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

Purpose: One of the primary objectives for investors and traders in financial markets is understanding stock price behavior. Accurate price predictions can assist individuals in making informed decisions about buying, selling, or holding stocks, thereby maximizing potential profits or minimizing losses through effective transaction timing. Due to the nonlinear behavior of prices and their dependence on various factors, predicting stock prices is challenging. To address this, the research employs deep learning models, a subset of machine learning models known for their capability to handle large datasets. Their structure, featuring multiple layers and interconnected nodes (neurons), allows for recognizing patterns and relationships, facilitating more accurate price predictions.
Method: This study utilizes two proposed models, LSTM-CNN and CNN-LSTM, to predict stock prices in the Tehran Stock Exchange. These models are optimized through hyperparameter tuning using the PSO algorithm, a population-based optimization technique, along with model enhancement techniques such as adversarial training, attention mechanism, and residual blocks. The proposed models are compared with CNN, LSTM, and CNN-LSTM models. Data from 10 stocks, spanning September 11, 2013, to September 11, 2023, were analyzed. The input data include adjusted stock prices, indicators, oscillators, the free market US dollar price, and the inflation rate.
Findings: Comparative results of evaluation metrics, including RMSE, MAE, R-squared, and MAPE, indicate that the two proposed models outperform other models. The LSTM-CNN model, in particular, has demonstrated the best performance. Evaluation of the proposed models, with and without the PSO algorithm, reveals that the algorithm aids in optimizing the models. Furthermore, strategy analysis based on the proposed models, applied to five stocks from five different industries over four time periods, shows superior financial performance. In essence, the LSTM-CNN and CNN-LSTM models have achieved notable success in terms of financial returns and the Sharpe ratio compared to other strategies, with the LSTM-CNN model showing the most favorable performance.
Conclusion: Hyperparameters are critical as they impact model outcomes, and varying these parameters can produce different results, which may be either suitable or unsuitable. These parameters are not learned during training and must be set before training. The findings suggest that hyperparameter optimization can fine-tune models based on stock behavior, leading to more accurate price predictions. Additionally, the LSTM-CNN model excels in feature extraction and understanding dependencies within the data compared to the CNN-LSTM model, thereby enhancing prediction accuracy. In summary, leveraging deep learning models, specifically LSTM-CNN and CNN-LSTM, optimized with PSO and advanced techniques, significantly improves stock price prediction in the Tehran Stock Exchange. The results indicate that these models, particularly the LSTM-CNN, offer superior performance in financial return metrics and prediction accuracy. This approach provides a robust framework for investors and traders seeking to make informed decisions in financial markets, highlighting the importance of advanced machine learning techniques in enhancing predictive analytics.