Commodity market behavior in different states of economy

The past decade witnessed wide swings in commodity prices, which has spurred renewed interest in non-fuel commodities. In 2008 Ben Bernanke, then chairman of the US Federal Reserve, identified commodity prices as one of the main outstanding issues in the analysis of inflation. However, interpreting commodity price cycles and providing factor attribution is still a widely unsolved riddle. Our project strives to shed some light on these issues. We plan to examine commodity price forecasting models with a focus on non-fuel commodities such as agricultural commodities, metals and minerals, where predictors include fundamental variables (like production, yields, inventories and weather), macroeconomic variables and financial variables. We aim to systematically compare a large battery of different statistical models, where we also address model uncertainty. In comparing the competing models, we evaluate the forecast performance not only in terms of traditional measures but also in terms of new measures, including, e.g., indicators that assess whether the direction of the price change was correctly forecasted. Under certain conditions these new measures could be economically more relevant than pure statistical measures (based on the difference between the prediction and the realized value). Our main objective is to find out whether the quality of commodity forecasts depends on the state of the economy and what variables are the key players in explaining different commodity classes in different states of the economy. For example, we would like to answer the question whether commodity forecast models provide better predictions in turbulent than in calm times (i.e., in periods of high/low uncertainty). Other examples of different states of the economy which we plan to investigate are recessions/expansions, periods of high/low inflation, periods of high/low interest rates, and periods of different market sentiment (investors attitudes as to anticipated price development in a market).

We examine whether more precise forecasts of commodity prices can be achieved by taking explicitly into account that their dynamics and linkages to their determinants change in different states of the economy. Such different states could be, for example, good or bad economic times, periods of high or low uncertainty, times of high or low interest rates (or inflation), periods of high or low oil prices, etc. In order to evaluate the quality of commodity price forecasts, we use different types of performance measures: traditional measures and more novel measures. While the traditional measures examine whether predictions are as close as possible to the actual realizations, the novel measures look at, for example, whether the direction of a price change was correctly forecasted or whether adverse movements (turning points) were correctly predicted. The latter type of measures appears more relevant than those based on pure forecast accuracy (i.e., the first type of measures) for a large number of applications in policy as well as in business. We study an aggregate commodity price index, namely the S&P Goldman Sachs Commodity Index, as well as price indices from five commodity sectors: energy, industrial metals, precious metals, agriculture, and livestock. As explanatory variables we consider macroeconomic and financial variables, as well as fundamental variables summarizing the particular market forces in the different commodity markets. We examine a non-linear econometric modelling framework, so as to capture the different dynamics in different states of the world (regimes) and compare the performance of forecasts implied by these regime-dependent models with the performance of forecasts obtained using traditional (linear) models where no regime-dependent dynamics are modelled. We find overwhelming evidence that allowing for regime-dependent dynamics leads to improvements in predictive ability for commodity prices for both types of performance measures and for the aggregate commodity index as well as the five sectoral indices. This is true for short-term and longer-term forecast horizons (one month, three, six, and twelve months). Our results suggest that an interesting trade-off appears between the two types of performance measures, which implies that the particular aim of the prediction exercise carried out (and thus the chosen measure) plays an extremely important role in identifying the best set of models to use. For instance, in regime-dependent models where uncertainty (volatility in equity markets) defines the states of the economy, the traditional measures perform better in times of low uncertainty than in times of high uncertainty, while the opposite is the case for the novel measures. While the suggested methodological toolkit has been thoroughly validated for modelling and forecasting commodity prices, it has the potential to be applied equally successfully in other fields of economics and finance.