EU-bonds yield curve forecast: comparing ARIMA and XGBOOST models

The goal of this dissertation is to model and forecast the yield curve of the bonds issued by the European Union, the EU-Bonds, using 2024 daily yield data for this product. Two models are used and compared for modelling the yield curve: the Nelson-Siegel model and the Svensson model. These models are calibrated daily during the year of 2024. A classic time series econometric model, the ARIMA, and a Machine-Learning model, the XGBoost, are used to forecast the yield curve. The forecast is performed daily for January 2025. The results presented focus on the forecasted yield curves for 1 day, 1 week (7th of January) and 1 month (31st of January) forecast horizons. The MSE, RMSE, MAE, and MAPE are the error metrics used to evaluate the forecasting performance of these models. For the 1 day forecast horizon, the models that present the best results are the Svensson model with the XGBoost forecasted parameters, according to the MSE and RMSE metrics, and the Svensson model with the ARIMA predicted parameters, according to the MAE and MAPE metrics. For the 1 week and 1 month forecast horizons, the models that produce the best results are the Svensson model with the forecasted parameters obtained with ARIMA, according to all error metrics used. It is concluded that the XGBoost performs better in smaller windows of time and ARIMA has a better forecast accuracy for larger forecast horizons.