English Français
Andreas Neuhierl - Washington University - St Louis
The Uncertainty of Machine Learning Predictions in Asset Pricing
We develop novel methodology to construct forecast confidence intervals (FCI) for machine learning predictions in asset pricing. We show FCIs for machine learning predictions obtained from sophisticated ML methods, such as neural networks, can be accurately approximated by simpler nonparametric methods such as B-splines. We prove that these FCIs provide correct coverage probabilities. In addition, we also establish the validity of a version of the wild bootstrap. We illustrate the practical use of the obtained confidence intervals in the context of a portfolio selection application for an uncertainty averse investor.