Forecasting Indian Banking Stock Prices: A Comparative Evaluation of ARIMA, ARCH, GARCH, RNN, and LSTM Models

Abstract

Objective: The main objective of this paper is to compare ARIMA, ARCH, GARCH, RNN, and LSTM forecasting models to determine the best approach for forecasting the adjusted prices.

Methodology: Daily stock price data covering the period from April 2015 to March 2025 were obtained from the National Stock Exchange to conduct the empirical analysis. days for each stock adjusted closing prices to determine the best model involves utilising the Augmented Dickey-Fuller test to determine the stationarity of each stock’s adjusted closing prices (Gayathri & Devi, 2025), thereby establishing a reliable foundation for fitting each ofthe five models to each stock’s prices, and utilising the mean squared error, root mean squared error, and mean absolute percentage error to rigorously evaluate the predictive accuracy of each model, providing a robust framework for comparing the relative performance of these statistical and deep learning approaches (Nafkha & Suchodolska, 2024). Furthermore, the empirical findings indicate how effectively hybrid architectures combining linear estimators with recurrent neural networks mitigate the volatility clustering inherent in emerging market equities (Kalange, 2025). Consequently, this investigation addresses existing methodological gaps by systematically contrasting parametric time series formulations with sequential deep learning paradigms (Satria, 2023). These quantitative insights equip institutional investors and portfolio managers with optimized risk-adjustment strategies tailored to high-volatility financial environments (Chaudhary & Minirani, 2025). Subsequent evaluations demonstrate that while recurrent architectures adeptly model complex long-term dependencies, conventional parametric frameworks often yield superior stability for near-term directional projections (Raissa, 2025).

Findings: The results indicate that the LSTM model performs best, with MAPE values for each of the six companies below 1%. For instance, the MAPE for the HDFC bank stock using the LSTM model was 0.0046%, with an MSE of 17.50; the MAPE for the Union Bank stock using the LSTM model was 0.0099%, with an MSE of 2.11. Furthermore, the empirical analysis highlights the substantial predictive advantage of deep learning methodologies; the LSTM models achieved a superior average MAPE of 0.0073% across all six banking stocks, contrasting sharply with the 19.53% average MAPE recorded for the ARIMA models. In the comparative hierarchy of model performance, the RNN architectures demonstrated the next highest level of precision, while the GARCH models, although robust in addressing volatility, followed the deep learning paradigms in overall predictive accuracy..

Originality: Finally, utilising the best models for each of the banks, forecasts for each bank’s stock 90 days into the future The LSTM model outperformed the statistical ARIMA baseline for these six Indian banks, achieving an average MAPE of 0.0073% compared to 19.53% for ARIMA, confirming the efficacy of deep learning paradigms for stock market prediction. The findings of this study may be helpful for the investors in these stocks and the analysts who monitor those stocks.

JEL Classification: G17, G53, G21, C22, G14, C45