Interpretable Machine Learning for Asset Pricing

We use deep neural networks to estimate time-varying equity risk premia. The key innovations are the nonlinear and non-parametric generalisation of Fama-Macbeth regressions through partial derivatives of an arbitrary estimator function with respect to its input and the introduction of Jacobian regularisation in the objective function. Our methodology not only contributes to the growing strand of literature investigating the usefulness of machine learning in economics and finance. We also show that the use of partial derivatives and derivative penalisation has the potential to promote interpretability of machine learning methods and offer useful insights for economic policy