Enhancing exchange rates forecasting: Leveraging long short-term memory with technical indicators

This paper provides a comprehensive evaluation of technical indicators in exchange rate forecasting by leveraging sequential architectures and demonstrating the adaptive capacity of LSTM in the presence of structural breaks. We show that the long short-term memory (LSTM) approach, when combined with technical indicators, exhibits stronger out-of-sample forecasting performance for five extensively traded currencies compared to using individual technical indicators or their combination. Our results also highlight that distinct technical indicators have varying effects on exchange rate forecasting in different countries. Additionally, we observe that the predictability is higher with high sentiment levels for Japanese yen and Swiss franc, and it generally strengthens across various currencies during recessions. We validate the robustness and persistent predictive power of the LSTM when integrated with technical indicators. This performance remains consistent across diverse model specifications and is robust to structural breaks.