Forecasting the yield curve for the Eurozone and USA using standard methods and machine learning approach

This paper investigates forecasting yield curve parameters using the Svensson model for the Eurozone and US markets, applying random walk, autoregressive, vector autoregression, and random forest models. Forecast accuracy was assessed using root mean square error over 1-day, 6-day, and 12-day horizons, and performance was compared using the Diebold-Mariano test. Results indicate that the random walk model consistently provided the most accurate forecasts, particularly for the level and slope of the yield curve. Autoregressive and vector autoregression models performed well for short-term forecasts, while random forest showed mixed results with competitive accuracy in certain cases but higher overall errors. The Diebold-Mariano test confirmed the superiority of the random walk model, especially for longer-term forecasts. Overall, random walk is a robust base model for longer-term forecasts, while other models contribute to short-term predictions, with random forest showing potential for further improvement.