Predicting Exchange Rate with FinBERT-Based Sentiment Analysis of Online News

The area of news-based sentiment is attracting growing attention in finance due to its recognized effect on traders’ sentiment and risk attitudes. Recently, contextualized word embedding models have been increasingly used to detect sentiment in online news. Despite this interest, few studies have considered the finance-specific news sentiment for financial prediction tasks. In this paper, we present a contextualized sentiment analysis model using a state-of-the-art FinBERT language model. We evaluate the performance of the fine-tuned FinBERT sentiment analysis model in the task of EUR/USD currency movement prediction. The results reported here confirm that the fine-tuned FinBERT is more effective than both the existing lexicon-based sentiment approaches and the survey-based Sentix index. In particular, substantial improvement in prediction performance is reported for different prediction horizons, ranging from 1-day-ahead to 180-day-ahead predictions.