Forecasting Foreign Exchange Rates in Response to Federal Reserve Communication: A Machine Learning Approach

Abstract

The release of Federal Open Market Committee (FOMC) statements has been shown to signi cantly increase volatility in foreign exchange markets, which motivates investors to make predictions about the directional spot movement of foreign exchange rates. This thesis seeks to quantify the content of FOMC communication by using natural language processing algorithms, and train support vector regressions( SVR) to predict the one-day directional spot movement of ve major currency pairs. We rst develop a GARCH(1,1) model which trains on historical currency data to generate volatility predictions. Then, we use Latent Dirichlet Allocation(LDA) to classify the sentences of each FOMC statement into ve unique mandates, and score each sentence according to their sentiment using a lexicon merging the Harvard IV-4 Psychosociological Dictionary with the Loughran and McDonald nancial tonal lists. These sentence level scores are then aggregated to the FOMC statement level, where each statement has a cumulative sentiment score based on the collective sentiment of each composite mandate. Four SVR models are created for prediction: GARCH, LDA, Historical, and LDA+. The GARCH model uses GARCH(1,1) predictions as input, the LDA model uses the sentiment scores as input, the Historical model uses lagged 5, 10, and 15 day relative di erence in percentage values as input, and LDA+ uses GARCH(1,1) predictions with sentiment scores as input. We nd that the LDA generated sentiment scores have a weak correlation with their corresponding oneday returns. Furthermore, we nd evidence that our LDA+ model can be a viable predictor of foreign exchange rate movement on days of FOMC statement releases.