Improving the Prediction of Asset Returns With Machine Learning by Using a Custom Loss Function

Abstract

Not all errors from models predicting asset returns are equal in terms of impact on the efficiency of the algorithm: a small error could trigger poor investment decisions while a significant error has no financial consequences. This economic asymmetry, critical for assessing the performance of algorithms, can usefully be replicated within the machine learning algorithms itself through the loss function to improve its prediction capability. . In this article: (a) we analyze symmetric and asymmetric loss functions for deep learning algorithms. We develop custom loss functions that mimic the asymmetry in economic consequences of prediction errors. (b) We compare the efficiency of these custom loss functions with MSE and the linear-exponential loss “LinEx”. (c) We present an efficient custom loss function that significantly improves the prediction of asset returns with improved risk-return metrics (like Sharpe ratio twice better), and which we confirm to be robust.