U.S. recession forecasting with extreme gradient boosting

---

In this study, I forecast U.S. recessions using the extreme gradient boosting algorithm (XGBoost) within a real-time out of sample prediction setting, considering factors such as the availability and delay of data releases for each training period. The macroeconomic time series used as predictors are based on the reference series of the OECD CLI for the U.S.. The predictive ability of the model is assessed up to 4 future horizons on quarterly basis. A naive forecast and a baseline logit are used as benchmark models. To identify downturns, I calculate GDP turning points using a simple method based on the four business cycle phases. The results show that compared to the benchmark models, one year forecast accuracy can be significantly increased when using the XGBoost model.

Course	Applied Econometrics (M.Sc.)
Grade	1.3