We present four models for predicting temperatures that can be used for pricing weather derivatives. Three of the models have been suggested in previous literature, and we propose another model that uses splines to remove trend and seasonality effects from temperature time series in a flexible way. Using historical temperature data from 35 weather stations across the United States, we test the performance of the models by evaluating virtual heating degree days (HDD) and cooling degree days (CDD) contracts. We find that all models perform better when predicting HDD indices than predicting CDD indices. However, all models based on a daily simulation approach significantly underestimate the variance of the errors.