Predicting FX market movements using GAN with limit order event data

Abstract

This study employs generative adversarial network (GAN) models to forecast 5-minute foreign exchange (FX) rate returns. Compared to the Long Short-Term Memory (LSTM) model, GAN demonstrates a significant economic advantage. Notably, the GAN that incorporates limit order events outperforms those that consider liquidity and market order variables. Additionally, the GAN with limit orders achieves tangible economic gains. Consequently, this study provides empirical evidence that adds to the existing literature on market structure regarding informed trading through limit orders.