A statistical evaluation was conducted of forecasts of electricity demand for New York City issued by the New York Independent System Operator (NYISO). Analysis focused on how NYISO refines its initial predictions progressively as the interval between forecast and observation closes. Analysis reveals that forecast adjustments exhibit a high degree of positive serial correlation: a forecast adjustment in a given direction (e.g., upward) tends to be followed the next day by a subsequent adjustment in the same direction. NYISO's forecast refinements are thus themselves substantially predictable, a pattern inconsistent with efficient forecasting. The results are consistent with NYISO's evident incentives, however, which involve higher penalties for over-forecasts than for under-forecasts. Load forecasters may rationally be seeking to minimize, not a standard statistical measure of error, but their own financial cost of error.