Machine Learning the Shortcomings of Financial Models

Abstract

This thesis investigates whether modern statistical learning methods can learn the shortcomings of financial models. In the context of this thesis, we consider option pricing models and define shortcomings as structural pricing errors. We split the learning task into two separate parts: prediction and inference. We find that complex statistical learning methods can predict a large part of pricing errors. Neural Networks explain upwards of 89% of pricing errors in financial models. Furthermore, under some simplifying assumptions, we can infer the effects of selected interesting variables. These results improve our understanding of theoretical models since they show areas of potential improvement.