Ending stocks play an important role in decision making by market participants and policy makers. The provision of accurate forecasts of ending stocks is critical as it can timely reflect the market situation and reduce the uncertainty faced by decision makers. Over the years, the USDA and private analysts have been providing ending stocks forecasts. However, few studies have addressed USDA forecasts, and researchers have not investigated analysts’ forecasts so far.

This dissertation focuses on analyzing the ending stocks forecasts issued by these two sources. It contains three essays which gradually delve into the USDA and private analysts’ forecasting behaviors. The first essay advances existing models and analyze USDA forecasts. The proposed model focuses on forecast revisions and retains the link between forecasts and the ending stocks. The essay also introduces an error covariance structure specifically for ending stocks forecasts. The model is estimated using Bayesian Markov Chain Monte Carlo methods. Results show that USDA forecasts are inefficient.

Given this finding, the model is applied to private analysts’ forecasts to find whether analysts can provide “better” forecasts. This part of analysis is performed in the second essay. Unlike USDA forecasts, analysts’ forecasts are often incomplete. Thus, Essay Two first investigates analysts’ forecasts as a group by combining them to create complete forecast data, and then proposes a method of integrating multiple imputation and MCMC estimations to analyze individual analysts’ forecasts. Results show that analysts, as a group, are inefficient in making ending stocks forecasts. Besides, forecasting behavior vary across individual analysts.

Essay Two also finds that analysts and the USDA have similar behavior in forecasting corn and soybeans ending stocks. Thus, it is possible that their forecasts affect each other. Hence, essay three proposes a method to analyze the forecasts from these two sources together, utilizing the overlooked the information in previous studies. Results show that analysts actually forecast the USDA forecasts instead of the ending stocks.

The proposed models and methods are designed for general purposes. Thus, they can be easily extended by including additional features, as well as can be applied to other fixed-events.