We provide a comprehensive study of the out-of-sample performance of a wide variety of existing spot interest rate models in density forecasts, which characterizes the full dynamics of interest rates. The models include continuous-time single-factor diffusion models, GARCH models, regime-switching models, jump diffusion models, as well as their various hybrid combinations. A new evaluation method for out-of-sample forecasts is employed to assess these models. We find that it is important, in terms of both in-sample and out-of-sample criteria, to model interest rate volatility clustering. It is also important to model for the conditional mean of the interest rate level. In the absence of regime-switching and jumps, linear and nonlinear drifts are dominated by no drift models in terms of out-of-sample forecasts, although the former give better in-sample fits. When regime-switching and/or jumps are introduced, linear drift specification becomes better than no drift models as well as nonlinear polynomial-type drift models in out-of-sample forecasts.