<p>Forecasting total Supplemental Nutrition Assistance Program (SNAP) spending in the United States is essential for effective budget planning and policy design. We develop panel econometric forecasting models that predict total SNAP spending using a bottom-up approach that models its key components—enrollment, monthly benefits, and administrative expenses—at the state level and aggregates them to the national level. This structure supports conditional forecasts under counterfactual state-level economic and policy conditions. We integrate economic theory, program rules, and empirical insights with a systematic model selection approach that searches over a large universe of candidate specifications and selects those with the most accurate forecasts for each component, based on root mean squared forecast error. We find that more parsimonious specifications consistently yield more accurate forecasts than more complex alternatives across all components. Importantly, for the enrollment component, simpler one-way state fixed effects models with deterministic trends outperform two-way (state and time) fixed effects (TWFE) models. A decomposition analysis suggests that TWFE underperformance largely reflects differences in estimated coefficients relative to models with deterministic trends, rather than uncertainty from projecting time fixed effects into the forecast horizon. Results also indicate that unobserved common factor models improve enrollment forecast accuracy, although the gains are modest relative to our best-performing models. We conclude by offering guidance for applying these models to baseline and conditional forecasting.</p>