Stock market prediction remains a complex task due to the nonlinear and volatile nature of financial time-series data. This study presents an LSTM-based equity price prediction system designed to capture long-term dependencies in historical stock data. Using five years of daily closing prices, the data were preprocessed through normalization and transformed into 60-day sequential windows for model training. A stacked LSTM architecture with dropout regularization was implemented to forecast next-day stock prices. The model demonstrated strong trend-learning capability, achieving low validation loss and a root mean square error (RMSE) of approximately 30.89 on test data. Visual comparison between actual and predicted prices showed close alignment, indicating effective temporal pattern recognition. The findings highlight the suitability of LSTM networks for short-term equity forecasting and provide a foundation for further enhancement through hybrid architectures and multi-feature integration in financial prediction systems.

The primary objective of this study was to develop and evaluate an LSTM-based system for forecasting equity prices, using historical closing prices of a major stock (Reliance Industries) as input. In doing so, we aimed to test whether deep recurrent neural networks could effectively learn and predict future price movements from past data. Consistent with prior research, our LSTM model successfully learned salient temporal patterns: after training on five years of daily data, it achieved a root-mean-square error (RMSE) of approximately 30.9 on the held-out test set. The predicted price series closely tracked the actual closing prices, capturing the overall upward trend of the stock (Figure 1). We also produced a point forecast (≈₹1423.12) for the next trading day. These results indicate that the LSTM system captured much of the stock’s volatility and trend. In summary, the key findings are that the LSTM network converged to a stable loss (validation MSE ~0.0018), produced coherent next-day predictions, and generated realistic price trajectories closely aligned with historical patterns. The broader significance of these findings is that they reinforce the theoretical promise of deep learning for financial time-series modeling. Equity markets are known to exhibit highly volatile, nonlinear behavior, and traditional linear methods (such as ARIMA) often struggle with such complexity. Our study confirms that an LSTM – a deep recurrent architecture designed to capture long-range dependencies – can approximate these complex dynamics. In particular, the ability of our LSTM model to mirror real price movements supports theoretical arguments that recurrent networks can “deal with complex patterns in stock prices” and achieve higher accuracy on nonlinear time-series data. Thus, our work contributes evidence that data-driven, learning-based models (especially LSTM networks) can effectively extract information from sequential price history without relying on strong a priori assumptions. This aligns with the modern view that deep sequence models, operating on raw price series (and potentially additional signals), offer a powerful approach to stock prediction. Looking forward, these findings suggest several implications for future research. First, our results indicate that enriching the input space could further improve accuracy. For example, recent studies have shown that combining price series with broader market indicators (e.g. macroeconomic data, technical indicators, or sentiment scores) can boost performance. In particular, Li et al. demonstrated that integrating additional features via symbolic genetic programming yielded significant gains in LSTM prediction accuracy. Future work could therefore explore multi-source data fusion – for instance, merging fundamental, technical, and news-sentiment inputs – to provide the LSTM with more comprehensive signals. Second, advanced model architectures merit exploration: hybrid and ensemble approaches (LSTM combined with other learning methods or newer architectures like Transformers) may capture market patterns more robustly. Attention-based models, for instance, could help the system focus on the most informative time periods or features. Third, rigorous validation on diverse asset classes and in real-world trading simulations would test the model’s generalizability. We encourage future research to build on our framework by conducting cross-market studies, longer-term forecasts, and integrating risk metrics, thereby deepening understanding of how sequence models behave under different market regimes. We acknowledge several limitations in our study. One key limitation is data scope: we used only one stock’s historical prices and did not include exogenous variables (such as macro signals or news). This restricts the generality of our conclusions and risks overfitting; indeed, LSTM models are known to be susceptible to overfitting when data are scarce. Relatedly, the internal mechanics of deep networks are “black boxes,” making interpretability challenging. Our system also used a fixed two-layer LSTM with simple hyperparameters; more elaborate tuning or architecture search was not performed. Finally, we evaluated performance primarily by RMSE and visual fit, without testing practical trading viability (e.g. cost-aware strategies). Addressing these limitations in future work could involve collecting larger and more varied datasets, applying regularization and explainability techniques, and benchmarking against stronger baselines or ensemble models. Such extensions would help to ensure robustness and real-world applicability. In conclusion, this study advances our understanding of LSTM-driven equity prediction by empirically validating that an LSTM network can learn and forecast realistic stock price movements from historical data. By demonstrating the model’s predictive capability and discussing its strengths and weaknesses, we lay the groundwork for more sophisticated forecasting systems. Future investigations that incorporate richer data sources, newer neural architectures, and rigorous evaluation will continue to refine the role of deep sequence models in financial theory and practice. Our work thus contributes to the evolving narrative that deep learning – and LSTM models in particular – holds promise for modeling complex financial time series, and it opens avenues for integrating these methods into next-generation trading and analytics frameworks.

[1] J. Zhang, Y. Li, and H. Wang, “LSTM-Based Stock Market Prediction Using Deep Learning Techniques,” in Proc. IEEE Int. Conf. Artificial Intelligence and Smart Systems (ICAIS), 2021, pp. 112–118. [2] A. Sharma and R. Kumar, “Comparative Analysis of LSTM and GRU Networks for Financial Time Series Forecasting,” in Proc. IEEE Int. Conf. Data Science and Advanced Analytics (DSAA), 2022, pp. 456–463. [3] M. Chen, T. Liu, and Q. Zhao, “Hybrid CNN–LSTM Model for Stock Price Forecasting,” Expert Systems with Applications, vol. 202, pp. 117–130, 2022. [4] S. Patel and D. Mehta, “Attention-Based LSTM Network for Financial Time-Series Prediction,” Neural Computing and Applications, vol. 35, no. 4, pp. 2891–2905, 2023. [5] Y. Kim and J. Lee, “Multivariate Time Series Forecasting Using Stacked LSTM Networks,” IEEE Access, vol. 11, pp. 76543–76555, 2023. [6] R. Gupta, P. Singh, and A. Verma, “Deep Learning Approaches for Stock Market Prediction: A Review,” Applied Soft Computing, vol. 134, 2024, Art. no. 110021. [7] L. Brown and M. Davis, “Optimized Recurrent Neural Networks for Financial Forecasting,” in Proc. ACM Int. Conf. AI in Finance (ICAIF), 2024, pp. 78–86. [8] H. Zhou, K. Tan, and W. Li, “Stock Market Forecasting Using Deep Learning with Sentiment Integration,” in Proc. IEEE Int. Conf. Big Data, 2025, pp. 1502–1510. [9] McKinsey & Company, The State of AI in Financial Services, Industry Report, 2023. [10] Deloitte, AI and Machine Learning in Capital Markets, Industry Report, 2024. [11] PwC, AI-Powered Financial Forecasting Trends, Industry Report, 2024.

© 2025 International Journal of Engineering and Techniques (IJET).

Digital Object Identifier (DOI)

IJET assigns a unique DOI to every accepted article via Zenodo for persistent linking and citation. Your article’s DOI: https://doi.org/{{doi}}

Open DOI About Zenodo DOI

Close