Forecasting European Interest Rates : A Comparative Study of Econometric and Machine Learning Models

Sammanfattning

Accurate forecasting of short-term interest rates is crucial for monetary policy analysis, financial stability, and investment decision-making. This thesis investigates the predictive performance of traditional econometric models and modern machine learning approaches in forecasting the European short-term rate (€STR) and the Swedish short-term rate (SWESTR). Using monthly data from January 2002 to February 2025, the study compares parametric models such as ARIMA, Random Walk and Diebold–Li hybrid specifications with non-parametric and semi-parametric machine learning methods, including Support Vector Regression, Random Forests and Bayesian Additive Regression Trees. Forecasts are evaluated across multiple horizons (1, 3, 6, and 12 months) using an expanding window framework that replicates real-time forecasting conditions. The analysis tests two core hypotheses: whether machine learning models outperform traditional approaches at longer horizons, and whether forecasting accuracy is higher for a sovereign central bank than for a supranational institution. The results reject the first hypothesis: Diebold-Mariano tests reveal that no model significantly outperforms the Random Walk for €STR at any horizon, while for SWESTR, ML models only achieve significant gains at the 12-month horizon (and only for point forecasts). McNemar tests show the Random Walk maintains decisive directional superiority across all horizons and both institutions (p < 0.001). The second hypothesis receives strong support: ablation studies confirm macroeconomic fundamentals contain systematically more predictive information for the sovereign Riksbank than for the supranational ECB