Multi-head attention based LSTM model for forecasting Bangladeshi stock market prices

Abstract

Forecasting stock prices is crucial for investment and risk management, offering insights into market trends and aiding investors in making informed decisions. Traditional time-series models, such as LSTM and GRU, have demonstrated promise in capturing temporal dependencies within stock market data. However, these models often fall short in identifying complex patterns due to limited attention mechanisms, particularly in emerging markets with unique economic factors like the Bangla stock market. Addressing this gap, we propose a novel Multi-Head Attention and LSTM-based architecture that combines the sequential processing capabilities of LSTM with multi-head attention layers to better capture contextual dependencies and subtle fluctuations in stock price movements. Our approach is evaluated on datasets from six major Bangladeshi banks, including Islami Bank, Dhaka Bank, Arab Bangladesh (AB) Bank, Bank Asia, Eastern Bank Limited (EBL), and City Bank. Results demonstrate that our model outperforms traditional models across all metrics. The integration of multi-head attention allows for nuanced attention allocation across time steps, significantly enhancing forecasting accuracy. This enhanced model can serve as a valuable tool for investors and financial analysts, supporting better forecasting for risk assessment and investment strategy formulation. The framework demonstrates potential scalability across other emerging markets, offering an adaptable solution for financial forecasting in regions with complex market dynamics.