Dynamic factors estimated from panels of macroeconomic indicators are used to predict future recessions using probit models. Three factors are considered: a bond and exchange rates factor; a stock market factor; a real activity factor. Three results emerge. First, models that use only financial indicators exhibit a large deterioration in fit after 2005. Second, models that use factors yield better fit than models that use indicators directly. Out-of-sample forecasting exercises confirm these results for 3-, 6-, and 12-month horizons using both ex-post revised data and real-time data. Third, results show evidence that data revisions affect factors less than individual indicators.

Latent factors estimated from panels of macroeconomic indicators are used to generate recession probabilities for the US economy. The focus is on current (rather than future) business conditions. Two macro factors are considered: (1) a dynamic factor estimated by maximum likelihood from a set of 4 monthly series; (2) the first of eight static factors estimated by principal components using a panel of 102 monthly series. Recession probabilities generated using standard probit, autoregressive probit, and Markov-switching models exhibit very different properties. Overall, a simple Markov-switching model based on the big data macro factor generates the sequence of out-of-sample class predictions that better approximates NBER recession months. Nevertheless, it is shown that the selection of the best performing model depends on the forecaster’s relative tolerance for false positives and false negatives.

The issue of testing for a unit root allowing for a structural break in the trend function is considered. The focus is on the construction of more powerful tests using the information in relevant multi‐variate data sets. The proposed test adopts the generalized least squares detrending approach and uses correlated stationary covariates to improve power. As it is standard in the literature, the break date is treated as unknown. Asymptotic distributions are derived, and a set of asymptotic and finite sample critical values are tabulated. Asymptotic local power functions show that power gains can be large. Finite sample results show that the test exhibits small‐size distortions and power that can be far beyond what is achievable by univariate tests.

No abstract

Short excerpts of these working papers may be quoted without explicit permission provided that full credit is given to the source.