Abstract Our understanding of farm-level decision-making is often constrained by sparse information about the local input and output prices faced by farmers operating under heterogeneous market conditions. We present a flexible, replicable approach that predicts wholesale prices for six staple cereals, common bean, and potato in Tanzania by leveraging relatively sparse price data for multiple locations and time periods. Exploiting spatio-temporal covariance, in which price patterns for one crop inform prices observed in other crops in neighboring locations and/or time periods, allows us to improve the prediction accuracy of prices for any individual crop. Random forest and autoregressive random forest models, trained on a multi-crop dataset, produce high-resolution price surfaces that show important local deviations from coarse regional averages. We discuss how this modeling framework could be used to design relatively low-cost monitoring systems for enabling regularly updated, national-scale price maps that support targeted interventions, ex ante impact assessments, and real-time advisory services for farmers and policymakers.
Madaga et al. (2026) studied this question.