Forex Price Predictions using Hybrid TCN-LSTM and LSTM-TCN Models

Abstract

Forecasting financial market prices, particularly foreign exchange (forex) rates, remains a substantial difficulty due to the market's inherent unpredictability, intricacy, and turbulent characteristics. By combining the Temporal Convolutional Network (TCN) and Long Short-Term Memory (LSTM) models into a hybrid framework, this study overcomes this difficulty and improves prediction accuracy.  The MinMaxScaler function was used to standardize the input data prior to training, bringing all values into a range between 0 and 1.  An 80% training segment and a 20% testing segment were then separated from the prepared dataset.  We tested two different hybrid architectures, the LSTM-TCN and the TCN-LSTM, with the EUR/USD, AUD/USD, and GBP/USD value pairs.  With uniform parameters applied to both models during training, the Root Mean Squared Error (RMSE) measure was used for all performance evaluations in order to ensure a fair comparison and determine which model was better. The LSTM-TCN architecture proved to be the superior predictor on the testing set. It recorded a lower average RMSE of 0.003911. This result contrasts with the TCN-LSTM model's performance, which yielded a higher average RMSE of 0.004181.