Recent research by authors from the IMF, ECB and the Bank of England has identified Random Forests as the most effective method for predicting crisis episodes and superior to the more traditional approach of probit/logit modelling. We challenge that finding when predicting recessions for 20 OECD countries. A customised algorithm that selects probit models, matches the performance of Random Forests in out-of-sample quarterly predictions using real-time data over a two-year horizon, including the period of the Global Financial Crisis. This enhanced performance is attributed to a "wisdom of crowds" feature whereby predictions are averaged from many well-fitting probit equations, comparable to Random Forests averaging many trees. Both country-specific and pooled Random Forests are estimated, and, although the latter has a superior out-of-sample performance, a disadvantage of the pooled approach is that recession risks are highly correlated across countries and rarely very elevated, so that it is unlikely that a recession is ever 'more likely than not'.
The estimation framework is also separately applied to 8 consecutive quarterly horizons and demonstrates that different explanatory variables matter at different horizons; activity variables are more important for immediate quarters, but financial cycle variables (especially credit and house prices) dominate at horizons beyond that.
