Predicting Asset Prices with Machine Learning

Abstract

This study examines whether machine learning techniques such as neural networks contain predictability when modeling asset prices and if they can improve on asset pricing prediction compared to traditional OLS-regressions. This is analyzed through measuring and comparing the out-of-sample R2 to find each models’ predictive power. Furthermore, we establish the loss metrics of root mean squared error and mean bias error to assess model strength. A sample of Swedish stocks ranging over a 40-year period is considered the dataset. We provide an analysis of various models to find indications of which models perform better from an economic viewpoint. Although we do not test for statistical significance, as forecasting returns infrequently exert this, the economic gains can prove relevant. We find that several neural networks outperform linear OLS regression in terms of out-of-sample R2. We believe that this might not be enough information to profitably transact upon as a considerable number of factors such as transaction costs are still unaccounted for. Our conclusion is therefore that further studying is required to fully allow for all factors to be considered.