ForExAI: Time Series Inference and News Article Analysis Reveal Profitable Foreign Exchange Signals

Forecasting foreign exchange rates over long time periods depends on economic fundamentals. Short-term predictions, by contrast, depend largely on emotions, governmental announcements, the flow of capital, and media commentary. This paper proposes a suite of methods, collectively referred to as ForExAI, to predict foreign exchange rates based on time series analysis and news article analysis. The time series analysis is based on classical statistical time series techniques, such as ARIMA, as well as machine learning methods using neural networks. Separately, ForExAI uses Large Language Models to analyze news articles based on two kinds of prompts: (i) expert-written based on econometric considerations, (ii) existing prompts documented in the literature. Our findings on time series of exchange rates indicate that there are signals in the time series that can be captured even by simple methods like ARIMA(1,1,1), as well as novel machine learning methods on the time series of foreign exchange rate trades. Further, an adaptation of the Kelly criterion can increase cumulative profits. Finally, an ensemble approach often delivers slightly lower profit, but also lower volatility, leading to a higher Sharpe ratio. Regarding news article analysis, shorter prompts yield far better results than complex ones derived from expert knowledge. This work shows a method for hyperparameter tuning a collection of models that forecast complex time series, as well as the relative virtues of different versions of the Kelly criterion. The results make a pragmatic contribution as well. Because we measure profits ignoring transaction costs, our work is not directly actionable by traders, but the insights could be useful. In addition, our results point to further areas of research for traders.