The paper introduces the Deep Factor Model, a unified deep learning approach for both return forecasting and risk attribution in quantitative portfolio management. Traditional multifactor models assume linear relationships between factors and stock returns, which is restrictive given market complexity. The authors replace the linear return model with a deep neural network that captures nonlinear interactions among factors. To address the black-box nature of deep learning, they apply layer-wise relevance propagation (LRP) to decompose the predicted return into contributions from individual factors, effectively creating an interpretable risk model. Empirical tests on Japanese TOPIX constituents from 1990 to 2016 show that the deep factor model significantly outperforms linear regression, support vector regression, and random forest in both prediction accuracy (MAE, RMSE) and portfolio profitability (Sharpe ratio). Notably, the deeper network (6 layers) achieves a higher Sharpe ratio (1.50) than the shallower network (1.41), while the shallower network has slightly better accuracy. LRP analysis reveals that quality and value factors dominate predictions, and that factor contributions differ from simple correlations, demonstrating the value of nonlinear attribution. The work bridges the gap between high-performance deep learning and the need for interpretability in financial applications.