Empirical studies of the modern theories of bond pricing typically choose proxies for the state variables in a rather arbitrary fashion. This paper empirically analyzes the question of the optimal spot rates to use as state variables. Our findings indicate that the four-year spot rate serves as the best proxy in the one-state-variable model. In the case of the two-state-variables model, the six-year rate and eight-month rate are identified as best. Tests of the out-of-sample prediction ability indicate that our model is superior to Macaulay's duration model and alternative proxies for state variables.