This paper develops a conceptual framework for estimating both nonlinear and time-varying weather impacts from customer-level billing histories. The framework is flexible insofar as multiple model specifications can be derived from it for real-world application. Next, through statistical tests the paper evaluates the performance of 10 progressively complex weather specifications applied to a sample of billing data. These tests reject many of the simple and commonly used weather specifications, demonstrating the importance of accurate weather modeling.¿ 1997 American Geophysical Union