Stock return prediction is a fundamental but difficult issue in financial research because of the noise and dynamics of market data. Although recent research has improved the ability of existing machine learning model prediction, these methods are difficult to interpret and lack practical application value due to their "black box" nature. This paper constructs an explainable machine learning framework to achieve high-accuracy predictions of short-term stock excess returns and strive for greater interpretability in economics; Compared with other complex models. Taking the daily trading data of Apple Inc. (AAPL) as a single-stock example, this paper constructs a set of economically meaningful features based on historical returns, volatility and trading volume. A linear regression model is used as the central prediction tool, and interpretability can be built into this specific form of a linear model. Out-of-sample forecasts use the rolling window method, and prediction performance is compared with a naive historical mean benchmark. Based on experiments, the current basis of explanation using an explainable linear model has been shown to be relatively stable in baseline prediction. According to the coefficient analysis, in addition to conforming to some basic Finance theory patterns. A return pattern that exists for one day appears. A Volatility-Return relationship also holds currently. This paper shows that interpretive models, in combination with carefully selected features, achieve satisfactory prediction performance; at the same time, it is also found through the analysis of economic implications that such an approach helps explore the essence of changes in stock returns.

Based on machine learning methods, this paper systematically studies the cross-sectional excess returns of A-share stocks. Taking 5,648 stocks in China's A-share market from 2020 to 2024 as a sample, a multi-dimensional factor system covering fundamentals and macro variables was constructed, and the out-of-sample performance of multiple models was compared under the rolling forecast framework of Expanding Window. The empirical results show that the nonlinear machine learning model continues to obtain positive information coefficients in most quarters, and its prediction signals can be converted into stable long-short portfolio returns, while the prediction ability and investment performance of traditional linear models decay significantly with time. In terms of risk control, LightGBM shows a relatively better trade-off between return level and drawdown magnitude. The research in this paper confirms that the machine learning method that combines a nonlinear structure and cross-sectional heterogeneity can effectively improve the return prediction ability and investment performance of the A-share market and provides practical empirical evidence for quantitative stock selection.