Electricity demand is strongly correlated with temperature and other weather variables. The New York Independent System Operator (NYISO) relies on weather forecasts and other information to produce its own forecasts of electricity demand. While NYISO does not disclose the algorithm they use to forecast load demand, and how it makes use of temperature forecasts, it is possible to investigate these relationships statistically. We examine the relationship between GFS temperature forecasts produced by NOAA, NYISO's forecasts of electricity load, and realized temperatures and electricity demand in New York City and other regions of New York State. Historical hourly power load forecasts and corresponding observational data were collected from NYISO for the period May 2001-May 2008 for all 11 geographical NYISO regions, including a distinct region for New York City, at lead times from zero to four days. GFS temperature forecasts and corresponding hourly temperature observations for that same period for a number of stations throughout New York were obtained through NOAA/NCDC.

We examine first the accuracy of NYISO's load forecasts as predictors of realized electricity demand, at different lead times. Results indicate that NYISO's forecasting system is mis-calibrated, persistently under-estimating load at all lead times. Viewing the sequence of load forecasts as a stochastic process, we find indications that the sequence of forecast updates—the changes to forecast levels made from one lead time to the next—may exhibit serial auto-correlation. If confirmed, this finding would indicate that the NYISO load forecasting system is not making efficient use of all available information, since in principle past forecast updates should not provide information useful in predicting subsequent revisions.

We examine also the volatility of the load forecast updating process, and the degree to which updates in NYISO load forecasts are correlated with updates in GFS temperature forecasts. Preliminary results indicate a strong positive correlation between temperature forecast revisions and load forecast revisions. This finding, while not surprising, suggests a potentially troubling implication: if providers disseminate weather forecasts at lead times longer than can be justified by demonstrated skill (i.e., if providers engage in “overforecasting”), and if users rely on these forecasts for planning, then users may in turn produce excessively volatile forecasts of weather-sensitive economic variables. The present study offers an opportunity to examine in detail this “propagation of overforecasting” in an area of first-order societal importance.