Presenting an Optimized CNN-LSTM Model for Stock Price Forecasting in the Tehran Stock Exchange

Purpose: Understanding stock price behavior is one of the primary objectives for investors and traders in financial markets. Accurate price predictions enable informed decisions about buying, selling, or holding stocks, thereby optimizing transaction timing to maximize profits or minimize losses. Stock price prediction is inherently challenging due to the nonlinear nature of price movements and their dependence on various factors. To address these challenges, this study employs deep learning models, a subset of machine learning techniques known for their capability to handle large datasets effectively. The layered architecture and interconnected nodes (neurons) of these models enable advanced pattern recognition and relationship analysis, resulting in more precise price forecasts.
Method: This study proposes two optimized deep learning models, LSTM-CNN and CNN-LSTM, for predicting stock prices in the Tehran Stock Exchange. The models are fine-tuned through hyperparameter optimization using the Particle Swarm Optimization (PSO) algorithm, a population-based method widely recognized in machine learning. Additionally, advanced techniques such as adversarial training, attention mechanisms, and residual blocks are employed to enhance the models’ performance. The proposed models are benchmarked against conventional approaches, including CNN, LSTM, and CNN-LSTM models without optimization. The dataset includes information from 10 selected stocks traded in the Tehran Stock Exchange over a 10-year period, spanning from September 11, 2013, to September 11, 2023. Input variables include adjusted stock prices, financial indicators, oscillators, the free market US dollar price, and inflation rates, providing a comprehensive representation of factors influencing stock price behavior.
Findings: The evaluation metrics used in this study—RMSE, MAE, R-squared, and MAPE—reveal that the proposed deep learning models significantly outperform traditional approaches. Among these, the LSTM-CNN model demonstrates superior performance across all metrics, achieving higher prediction accuracy compared to the CNN-LSTM model. Analysis of the models with and without the PSO algorithm confirms that incorporating PSO significantly enhances model optimization. Strategy-based evaluations conducted on five selected stocks from five different industries, over four distinct time periods, further validate the superior financial performance of the proposed models. Specifically, the LSTM-CNN and CNN-LSTM models excel in terms of financial return metrics and the Sharpe ratio. Among the two, the LSTM-CNN model consistently achieves the most favorable outcomes, demonstrating its reliability and robustness for stock price forecasting.
Conclusion: Hyperparameters, which must be set before training, significantly influence the performance of deep learning models. The findings emphasize that hyperparameter optimization tailors models to stock-specific behaviors, enabling more accurate price predictions. The LSTM-CNN model outperforms the CNN-LSTM model in feature extraction and identifying dependencies within the data, resulting in improved accuracy. By integrating advanced techniques such as adversarial training, attention mechanisms, and the PSO algorithm, the proposed models provide a robust framework for stock price forecasting in the Tehran Stock Exchange. These results underscore the critical role of deep learning in financial markets, offering investors and traders tools for more informed decision-making. The LSTM-CNN model, in particular, demonstrates exceptional performance in prediction accuracy and financial metrics, making it a valuable approach for enhancing predictive analytics in the capital market.