Text-Based Recession Probabilities

This paper proposes a new methodology based on textual analysis to forecast U.S. recessions. Specifically, the paper develops an index in the spirit of Baker et al. (2016) and Caldara and Iacoviello (2018) which tracks developments in U.S. real activity. When used in a standard recession probability model, the index outperforms the yield curve based forecast, a standard method to forecast recessions, at medium horizons, up to 8 months. Moreover, the index contains information not included in yield data that are useful to understand recession episodes. When included as an additional control to the slope of the yield curve, it improves the forecast accuracy by 5% to 30% depending on the horizon. These results are stable to a number of different robustness checks, including changes to the estimation method, the definition of recessions and controlling for asset purchases by major central banks. Yield and textual analysis data also outperform other popular leading indicators for the U.S. business cycle such as PMIs, consumers' surveys or employment data.