Forecasting cotton prices using statistical methods has advantages and disadvantages. The limitations are mainly a matter of data: for structural models, we often don’t have data to model the variables that are really influencing prices. Even if we do have data, our model prediction of price is tied up with the historical set of data that we used to generate the prediction. This can break down when we are in exceptional time periods such as 2010-11.
The advantages of using regression forecasts is that it is a generally acceptable, repeatable, systematic method of measuring the influence of important (or so we hypothesize) variables on cotton prices. For example, we specify that the U.S. cotton farm price is a function of 1) U.S. cotton ending stocks-to-use 2) a dummy variable accounting for the abnormally high price spike of 2010 and 2011, and 3) a trend variable to capture patterns like population growth and technology improvement. Using annual data between 2000 and formal results of this regression are summarized here:
A quick interpretation of this output is as follows. First, this relationship of the specified independent variables does a fairly decent job explaining most (77%) of the variability in the U.S. average farm price of cotton.
Concerning those variables, the results indicate that U.S. cotton farm price move in the opposite direction of U.S. ending stocks-to-use. More specifically, for every 1% increase in U.S. ending stocks, the U.S. farm price of cotton would be expected to weaken by 0.223 cents per pound, or $0.00223 per pound. Both the price spike dummy variable and the trend variable were also significant and positive.
From a price forecasting standpoint, using the USDA July forecasts of stocks-to-use and Chinese imports, this model predicts an average annual U.S. farm price of 72.02 cents per pound for the 2014/15 marketing year. This estimate is more than a few cents above the high end of USDA’s August forecast of 56 to 64 cents per pound. This illustrates the discrepancy that can sometimes occur between strict regression forecasts (like this one) and ad hoc forecasts like USDA’s. The latter may be harder to replicate, but it could be more accurate by incorporating additional information.
The 72.02 cent price forecast is visible by the black line below on the far right side of the chart. The associated risk bounds imply, unhelpfully, that there is a 95% probability of the average U.S. cotton farm price being between 50 cents and 94 cents. That is unfortunately what you get statistically from having such variable prices in recent years.
