Machine Learning Forecasting of EUR/USD Trends with Qualitative Cross-Validation from Institutional Reports

Abstract

This study presents a comparative methodology for forecasting weekly trends in the EUR/USD exchange rate using daily financial data. A first model is calibrated using a Classification and Regression Tree (CART), followed by an enhanced version based on the Random Forest algorithm. The analysis evaluates statistical performance, model robustness, interpretability, and decision-making quality through systematic backtesting.

The machine learning models are trained to classify weekly currency movements into two categories: +1 for an expected rise and -1 for an expected decline. A prediction is made every 5 days to anticipate the market direction for the subsequent 5-day period, using only information available at the time of the prediction (ensuring strict causality).

To complement and validate the quantitative results, a qualitative layer is introduced using a large language model (LLM) to extract directional sentiment from institutional FX research reports. This dual approach enhances the interpretability and contextual relevance of the forecasting framework.

Keywords: EUR/USD, machine learning, Random Forest, CART, forecasting, large language models, sentiment analysis, time series

DOI: 10.7176/EJBM/17-4-06

Publication date: May 30^th 2025