The Economics of Recession: A Survey
Part 8/10, November 2022 [1]
Arturo Estrella

8. Identifying recessions in real time

          The term “real time” has to be loosely interpreted when applied to the identification of recessions. For one, it takes a lot of time and effort to collect aggregate output data. For example, U.S. GDP, which is probably available as promptly as any in the world, is reported by the government in dribs and drabs, first as advance estimates, then as revised estimates. Eventually, the final figures are released, but even then there is a possibility of further revisions, perhaps years down the line. A true direct real-time measure of U.S. recessions is thus all but impossible. The best that economists have been able to accomplish so far is to produce estimates of current economic growth from available data for past periods, relying as much as possible on data with shorter reporting lags.

          Chauvet and Piger [R57] construct real time recession estimates using two alternative models that extend, respectively, the nonparametric approach of Harding and Pagan [R8] and the Markov switching approach of Hamilton [R6]. The models are used to simulate real-time identification of NBER turning point dates associated with the four recessions that started in 1980, 1981, 1990, and 2001. Results show that the Markov switching model has an edge as far as conforming more closely to the NBER dates, with estimates generally within one month of the corresponding NBER turning point. Both models are quicker than the NBER to identify cyclical troughs (ends of recessions), in one case beating the NBER announcement by more than one year. Still, identification of troughs comes at best about 6 months after the trough itself. For NBER peaks (starts of recessions), which may be more important to forecast, the results are not as impressive. Lead times over the NBER announcements, if any, are relatively small and identification of peaks lags the NBER announcement more often than not.

          Hamilton [R58] examines a greater variety of model specifications that may be used to identify business cycle turning point dates in real time. Comparing the models’ timing with the NBER as in the previous reference, he expresses cautious preference for a Markov switching model based on GDP alone, which nevertheless “could clearly be improved upon” according to the author. Qualitatively, results for this model are generally similar to those of the previous reference. The GDP-based model identifies troughs earlier than the NBER in 3 of 5 cases, in one case by one year. For peaks, the model’s identification predates the NBER announcement in only 2 of 5 cases. In the cases in which the model comes in first, however, it estimates the date of the peak to be 3 quarters earlier than the NBER date. In light of the disappointing results of these articles, Hamilton’s call for further research in this area seems very sensible.

Readings referenced from book The Economics of Recession

R57    Marcelle Chauvet and Jeremy Piger (2008), ‘A Comparison of the Real-Time Performance of Business Cycle Dating Methods’, Journal of Business and Economic Statistics, 26 (1), January, 42–9

R58    James D. Hamilton (2011), ‘Calling Recessions in Real Time’, International Journal of Forecasting, 27 (4), October–December, 1006–26