Enhancing Forex Market Forecasting with ConvLSTM2D: A Comprehensive Analysis of Spatiotemporal Dependencies and Data Preprocessing Techniques

Abstract

Accurate forecasting in the FX (FX) is critical due to its significant role in global economic activities and the intricate interrelations between currencies. These relationships are inherently nonlinear and spatiotemporal, reflecting complex patterns influenced by economic, political, and social factors. This study leverages Convolutional Long Short-Term Memory Two-Dimensional (ConvLSTM2D), a Deep Learning (DL) model uniquely capable of capturing such dependencies, to forecast multiple currency pairs simultaneously, capturing interdependencies that would be missed in separate analyses. The primary goal is to predict the next hour's price values for multiple currency pairs in a 1-hour timeframe, visualizing FX data as sequences of frames for simultaneous analysis. To optimize prediction accuracy, different combinations of hyperparameters in the preprocessing phase were tested, including hierarchical clustering for data arrangement, frame heights, time steps, and two scaling methods: Value at Risk (VaR) and Rolling Window (RW). To assess the importance of individual hyperparameters and their combinations on forecasting accuracy, Random Forest (RF) was employed, and SHAP (SHapley Additive exPlanations) values were subsequently applied. The study identified the optimal configuration for accurate predictions: hierarchical single arrangement, rolling window scaling (window size of 16), time steps of 8, and frame height of 8. This setup achieved a Mean Absolute Percentage Error (MAPE) of 0.037 on the validation dataset and 0.04 on the test dataset. RF highlighted data arrangement, scaling method, time steps, and frame height as the most critical factors, while SHAP verified their impact on the results, further validating the robustness of the optimal configuration. These findings demonstrate the effectiveness of ConvLSTM2D in enhancing the accuracy of FX predictions and provide new insights into the role of data preprocessing techniques in financial forecasting.