On searching the best mode for forex forecasting: bidirectional long short-term memory default mode is not enough

Abstract

Presently, the Forex market has become the world’s largest financial market with more than US$5 trillion daily volume. Therefore, it attracts many researchers to learn its traded currency pairs characteristics and predict their future values. Here, we propose simple three layers Bidirectional long short-term memory (Bi-LSTM) networks for Forex forecasting with four different merge modes. Moreover, the proposed model is also compared to the conventional long short-term memory (LSTM) networks with the same architecture. Five major Forex currency pairs, namely AUD/USD, EUR/USD, GBP/USD, USD/CHF, and USD/JPY, with more than ten years of historical records are considered in this study. It is revealed from the experimental results that among four available merge modes, the concatenation mode as the default merge mode in Bi-LSTM networks is actually the less preferred mode for Forex forecasting (Root mean square error 0.30685517, mean absolute error 0.27442235, mean absolute percentage error 0.827108%). Moreover, Bi-LSTM average mode gets the highest R2 score that could achieve 89.579%. Therefore, the proposed three layers Bi-LSTM networks could provide a baseline result for developing a good trading strategy in Forex forecasting.

Application of Recurrent Neural Network Bi-Long Short-Term Memory, Gated Recurrent Unit and Bi-Gated Recurrent Unit for Forecasting Rupiah Against Dollar (USD) Exchange Rate

“…Foreign Exchange is one of the largest financial markets in the world [9], [12], [19]. Forex is a core component of every country's financial market, and exchange rate stability is essential for macroeconomic stability [7] Currency rates are influenced by several factors, such as the country's economy, politics, society, international situation [9], [23].…”

“…There are several types of RNN architecture, such as Bi-Long Short-Term Memory (Bi-LSTM), Gated Recurrent Unit (GRU), and Bi-Gated Recurrent Unit (Bi-GRU). These three RNN architectures offers good performance for time-series data, Hansun et al conducted research in 2022 [12] that focused on implementing Long Short-Term Memory (LSTM) and Bi-LSTM for forex forecasting with five major currencies of USD exchange rate finds LSTM has less training time than Bi-LSTM. However, Bi-LSTM has greater predictive performance than LSTM models.…”

“…Their findings revealed the LSTM model's proficiency in predicting the Rupiah exchange rate, exhibiting an error margin of 4.765% in day 5 predictions. In a separate study in 2022, Hansun et al identified the superior predictive performance of Bi-LSTM over LSTM models for forecasting forex market movements [12]. Additionally, Islam and Hossain's research in 2021 compared GRU-LSTM hybrid models against LSTM and GRU individually, concluding that GRU outperformed LSTM, while the GRU-LSTM hybrid model surpassed GRU's performance [19].…”