Stock Market Prediction Model Based on Deep Learning with Hybrid CNN-LSTM Model

This work was undertaken to investigate whether using the CNN-LSTM combination model on the principled, return-based feature set from S&P 500 equity data yields any forecasting performance gains over the pure LSTM approach. The high R values achieved even for Ridge Regression indicates that the return-based feature engineering approach is powerful on its own. From the experimental results we see that the CNN-LSTM obtain prediction accuracy of 98.52% and R of 0.9507 on the independent AAPL test split, which still outperforms the two classical baseline methods and the basic LSTM in all the presented aspects with slight and consistent advantage. A variety of directions still could be followed into the future. Test the pipeline's viability for application into multiple stocks, in relation to varying industries, and market capitalization, would also be interesting. Feature representation can be enriched by including additional information sources: eg. Macroeconomic statements, earning announcements, sentiment derived from financial news. Finally, attention based architecture such as Temporal Fusion Transformer can be seen as promising comparable. As there explicitly model temporal weights of attention, and their performance on multivariate time-series benchmarks is high. In summary, this study demonstrates that with disciplined feature engineering and rigorous experimental design, deep learning models can achieve strong predictive performance on individual equity time series. The CNN-LSTM architecture, in particular, offers a compelling balance between model complexity and predictive accuracy, making it a promising candidate for deployment in systematic trading and risk management applications.

[1]  P. Pilla and R. Mekonen, "Forecasting S&P 500 Using LSTM Models," arXiv preprint arXiv:2501.17366, Jan. 2025. [2]  H.H. Htun, M. Biehl, and N. Petkov, "Forecasting relative returns for S&P 500 stocks using machine learning," Financial Innovation, vol. 10, no. 1, pp. 1–28, 2024. [3]  X. Zhong and D. Hitchcock, "S&P 500 stock price prediction using technical, fundamental and text data," arXiv preprint arXiv:2101.09897, 2021. [4]  Z. Lanbouri and S. Achchab, "Stock Market Prediction on High Frequency Data Using Long-Short Term Memory," Procedia Computer Science, vol. 175, pp. 603–608, 2020. [5]  T. Fischer and C. Krauss, "Deep learning with long short-term memory networks for financial market predictions," European Journal of Operational Research, vol. 270, no. 2, pp. 654–669, 2018. [6]  H.Y. Kim and C.H. Won, "Forecasting the volatility of stock price index: A hybrid model integrating LSTM with multiple GARCH-type models," Expert Systems with Applications, vol. 103, pp. 25–37, 2018. [7]  S. Siami-Namini, N. Tavakoli, and A. Siami Namin, "A comparison of ARIMA and LSTM in forecasting time series," in Proc. IEEE ICDMW, pp. 1394–1401, 2018. [8]  C. Krauss, X.A. Do, and N. Huck, "Deep neural networks, gradient-boosted trees, random forests: Statistical arbitrage on the S&P 500," European Journal of Operational Research, vol. 259, no. 2, pp. 689–702, 2017. [9]  A.A. Adebiyi, A.O. Adewumi, and C.K. Ayo, "Stock Price Prediction Using the ARIMA Model," in Proc. 16th Int. Conf. Computer Modelling and Simulation (UKSim-AMSS), pp. 106–112, 2014. [10] L. Cao and F.E.H. Tay, "Support vector machine with adaptive parameters in financial time series forecasting," IEEE Transactions on Neural Networks, vol. 14, no. 6, pp. 1506–1518, 2003.