Tree-Based Machine Learning Approaches for Equity Market Predictions

We empirically analyze equity premium predictions with ‘traditional' linear regression models and tree-based machine learning approaches. Based on a commonly used dataset of equity market predictors extended by additional fundamental, macroeconomic, sentiment and risk indicators, we find mixed results for machine learning algorithms for equity market predictions. In contrast to sophisticated linear regression models such as penalized least squares or principal component regressions (PCR), the analyzed machine learning algorithms fail to significantly out-perform the historical average benchmark forecast. However, an investment strategy that uses machine learning predictions in a market-timing strategy outperforms a passive buy-and-hold investment. Compared to sophisticated linear prediction models machine learning algorithms do not improve forecast accuracy in our problem set