This thesis describes a novel method for estimating the yield on agricultural commodities such as soybeans, corn, and wheat. The initial approach relies on an optimal (non-parametric) hyper-plane strategy to determine if a particular geographic region is likely to produce an above (or below) average yield. We evaluate several underlying weather related features in our quest to find the most parsimonious and predictive model. We show the advantages of the approach with historical data on soybeans in the United States over the past 20 years.

We further model weather in a network of sites in a stochastic manner. The methods involved both top-down and bottom-up approaches such as Markov modulated semi-parametric copula techniques, hidden Markov models, non-parametric resampling techniques, and correlated binary simulation techniques. Extensive evaluations of these methods are carried out.

Simulated weather scenarios are utilized in a stochastic optimization method. In particular, a futures/options portfolio of Soybeans November contract is optimized with conditional value at risk constraints.