A model to predict the probability of large hail in a thunderstorm using pre-storm environmental parameters is developed with an ensemble decision tree method known as Random Forest. Variables integral to hail forecasting are investigated and selected for inclusion in the model. Using a predefined database of observed atmospheric soundings, the model is tested and found to have positive skill, similar to current operational techniques. Applying the model to North American Regional Reanalysis (NARR) data, a recent climatology of hail threat in the central United States is created and compared to observations. Model statistics, generated by defining probability thresholds to detect large hail and comparing the results to observed hail reports, show that the model has an advantage over other techniques in correctly forecasting hail occurrence. Trends in the statistics are explored and found to have a dependence consistent with spatial and temporal variability in atmospheric variables over the US, and improvements to the model based on these findings are proposed.