Testing characteristics importance with neural network gradients: Evidence from the China A-share market

Abstract

This paper develops a hypothesis testing framework based on neural network gradients to evaluate the importance of firm characteristics in predicting monthly excess returns in the China A-share market. Using smooth neural networks with one to five hidden layers, we find that many characteristics, including several traditionally viewed as important, are not statistically significant. Growth and valuation characteristics dominate return prediction, whereas operational and solvency characteristics have limited explanatory power. A comparison with the traditional $R^2$-reduction approach reveals that conclusions on characteristics importance depend on the analytical method. Monte Carlo simulations confirm the robustness of the proposed discrete testing procedure. The framework offers a rigorous and interpretable approach to assessing characteristics importance, contributing to model transparency and financial forecasting.