Are the USDA’s Soybean Ending Stocks Projections Influencing the Market?
by Qianrong Wu, Ryan Goodrich, and Dermot Hayes | 2016
Executive Summary
The U.S. Department of Agriculture (USDA) provides monthly forecasts of the supply and utilization of soybeans, these reports have an impact on market prices. Recent research at ISU suggests that the USDA’s projections are statistically inefficient, with a statistically significant upward bias in the estimate of closing stocks over the last 18 years. The average annual overestimate of stocks is approximately 16%. The USDA’s biased forecasts puts downward pressure on soybeans futures prices with the largest and most significant impact being seen in the July and August contracts. The bias is about 10% for the July contract one year before maturity. This July bias fall off gradually as maturity nears and averages approximately 5% over the entire period. Farmers who presell soybeans using these contracts receive less about 2% to 5% less that they would have had the bias not been an issue.
Establishing Bias USDA Forecasts
The efficiency of USDA soybean ending stocks from the 1985/86 through 2013/14 marketing years was part of a recent study done by Jinzhi Xiao at Iowa State University. Since we are looking to establish the link between the USDA’s bias projections and their influence on the market, and because these results have not yet been made public, we will provide a brief summary of this work here.
Following from the methodology developed by Davies and Lahiri, Xiao breaks forecasting errors into three components: unforecastable shocks, the idiosyncratic errors of the forecaster, and bias. If we denote St as the soybean ending stocks in marketing year t and Ut,n as the USDA forecasts of the ending stocks n months before the end of the marketing year t, we can visualize the forecasting error as following a form such as:
Where the right-hand side includes the following components:
By subtracting the forecasting error from a given month from the forecasting error of its preceding month, a system of equations is developed to perform the statistical tests. Following the visualization, a first difference equation can be thought of as following a form such as:
In addition to the shock, bias, and idiosyncratic error, Xiao introduces a variable to account for whether or not the USDA has failed to take account for all of the information known at the time of forecasting. Since it is impractical to account for all of the information known at any one time, the measurement he uses is the difference between the original forecast and its revision in the next month. Any difference between these two values implies that after posting their monthly forecast, the USDA gained further information that would have had an influence on their forecasted value.
With this addition, he develops the following system of equation for each marketing year, where N represents the forecast furthest away from the marketing year end date:
The coefficient represents the amount the USDA adjusts its forecasts down each month and the coefficient measures how well the USDA utilizes the arrival of new information. The null hypothesis of
implies that Ut,n is an efficient forecast of the end stock St.
To estimate the system of equations, Bayesian Markov Chain Monte Carlo methods are used with priors chosen from existing literature (Gelman (2006)). This method was chosen to accommodate for the heteroskedasticity of the monthly shock variables
and autocorrelation of the equations due to idiosyncratic error variables. For example,
The resulting coefficients for soybeans are significantly different from zero, implying that the USDA estimates of ending stocks are inefficient. He finds that α=1.37%, meaning that the USDA adjusts its forecast down approximately 1.37% each month, or equivalently that the USDA tends to overestimate ending stocks of soybeans. The furthest out forecast of ending stocks for soybean (June-16 months) will overestimates ending stocks by 21.92% (formular below). On an annual basis, the overestimate is 16.44%. The estimate of for soybeans is 51.37%, implying that if the USDA adjusts its forecast up by 1% last month, it will adjust its forecast this month up by 0.5137%. This serial correlation in updates suggests that when the USDA realizes it has over estimates stocks it is slow to let market participants know.
Biases of Future Prices Associated with USDA Soybean Ending Stocks Estimates
With a bias established in the USDA soybean ending stock forecasts, the next question is whether the bias causes soybean prices in the futures market to be less than would otherwise be the case. To explore this possibility, we make use of historical soybean weekly future prices from 2001 to 2015 from the Moore Research Center.
In the results shown below, we provide normalized weekly prices for each year (2001–2015) for each maturity date setting the maturity price equal to 100. Since the futures price must necessarily converge to the cash price at maturity, a significantly positive trendline for the weekly prices would indicate that prices are downwardly biased prior to maturity and that this bias is eliminated as the contracts mature.
The figure shown below summarizes the results. It shows the average change in soybean futures contracts across contract months and years. On average soybean futures markets under estimate the closing price on the same market. The under estimate is greatest one year prior to maturity and gradually disappears as the USDA reduces the size of the carryout.
The summary slide shown above hides enormous year to year and contract o contract variability. These charts are shown next. As the figures show, the slope of the average trendlines for each maturity date are positive, indicating that futures prices are in fact lower than the maturity prices on the same contracts.
Fig.1. Normalized annual weekly soybean future prices till maturity for January, March, May, July, August, September, November soybean future contracts, 2001–2015.
Fig.2. Average of 15 years normalized annual weekly soybean future prices till maturity for January, March, May, July, August, September, November soybean future contracts, 2001–2015.
To establish the statistical significance of the downward bias displayed in the price trends, a t-test was conducted for each of the January, March, May, July, August, September, and November future contracts. The results show that the bias in the July contract is statistically significant (p-value: 0.03). The lack of significance for the other months may come as a surprise given the presence of an upward trend in the average price trendline for each contract maturity date. This is due in part to the noise that is introduced by the inclusion of 2012 in the data.
Reasons for the Bias
It is very unusual to see a consistent bias in a futures market. An investor who simply went long on soybean futures over the past 15 years would have made a substantial profit. This is in direct contradiction to the efficient market hypothesis. It seems highly likely that this bias is due in part to USDA over predicting the size of the carryout. But this does not mean that the USDA forecasters were not doing their best when making these forecasts available. US soybean exports skyrocketed during this period and it is possible that the USDA forecasters and other market participants were taken by surprise. The bulk of the export increase was to China and it is incredibility difficult to predict or even understand this market.
The evidence that the USDA gradually updates the forecast is more troubling. The evidence suggests that once the forecasters realize they are wrong, they spread the adjustment out over several forecasts. This may help reduce market volatility and the attention that is focused on forecasters, but it introduces a distortion into the marketplace that could easily be eliminated.
