Interest Rate Forecasting with Dynamic Nelson-Siegel Models, Evaluating Traditional and Distribution-Free Methods for Risk Assessment

Sammanfattning

Accurate yield curve forecasts are essential for financial planning and risk management. This thesis evaluated the predictive performance of three variants of the Dynamic Nelson-Siegel (DNS) models: a one-step DNS using the Kalman filter, a two-step DNS with a Vector Autoregressive VAR(1) model, and a simulation-based method using block-bootstrap resampling. To quantify predictive uncertainty, conformal prediction intervals were applied, offering finite-sample coverage guarantees without relying on distributional assumptions. The models were assessed on their forecast accuracy and ability to replicate key statistical properties of the observed curve dynamics, including autocorrelation and distributional moments. The analysis was conducted using daily interest rate data from 2010 to 2024 across three markets: Stockholm Interbank Offered Rate (Stibor), Euro Interbank Offered Rate (Euribor), and the US federal funds rate. The results showed that the block-bootstrap method consistently achieved the lowest forecast error and best replicated the statistical properties. The one-step DNS produced the widest prediction intervals, was most sensitive to market volatility, and struggled to accurately replicate the variance and autocorrelation. The two-step DNS performed reasonably well overall, except for the poor estimation of the higher-order moments. Across all models, the conformal prediction intervals delivered empirical coverage well above the nominal level, with the bootstrap method producing the narrowest and most stable intervals. These findings demonstrate the practical value of combining DNS models with distribution-free uncertainty quantification and offer a comparative view of model strengths and limitations in the context of risk management.