Stock price forecasting is a challenging task due to the non-stationarity, volatility, and nonlinearity of financial time series. Hence, it is necessary to evaluate the performance of stock price forecasting under varying market conditions using both conventional econometric and deep learning models. In this study, a comparative analysis of ARIMA, GARCH, LSTM, transformer encoder, TCN, and neural ensemble models is conducted for next-day stock price forecasting of NSE equities. Experiments are conducted on both large-cap and small-cap stocks using daily OHLCV data from 2015 to 2025. Per-ticker models are trained on log-return sequences augmented with rolling volatility features and evaluated using performance metrics including MAE, RMSE, and MAPE. The results indicate that ARIMA and GARCH models achieve higher forecasting accuracy for relatively stable large-cap stocks. In contrast, deep learning models significantly outperform classical methods in highly volatile small-cap equities. These findings suggest that forecasting performance is strongly influenced by market segment characteristics, highlighting the importance of volatility-aware model selection in equity price prediction.

In this work, a comparative analysis of statistical and deep learning models for next-day stock price forecasting on NSE equities was conducted. The results indicate that ARIMA and GARCH models achieve strong performance on large-cap stocks characterized by stable trends and moderate volatility. However, their effectiveness is reduced when applied to small-cap stocks with higher volatility and nonlinear price movements. Deep learning models, particularly LSTM and TCN architectures, demonstrate improved robustness on small-cap equities by effectively capturing nonlinear temporal dependencies. The neural ensemble model provides competitive results but does not consistently outperform the best individual models across all market segments. Overall, the findings confirm that forecasting performance is strongly dependent on market characteristics and volatility levels. Consequently, the selection of forecasting models should be guided by the structural properties of the target market segment. Future research may focus on adaptive hybrid frameworks and volatility-conditioned model selection strategies to further enhance forecasting reliability.

[1]M. S. Chowdhury, N. Nabi, M. N. U. Rana, M. Shaima, H. Esa, A. Mitra, M. A. S. Mozumder, I. A. Liza, M. M. R. Sweet, and R. Naznin, “Deep learning models for stock market forecasting: A comprehensive comparative analysis,” Journal of Business and Management Studies, vol. 6, no. 2, p. 95, 2024. [2]W. Long, Z. Lu, and L. Cui, “Deep learning-based feature engineering for stock price movement prediction,” Knowledge-Based Systems, vol. 164, pp. 163–173, 2019. [3]A. Borovykh, S. Bohte, and C. W. Oosterlee, “Dilated convolutional neural networks for time series forecasting,” Journal of Computational Finance, vol. 22, no. 4, pp. 73–101, 2018. [4]J. Sirignano and R. Cont, “Universal features of price formation in financial markets: perspectives from deep learning,” in Machine learning and AI in finance. Routledge, 2021, pp. 5–15. [5]S. Gu, B. Kelly, and D. Xiu, “Empirical asset pricing via machine learning,” The Review of Financial Studies, vol. 33, no. 5, pp. 2223– 2273, 2020. [6]A. Tsantekidis, N. Passalis, A. Tefas, J. Kanniainen, M. Gabbouj, and A. Iosifidis, “Using deep learning to detect price change indications in financial markets,” in 2017 25th European signal processing conference (EUSIPCO). IEEE, 2017, pp. 2511–2515. [7]O. B. Sezer, U. Gudelek, and A. M. Ozbayoglu, “Financial time series forecasting with deep learning: A systematic literature review,” Applied Soft Computing, vol. 90, p. 106181, 2020. [8]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, 2020. [9]J. Patel, S. Shah, P. Thakkar, and K. Kotecha, “Stock price forecasting using deep learning: A comparative study,” Journal of King Saud University – Computer and Information Sciences, vol. 33, no. 8, pp. 1018–1031, 2021. [10]B. Lim, S. O. Arık, N. Loeff, and T. Pfister, “Temporal fusion transformers for interpretable multi-horizon time series forecasting,” International Journal of Forecasting, vol. 37, no. 4, pp. 1748–1764, 2021. [11]H. Wu, J. Xu, J. Wang, and M. Long, “Autoformer: Decomposition transformers with auto-correlation for long-term series forecasting,” Advances in Neural Information Processing Systems, vol. 34, pp. 22419–22430, 2021. [12]T. Zhou, Z. Ma, Q. Wen, X. Wang, L. Sun, and R. Jin, “Fedformer: Frequency enhanced decomposed transformer for long-term series forecasting,” Proceedings of the 39th International Conference on Machine Learning, 2022. [13]S. Bai, J. Z. Kolter, and V. Koltun, “An empirical evaluation of generic convolutional and recurrent networks for sequence modeling,” Journal of Machine Learning Research, vol. 21, no. 113, pp. 1–33, 2020. [14]A. Borovykh, S. Bohte, and C. Oosterlee, “Conditional time series forecasting with convolutional neural networks,” Journal of Computational Finance, vol. 24, no. 3, pp. 1–39, 2021. J. Sirignano and R. Cont, “Universal features of price formation in financial markets: Perspectives from deep learning,” Quantitative Finance, vol. 20, no. 9, pp. 1449–1459, 2020.