Abstract:
This study evaluates whether feature selection improves machine learning forecasts of German business cycles. Using a high-dimensional dataset with 73 indicators, primarily from the OECD Main Economic Indicator Database, covering a period from 1973 to 2023, Sequential Floating Forward Selection (SFFS) is applied to build compact, explainable, and performant models. The focus is on regularized regression models (LASSO, Ridge, Elastic Net) and tree-based classification models (Random Forest, Gradient Boosting and AdaBoost). SFFS yields models with up to eleven indicators that outperform a standard term-spread probit model—especially during Quantitative Easing. Regularized regressions provide the most accurate recession signals. Feature selection increased the forecasting power of tree-based models, while marginally reducing the performance of regression models. The findings contribute to the ongoing discussion on the use of machine learning in economic forecasting, especially in the context of limited and imbalanced data.