Forecasting Food Crisis with High Frequency Household Surveys: A Bayesian-Neural Approach

High-frequency household surveys offer a promising data source for early warning in food security, but their operational utility is limited by spatial sparsity and sampling non-representativeness at lower geographic granularity such as the Admin-2 level. In collaboration with ANON (name hidden for anonymity), a large humanitarian organization, we develop early warning tools that address these challenges. We formulate the problem as a two-step probabilistic framework: first, we forecast Admin-2 level crisis prevalence three months ahead using a Beta-Binomial likelihood; second, we assess whether the projected three-month change exceeds operationally relevant thresholds. To this end, we propose BYM-N, a hybrid Bayesian-neural model that combines conditional autoregressive spatial priors with a Directed Acyclic Graph (DAG)-structured neural network to jointly capture spatial dependence and nonlinear covariate effects in data-sparse regions. Evaluated across 967 Admin-2 locations in Cameroon, Nigeria, and Yemen, BYM-N outperforms all comparators in deterioration detection and correctly anticipates approximately 9.3 million additional food-insecure people that historical baselines miss over a 6-month holdout period. BYM-N is designed for integration into ANON's operational early warning infrastructure, enabling earlier anticipatory action across data-sparse fragile states.