Classifying high-frequency FX rate movements with technical indicators and inception model

Abstract

Recent advances in the adoption of AI to forecast stock price movements highlight (i) new ways to encode financial data and technical indicators conveniently; (ii) a state-of-the-art architecture based on inception networks; and, (iii) the existence of across-asset universal features that can be leveraged to improve performances. We combine these three observations and investigate the extent to which this new pipeline can guarantee good predictive power in FX markets. Ultimately, we wonder if these approaches continue to work in a quote-driven market, wherein the AI cannot rely on the (micro) structure of the limit order book. More precisely, we develop a neural network based model for classifying high-frequency FX price movements. The architecture utilises inception modules to capture useful spatial structures from an image-like matrix that consists of different technical indicators. Gated Recurrent Units (GRU) are also implemented to identify longer time dependencies. We assess the model by testing its out-of-sample classifications on future price movements with ten FX pairs - we show how it outperforms Linear Discriminant Analysis (LDA) model on both accuracy and F1 score. We also found that training a universal model with all FX pairs could further improve the classification performances, which indicates that universal features exist among FX tick data.