Profit scoring with machine learning in marketplace lending: Incorporating soft information

This paper examines the impact of non-semantic and semantic textual features on profit scoring in marketplace lending and compares the out-of-sample performance of machine learning models to that of traditional linear regression models. We find that the inclusion of soft information modestly enhances the model’s explanatory power, with certain aspects of soft information demonstrating significant predictive power for return rates. While linear machine learning models that employ regularization, such as Lasso, Bayesian Ridge, and Elastic Net, offer some advantages, none of them significantly outperform the baseline OLS model in out-of-sample performance. In contrast, most non-linear machine learning models, including the tree-based and instance-based models as well as feedforward neural network models, significantly outperform the OLS models, particularly when soft information is incorporated. Our findings shed lights on improving decision-making efficiency by incorporating soft information and employing non-linear machine learning models.