Applying aid optimization beyond prediction 

Project Description

Beyond model performance, a critical gap persists in the application of these predictions to policy design. Zambia Multimodal Project examines how AI/ML based poverty prediction outputs (generated at a 1km-by-1km grid cell) from multi and unimodal data can be applied to guide resource optimization. This research leverages actual budget allocation of food aid across the finest geographic boundaries in Zambia. The research translates the prediction into concrete budget reallocation based on poverty reduction objectives. 

Future study ​

Our study plans to further develop a systematic approach to balancing multiple objectives, with weights guided by community insights. For the poverty reduction objective, wealth predictions will be further refined through the initial three steps of this proposal. With respect to the regional parity objective, we plan to formulate a fairness measure for resource optimization as there is no widely accepted standard for measuring regional fairness, analogous to the FGT poverty measures per. Community input on the importance of each objective, elicited during the local validation stage, will guide the weighting in the final optimization. By integrating refined poverty estimates, a new regional fairness metric, and community-informed trade-offs between poverty reduction and geographic parity, we will advance the basic optimization approach discussed. 

​