ABSTRACT This paper introduces a financial risk forecasting model that effectively exploits information from a large set of economic and financial predictor variables. The model is built using generalized quantile random forests, a nonparametric machine learning method that naturally permits variable interactions and nonlinear relationships. We use a model‐free variable screening technique and a robust cross‐validation approach to minimize the risk of overfitting. Our risk model produces competitive value‐at‐risk and expected shortfall forecasts at both 1‐day‐ahead and 10‐day‐ahead horizons. A dynamic portfolio insurance strategy that uses the VaR and ES forecasts from our risk model generates attractive Sharpe, Sortino, and Omega ratios, particularly at the 10‐day forecast horizon. Additionally, we provide a detailed analysis of the dynamic importance of our predictor variables. Our findings demonstrate the utility of combining large datasets with tree‐based algorithms for financial risk forecasting.
James et al. (Mon,) studied this question.