The most common models to assess asset returns are a linear combination of risk factors. We have employed tree-based machine learning algorithms to capture nonlinearities and detect interaction effects among risk factors in the EUR and USD credit space. We have built a nonlinear credit pricing model and compared it to our baseline linear credit pricing model using error metrics on training and testing sets and during different periods. In-sample error metrics revealed the benefit of treebased regressions.
Then, we analysed the explanatory and predictive power measure by factor category and by period in order to evaluate the contribution of each factor in the explanation and prediction of credit excess returns. We found value in adding alternative factors to a traditional factor model and point out which of them prevail across different time horizons and during market crisis periods.
Finally, tree-based regressions methods assisted us in improving our understanding of prices through the interaction between features and between each feature and the output of the model.
To find more, download the full paper.