Probabilistic Forecasting of Agricultural Yield

Haselmann Arakawa, Heitor, Systems Engineering - School of Engineering and Applied Science, University of Virginia
Krzysztofowicz, Roman, EN-Eng Sys and Environment, University of Virginia

Forecasting agricultural yield at a local or regional level is of utmost importance to decision makers in the food supply chain sector. Growers must make decisions based on projected yields. For example, they might be interested in selling their production in advance to cover part of their costs or to hedge potential price volatility. In Brazil, these stakeholders rely on public forecasts provided by the Companhia Nacional de Abastecimento (CONAB) and the Instituto Brasileiro de Geografia e Estatística (IBGE). However, the forecasts published by these sources have something in common: they are deterministic and discount or omit the uncertainty associated with their estimates. Moreover, forecasts for the same crop, region, and time may differ from source to source. This research develops a methodology to quantify the uncertainties associated with deterministic forecasts of soybean crop yields in the state of Mato Grosso, Brazil. The theory of Bayesian Processor of Forecasts (BPF) is reviewed and expanded to incorporate a judgmental prior distribution function modeled from the farmers’ assessments. Farmers in Mato Grosso were interviewed and a set of quantiles of yields was assessed for each one. Individual prior distribution functions were modeled using these sets of quantiles and then combined into a single prior distribution function. The deterministic forecasts were collected from reports issued by CONAB and IBGE in October, February, and May annually between 1993 and 2017. The BPF model is able to merge these deterministic forecasts, and produce probabilistic forecasts of the yield. Various BPF models were developed for different lead times and using different prior information. The empirical and simulated results of this study exemplify the advantages of using the BPF theory and provide a guideline on how to apply this methodology to combine prior distribution functions, fuse information from different sources, and produce probabilistic forecasts.

PHD (Doctor of Philosophy)
agricultural yield, probabilistic forecasting, Bayesian forecaster, Bayesian Processor of Forecasts, data modeling, judgmental assessment, expert uncertainty
Sponsoring Agency:
Coordenação de Aperfeiçoamento de Pessoal de Nível Superior (CAPES)
Issued Date: