The objective of this study is to identify the preferred model structure that enhances the predictability of KTB government bond yields, based on the Dynamic Nelson-Siegel (DNS) model and its variants. We investigate the effects of varying the number of factors (3, 4, or 5), the arbitrage-free restriction, the 1-step or 2-step estimation methods, the shifting endpoints (SE) technique, factor forecasting models such as AR, VAR, or VECM, and the sample types (rolling or expanding). Empirical analysis using KTB yield data on 14 maturities (ranging from 3 months to 20 years) from January 2006 to December 2023 reveals four main findings. First, the DNS and its extended models lead to an overall improvement in predictive power relative to the random walk (RW) model. Second, the VAR model is suitable for shorter maturities, while the VECM model is better for longer maturities. Third, in the short-term prediction of long-term interest rates, it is difficult for the DNS model or its variants to outperform the RW model.
Fourth, the SE technique was found to be relatively less helpful in forecasting the KTB yield curve. In practice, given the difficulty of model estimation, it is useful to extract yield factors using linear regression models such as the NS, NSS, and GNS models, and to predict the yield curve using the VAR model for short and medium maturities and the VECM model for long maturities.