Forecasted commodity prices are important to the success of many investment strategies. The authors aim to evaluate the extent to which a simple static factor model captures co-movements in commodity prices. They find that commodity prices display a tendency to revert toward two factors, with one factor being very dominant.
What’s Inside?
The authors model co-movements of real commodity prices via a static factor model. They use a principal component method to generate the factors and find that commodity prices tend to revert toward a factor. The speed of reversion, however, is slow. The authors compare the factor model with the “no-change” model and with two simple models that use changes in commodity prices and changes in global industrial production or the US dollar. They conclude that factors have information about changes in commodity prices that is not included in current values of industrial production and exchange rates.
How Is This Research Useful to Practitioners?
Forecasted commodity prices are critical to the success of practitioners. The authors present a simple but powerful model for forecasting commodity prices. Their results indicate that their model is best for the 12-month horizon and for energy prices and worst for metals.
How Did the Authors Conduct This Research?
The authors use a monthly panel of US-dollar-denominated commodity prices deflated by lagged US CPI. They use monthly prices of 13 commodities for the period of 1989–2012 as well as 9 additional commodities (for which 1989 data are not available) for the period of 1996–2012. They calculate the factor loadings for 1-, 3-, and 12-month horizons for the two commodity price panels. All commodities load positively on the first factor. The sign for the loading on the second factor is mixed. The results show that the first factor explains 71% (81%) of the variance of the 1996–2012 (1989–2012) series and the second factor explains 14% (9%) of the variance of the 1996-2012 (1989–2012) series.
Beginning with six-year samples, they estimate a sequence of regressions in which they use a factor to forecast subsequent percentage changes in the prices of each of the commodities in the sample. They find that the forecasts based on the factor models are highly correlated with subsequent price changes in the commodity. The authors compare the factor model with the no-change model and with two simple models that use, respectively, changes in global industrial production and the US dollar. They use root mean-squared prediction error (RMSPE) as the measure of forecast performance.
The authors find that a no-change forecast is optimal when the commodity price follows a random walk and that the factor model, industrial production model, and no-change model are comparable. The exchange rate model performs the worst. The factor model is best for a 12-month horizon and worst for a 1-month horizon. For three-month horizons, the factor, industrial production, and no-change models are comparable in terms of their ability to predict the commodity prices.
Abstractor’s Viewpoint
The results of this study are very robust. The authors use in- and out-of-sample data to develop their conclusions. They acknowledge some limitations in their study. Inventories and such financial market data as futures prices are not used. Macroeconomic variables other than industrial production and exchange rate could have been used. Overall, the study and its results are very useful to the practitioner community.