Currency Exchange Rate Prediction with Long Short-Term Memory Networks Based on Attention and News Sentiment Analysis

Currency exchange rate prediction is a typical time series prediction problem which has been solved by time-series models, such as Autoregressive Integrated Moving Average (ARIMA), Seasonal ARIMA (SARIMA) as well as machine learning methods, such as Single Layer Perception (SLP) and Long Short-Term Memory (LSTM). In this work, we aim to predict the future currency exchange prices in collaboration with news sentiment analysis. We use the Australian dollar (AUD) against the US dollar as a case study and study the prediction AUD rate for next day, week, two-week, and month. We conduct a comparative study of the proposed attention-based LSTM with typical models, including ARIMA, SARIMA, SLP, and classical LSTM. The numerical results showed that adding sentiment score of the news articles and matching keywords of 'up/increase' can reduce prediction error by at least 15%. For more extended future prediction, the newly trained model is better than the strategy that reuses the next-day model.