A Study on the Incremental Value of Technical Indicator Features in Predicting the Next-Day Direction of the CSI 300 Index

Whether technical indicators can provide incremental information for predicting the direction of stock price movements is a classic topic in financial market research. Using the CSI 300 Index as our research setting, we select daily data from 1,073 valid trading days from July 2020 to December 2024, construct a dataset containing 19 features such as historical returns, moving averages, RSI, MACD, Bollinger Bands, KDJ, and trading volume, and use Logistic regression and random forest models to rigorously evaluate the incremental value of technical indicators in prediction. We adopt a rigorous time-series split. In the main experiment, July 2020 to December 2023 is used as the training set (831 samples), the whole year of 2024 is used as the test set (242 samples), and five rolling windows are used to test the temporal robustness of the conclusions. We find that: (1) the best model, the random forest model using only historical return features, has a test-set accuracy of 56.20%. The Bootstrap test shows that this accuracy is higher than the random-guessing baseline at the 95% confidence level, but its difference from the naive baseline is not statistically significant (one-sided p = 0.123), indicating that the incremental predictive value of technical indicators is statistically fragile; (2) after technical indicators are added, the training-set accuracy of the random forest increases from 0.7052 to 0.7858, but the test-set accuracy decreases from 0.5620 to 0.5372, and the generalization gap expands from 14.32 percentage points to 24.86 percentage points. This paradox of feature increase, training improvement, and generalization deterioration is our core finding. The redundancy-removal experiment and the regularization control experiment further show that feature redundancy and limited samples jointly lead to the decline in generalization performance; (3) model prediction shows significant and stable asymmetry between up days and down days. The recall rate for down days (78.05%) is much higher than that for up days (33.61%), the predicted proportion of up days is only 27.69%, and this bearish bias persists in all five rolling windows. Market-state analysis shows that the weak market environment in the training period, with a cumulative return of -26.92%, is an important cause of the bearish bias.