Deep learning-based predictive models for forex market trends: Practical implementation and performance evaluation

Abstract

In recent years, there has been growing interest in the prediction of financial market trends, due to its potential applications in the real world. Unlike traditional investment avenues such as the stock market, the foreign exchange (Forex) market revolves around two primary types of orders that correspond with the market's direction: upward and downward. Consequently, forecasting the behaviour of the Forex behaviour market can be simplified into a binary classification problem to streamline its complexity. Despite the significant enhancements and improvements in performance seen in recent proposed predictive models for the forex market, driven by the advancement of deep learning in various domains, it remains imperative to approach these models with careful consideration of best practices and real-world applications. Currently, only a limited number of papers have been dedicated to this area. This article aims to bridge this gap by proposing a practical implementation of deep learning-based predictive models that perform well for real-world trading activities. These predictive mechanisms can help traders in minimising budget losses and anticipate future risks. Furthermore, the paper emphasises the importance of focussing on return profit as the evaluation metric, rather than accuracy. Extensive experimental studies conducted on realistic Yahoo Finance data sets validate the effectiveness of our implemented prediction mechanisms. Furthermore, empirical evidence suggests that employing the use of three-value labels yields superior accuracy performance compared to traditional two-value labels, as it helps reduce the number of orders placed.