Deep models for empirical asset pricing (risk-premia forecast) and their interpretability

Abstract

Risk premia measurement is an essential problem in Asset Pricing. It is estimation of how much an asset will outperform risk-free assets. Problems like noisy and non-stationarity of returns makes risk-premia estimation using Machine Learning (ML) challenging. In this work, we develop ML models that solve the associated problems with risk-premia measurement by decoupling risk-premia prediction into two independent tasks and by using ideas from Deep Learning literature that enables deep neural networks training. The models are tested robustly using different metrics where we observe that our model outperforms existing standard ML models. One another problem with ML models is their black-box nature. We also interpret the deep neural networks using local approximation based techniques that make the predictions explainable.