Counterfactual explanations for forex prediction using deep learning methods

Abstract

Fund managers and traders in the Forex trading market have increasingly moved to electronification of their trade execution which provides an opportunity and incentive to perform time series forecasting. However, black-box models lack explainability in their decision-making, especially in the domain of the target audience. This study presents a Deep Learning (DL) approach to time series forecasting and thereafter applies a novel counterfactual explanation as an alternative approach to model explainability. Past research has shown the effectiveness of DL Models such as LSTM in performing time series forecasting in a financial context. This study aims to build on foundational DL models to generate ‘What-If’ alternative scenarios based on a desired outcome. The study optimizes different parts in the counterfactual generation process to the context of a Forex data stream. The research has conclusively achieved high accuracy in a three-day forecast of its generated counterfactual scenario with a validity score of 0.8201. This study and its areas of future work have the potential to create an invaluable impact on the workflows of stakeholders in the fund management and trading sectors, thereby making a breakthrough for explainable AI in the world of finance.