Exchange rate forecasting is crucial for informed decision-making in financial markets, but significant challenges arise due to the high volatility and non-linear nature of economic time series. Traditional statistical models (ARIMA), state-of-the-art deep learning methods (LSTM, GRU), and hybrid models (TSMixer, in addition to AB-LSTM-GRU) all exhibit low adaptability to dynamic market conditions, as they cannot perform iterative optimization based on real-time feedback. To bridge this gap, this work presents an innovative hybrid framework that combines Long Short-Term Memory (LSTM) networks and a Deep Q-network (DQN) agent. Precisely, LSTM models capture temporal dependencies in time series data, and DQNs introduce a reinforcement learning mechanism that optimizes prediction adaptively based on feedback. The algorithm leverages the strengths of both deep learning and reinforcement learning to achieve improved predictive accuracy and adaptability. The effectiveness of the proposed model is substantiated by an experimental study based on USD/INR exchange rate data, which outperformed five existing state-of-the-art models in terms of lower Mean Squared Error (0.37) and Root Mean Squared Error (0.61). These quantitative achievements demonstrate the model's power and robustness in minimizing forecast errors. The model proposed in this study has significant implications for financial forecasting, improving the decision-making capabilities of traders, investors, and policy-makers. Its robust framework allows for greater flexibility in response to market changes, making it a potential instrument for complex financial systems.
Madhulatha et al. (2025) studied this question.