<![CDATA[Forecasting stock market indices is challenging due to the complex, nonlinear, and dynamic behavior of financial markets. Although statistical, machine learning, and deep learning methods have been widely applied, their comparative performance for daily Composite Stock Price Index (CSPI) forecasting remains insufficiently explored. This study systematically compares nine forecasting models representing statistical (ARIMA, SARIMA, ETS), machine learning (Random Forest, Support Vector Regression, XGBoost), and deep learning (LSTM, BiLSTM, GRU) approaches using the CRISP-DM (Cross-Industry Standard Process for Data Mining) framework. Historical daily CSPI closing prices from 1995 to 2019 were preprocessed through missing-value handling, feature engineering, normalization, and chronological data partitioning. Statistical models were trained on the original price scale, whereas machine learning and deep learning models used normalized data. Hyperparameters were optimized using Bayesian Optimization with the Optuna framework. Because the models were evaluated on different scales, cross-paradigm comparisons primarily relied on the coefficient of determination (R²) and Directional Accuracy (DA). The results show that machine learning models generally outperformed statistical and deep learning approaches on the testing dataset. Random Forest achieved the highest predictive performance with an R² of 0.902, while Support Vector Regression produced the lowest MAE and MAPE among normalized models. Although LSTM achieved the lowest validation error, its performance did not generalize consistently to unseen data. Directional Accuracy remained relatively low (31–41%), indicating that predicting market direction is more difficult than forecasting price magnitude. Future studies should incorporate macroeconomic indicators and market sentiment to improve forecasting performance.]]>
