Forecasting Realized Volatility : An Automatic System Using Many Features and Many Machine Learning Algorithms

We propose an automatic machine-learning system to forecast realized volatility for S&P 100 stocks using 118 features and five machine learning algorithms. A simple average ensemble model combining all learning algorithms delivers extraordinary performance across forecast horizons, and the improvement in out-of-sample R2's translates into nontrivial economic gains. We further augment the feature set by including firm characteristics and pure noise terms, and find that the system continues to perform well after including weak or noisy features. Finally, we demonstrate that our learning system is scalable to a broader S&P 500 stock universe via hyperparameter transfer learning for nonlinear models