In bet hedging, organisms sacrifice short-term success to reduce the long-term variance in success. Delayed germination is the classic example of bet hedging, in which a fraction of seeds remain dormant as a hedge against the risk of complete reproductive failure. Here, we investigate the adaptive nature of delayed germination as a bet hedging strategy using long-term demographic data on Sonoran Desert winter annual plants. Using stochastic population models, we estimate fitness as a function of delayed germination and identify evolutionarily stable strategies for 12 abundant species in the community. Results indicate that delayed germination meets the criteria as a bet hedging strategy for all species. Density-dependent models, but not density-independent ones, predicted optimal germination strategies that correspond remarkably well with observed patterns. By incorporating naturally occurring variation in seed and seedling dynamics, our results present a rigorous test of bet hedging theory within the relevant environmental context.