FORECASTING WILDFIRE SUPPRESSION EXPENDITURES FOR THE UNITED STATES FOREST SERVICE

The wildland fire management organization of the United States Forest Service (USFS) operates under policy and budget legacies that began nearly -100 years ago and a forest fuel situation that is all too current. The confluence of these three factors contributes to increased burning and firefighting costs for the agency, and increased concern from both the U.S. Congress and the public. Historically, the 10-year moving average of suppression expenditures has been used in USFS annual budget requests to Congress. But in a time when fire activity and costs are steadily rising, the IO-year moving average budget formula has translated into shortfalls in available suppression funds nearly every year since the mid-1990s. When the budgeted amount is insufficient, the agency continues to suppress fires by reallocating funds from other land management programs and by making subsequent requests to Congress for additional funding. A recent report from the U.S. General Accounting Office (renamed the Government Accountability Office in 2004) recommended a reevaluation of the budgeting system for wildfire suppression expenditures by the federal land management agencies (U.S. GAO, 2004). While many of the issues and critiques made by GAO are beyond the control of the agencies, the USFS has explored alternatives to current practices used in developing out-year budget requests for emergency fire suppression. We have two primary objectives in th is chapter. First, we seek to evaluate candidate forecast models of wildfire suppression expenditures. These time series models are constructed to allow suppression budget forecasts up to 3 years in advance of a coming fire season. These models are evaluated for their suit­ ability for budget documents presented to Congress. The structure of estimated models highlights the importance of accounting for intertemporal dynamics and stochasticity in wildfire suppression expenditures. Second, we demonstrate a method from the forecasting literature that quantifies some of the factors poten­ tially important in choosing among alternative models. The method applies loss functions to errors in forecasts, and our comparisons are between the 10-year moving average and our estimated time series models.