This paper addresses a critical limitation in quantitative finance: the trade-off between predictive accuracy and interpretability in multi-factor models. While traditional linear multi-factor models offer transparency but limited accuracy due to their inability to capture non-linear market dynamics, recent deep learning approaches improve predictions but operate as black boxes that institutional investors cannot use in practice due to accountability requirements.

The authors propose a novel deep recurrent factor model that combines Long Short-Term Memory (LSTM) networks with Layer-wise Relevance Propagation (LRP) to create an interpretable non-linear and time-varying multi-factor model. The LSTM component captures complex temporal patterns and non-linear relationships in financial markets, while LRP provides a mechanism to approximate and linearize the learned models, making them interpretable for investment professionals. Through empirical analysis of the Japanese stock market, the authors demonstrate that their approach not only outperforms traditional linear models but also surpasses fully-connected deep learning methods while maintaining the transparency required for institutional investment practice.