Predicting Index Trend Using Hybrid Neural Networks with a Focus on Multi-Scale Temporal Feature Extraction in the Tehran Stock Exchange

Message:
Article Type:
Research/Original Article (دارای رتبه معتبر)
Abstract:
Objective

Predicting the future trends in financial markets stands as a critical task for both investors and researchers, given its pivotal role in enabling well-informed decision-making processes and effective risk management strategies. Nevertheless, the realm of stock market dynamics is fraught with inherent complexities and uncertainties, posing a formidable challenge when it comes to achieving accurate predictions. A wide array of predictive modeling techniques have been meticulously investigated, spanning from conventional statistical methodologies to more sophisticated machine learning algorithms. The primary focus of this research endeavor revolves around the predictive analysis of the Tehran Stock Exchange (TSE) Composite Index, wherein a novel hybrid neural network framework is employed. This approach seamlessly integrates multiscale temporal features, with the ultimate objective of bolstering prediction precision and offering profound insights into prevailing market trends and dynamics.

Methods

The hybrid neural network architecture that has been put forward integrates the unique capabilities of convolutional neural networks (CNNs) in the realm of feature extraction with the effectiveness of long short-term memory (LSTM) networks in capturing temporal dependencies. The dataset used in this study consists of daily historical data pertaining to the TSE Composite Index, covering a substantial period from the year 1998 to 2022, which has been meticulously gathered, preprocessed, and subsequently partitioned into distinct sets for training and validation purposes. Within the framework of this hybrid neural network model, a sophisticated approach is adopted to harness multiscale temporal features derived from the input data, enabling the generation of highly accurate predictions regarding the future trends of the index. Moreover, to further enhance the performance and resilience of the model, sophisticated feature engineering methodologies are implemented to optimize its overall functionality.

Results

The results of the study reveal that while the hybrid neural network model, integrating CNN and LSTM components, demonstrates promising capabilities in predicting the TSE Composite Index, its accuracy falls short compared to competing models, particularly at weekly and monthly time scales. Conversely, the standalone CNN model exhibits superior performance, yielding more accurate predictions of the index's movements. These findings challenge the prevailing notion regarding the efficacy of hybrid neural network models in financial market prediction, highlighting the importance of evaluating alternative modeling approaches based on their specific strengths and limitations.

Conclusion

Despite the potential of hybrid neural network models, as demonstrated in previous research, the findings of this study suggest that simpler neural network architectures, such as CNNs, may offer better prediction performance in certain scenarios. To address the limitations identified, future research endeavors could explore alternative model configurations, ensemble methods, or hybrid architectures that combine the strengths of different predictive models. Additionally, incorporating additional market indicators and exploring intraday data sources could further enhance prediction accuracy and robustness. This abstract encapsulates the key findings and implications of the research, providing valuable insights for investors, researchers, and practitioners in the field of financial market prediction.

Language:
Persian
Published:
Financial Research, Volume:27 Issue: 77, 2025
Pages:
85 to 113
https://www.magiran.com/p2840540