NBER.Diebold, F.X. and G.D. Rudebusch (1991), “Forecasting Output with the Composite LeadingIndex: A Real-Time Analysis,”
Journal of the American Statistical Association , 86, 603–610.Doz, C., D. Giannone, and L. Reichlin (2012), “A QuasiMaximum Likelihood Approach forLarge, Approximate Dynamic Factor Models,”
Review of Economics and Statistics , 94,1014–1024.Durbin, J. and S.J. Koopman (2001),
Time Series Analysis by State Space Methods , Oxford:Oxford University Press.Gunsilius, F. and S. Schennach (2019), “Independent Nonlinear Component Analysis,”Manuscript, Departments of Economics, MIT and Brown Univeristy.Harvey, A.C. (1991),
Forecasting, Structural Time Series Models and the Kalman Filter ,Cambridge: Cambridge University Press.McCracken, M. and S. Ng (2016), “FRED-MD: A Monthly Database for MacroeconomicResearch,”
Journal of Business and Economic Statistics , 36, 574–589.22argent, T.J. and C.A. Sims (1977), “Business Cycle Modeling Without Pretending to Havetoo Much a Priori Theory,” In C.A. Sims (ed.),
New Methods in Business Cycle Research:Proceedings from a Conference, Federal Reserve Bank of Minneapolis, 45-109 .Stock, J. H. and M.W. Watson (1989), “New Indexes of Coincident and Leading EconomicIndicators,” In O. J. Blanchard and S. Fischer, eds.,
NBER Macroeconomics Annual , MITPress, 351-393.Stock, J. H. and M.W. Watson (2002), “Forecasting Using Principal Components from aLarge Number of Predictors,”