This study evaluates the benefits of integrating return forecasts from a variety of machine learning and forecast combination methods into an out-of-sample asset allocation framework. The economic evaluation of the forecasts shows that model complexity translates to improved results in the majority of cases considered, with shrinkage methods and shallow neural networks generating the highest individual performance. Overall, an investor would consistently realize superior out-of-sample gains by incorporating forecast combinations of machine learning models in the portfolio formation process.
