Predicting Forex Rates using Sentiment Analysis on Financial Articles

Sammanfattning: We develop a deep learning model that uses financial news articles and historical exchange rates to predict the rate for the Euro/US Dollar currency pair one hour ahead. From the articles’ titles and bodies we extract sentiment scores using two different transformer based classifiers – one for short text and one for long text. The long-text classifier is built and trained by us. These sentiment scores are used as a proxy for the market participants’ mood. We combine the mood with technical indicators derived from historical exchange rates to make the future prediction. Our long-text sentiment classifier obtains 72% test accuracy, and our predictive Forex model achieves 44% improved test loss compared to a similar model not leveraging sentiment from article bodies. Moreover, the Forex model heavily outperforms a model not using financial articles at all. Using these results we discuss the efficiency of the EUR/USD currency market. It hints at some inefficient tendencies in the market, but more data and further experiments are needed before any conclusion can be drawn.