[3Xin4] Poster session 1

[3Xin4-31] Locally Guided Global Surrogate Model for Deep Multi-Factor Strategy

〇Yugo Fujimoto¹, Kei Nakagawa¹ (1.Nomura Asset Management Co., Ltd.)

Keywords:Finance, Deep leaning, Multifactor model, Interpretability

Deep multi-factor models using deep learning(DL) have been proposed to simultaneously handle a large number of factors for stock price prediction. However, the interpretability of the output of DL is generally difficult. When considering its practical application in actual investment management, we need interpretability from the standpoint of explanatory responsibility. While local surrogates cannot explain the behavior of the model on the entire input space, global surrogates provide it with more interpretable models. However, even if the output of global surrogates is similar to that of the original model, the evidence of the decision is not always consistent. Therefore, in this study, in order to improve the interpretability of deep multi-factor models, we propose an interpretive method that applies the global surrogate models and local surrogate technique, Layer-wise Relevance Propagation (LRP). Our method learns global surrogates regularized by LRP to preserve consistency in predictions and local explanations. The experimental results with real market data demonstrate that the proposed method learns global surrogates consistent with local explanations and can be interpreted more naturally.