A New Approach to Forecasting Food Stamp Caseloads

Forecasting the Food Stamp caseload has become increasingly important. One reason is the sheer frequency and magnitude of recent fluctuations. Over a span of less than fifteen years, the FSP caseload has exhibited three prominent turning points. The caseload declined from 22.4 million participants in 1981 to 18.6 million in 1988. Between 1989 and 1994, it rose to an historical high of nearly 27.5 million persons, before plummeting to 17.2 million persons in 2000. Between 2000 and 2004 it rose to 23.9 million persons.1 In the presence of such large-scale fluctuations, long-term forecasts would be an invaluable aid to managing and budgeting the program. Yet even short-run forecasts could serve useful purposes. County-level forecasts could help administrators deploy resources needed to provide services and to process applications and redeterminations. State-level agencies use forecasts to estimate the sample size that they need to satisfy the conditions of the Quality Control program. Yet Food Stamp caseloads are notoriously difficult to predict. Despite an extensive modeling effort, Dynarski et al. (1991) concluded that their model did not yield "highly accurate" forecasts of the Food Stamp caseload, and that "none of the … models would have captured the increase in participation that began in 1989" (p. xi). Turning points were particularly difficult to forecast. The purpose of this paper is to introduce a new method for forecasting Food Stamp caseloads. I refer to the technique as Markov forecasting because it is motivated by results from the theory of Markov chains. I apply the technique to caseload data from California.