Covid broke a lot of forecasting models.
The New York Fed suspended their GDP Nowcasting model because the data were so volatile that it predictions turned to rubbish.
How will economic forecasters re-tool their models?
Here are a few interesting papers:
Schorfiede and Song (2021) find that a mixed frequency vector autoregression (MF-VAR) estimated as a state space with 1Q 2020 excluded leads to stable predictions.
Götz and Hauzenberger (2021) modify the Schorfiede-Song MF-VAR so that the intercept and covariance matrix (common stochastic volatitly) are time-varying, but not the slope coefficients. This allows for a novel interpretation wherein the slope coefficients preserve a fundamental economic relationship while the intercept acts as an evolving recession dummy and stochastic volatility insures against craziness the intercept cannot detect.
Ng (2021) “de-COVIDs” data by using data on positive cases, hospitalizations, and deaths. The de-COVID data then can be used to estimate a more standard model. This approach helped inform my personal approach to forecasting the labor market into year-end 2021.
Lenza and Primiceri (2020) a priori specifies extreme shocks to the conditional covariance structure during the outbreak, softening damage to VAR coefficients and preserving some resemblance of pre-COVID behavior.
Primiceri and Tambalotti (2020) “bends” forecasts to a prior belief supposing Covid as a one-quarter chock with a polynomial recovery trajectory. However, the Covid recession, not being the result of a typical business cycle, broke the mold those prior beliefs were based upon, under-estimating the strength of the recovery.
It seems that one’s way of viewing COVID itself is essential to how the forecaster overcomes these issues.
Schorfiede and Song treat the COVID episode as observations to be ignored or interpolated: outliers. Ng treats it as a unique economic shock whose intensity can be modeled non-economically through time. Lenza and Primiceri treat it as a sudden evolution move in the conditional covariance structure.
Conceptually, I like Ng’s approach of trying to endogenize COVID, rather than treat the outbreak as too extreme to model explicitly or just chalking it up to volatility. But which one yields the best forecasts is an empirical question. That much has yet to be seen.
Götz, T. B., & Hauzenberger, K. (2021). Large mixed-frequency VARs with a parsimonious time-varying parameter structure. The Econometrics Journal, 24(3), 442-461.
Lenza, M., & Primiceri, G. E. (2020). How to Estimate a VAR after March 2020 (No. w27771). National Bureau of Economic Research.
Primiceri, G. and Tambalotti, A. 2020, Macroeconomic Forecasting the Time of COVID-19, mimeo, Northwestern University.
Ng, S. (2021). Modeling macroeconomic variations after COVID-19 (No. w29060). National Bureau of Economic Research.
Schorfheide, F., & Song, D. (2021). Real-time forecasting with a (standard) mixed-frequency VAR during a pandemic (No. w29535). National Bureau of Economic Research.