Ensembling Machine Learning Models For Predicting Stock Index Returns

We develop a conditional machine learning (CML) approach to estimate factors in a conditional latent-factor model for expected-returns. We extract the factors from deep neural network-based estimates of individual stock returns to construct stock index return forecasts and derive a closed form expression for their associated uncertainty. The uncertainty of forecast-uncertainty is heightened during market-disruptions with potentially unstable latent-factor-structure, indicated by lower cross-sectional correlation between firms’ ranks based on their sales revenues and market values. We therefore combine the CML approach, and the standard unconditional-machine-learning approach that does not assume the latent-factor structure, to construct robust ensemble forecasts.