This paper examines the relative performance of models in the affine term structure family which includes both complete and essential affine models using Korean government bond yield data. Principal component analysis with Korean government bond yield data shows that the first three components of yields explain 97% of the total yield curve variation, and the components can be characterized as "level", "slope", and "curvature." We also estimate all three-factor affine models using a Kalman filter/quasi maximum likelihood (QML) approach. An exhaustive comparison shows that the three-factor essential affine model, A(1) (3)E, in which only one factor affects the instantaneous volatility of short rates but all three factors affect the price of risk, appears to be the best model in Korea. This finding is consistent with results in Dai and Singleton (2002) and Duffee (2002) on US data and in Tang and Xia (2007) on Canadian, German, Japanese, UK and US data.