Commodity Forecasting for Tactical Asset Allocation

This analysis aims at developing forecasting models for both commodity volatility and commodity returns based on monthly data. I show that standard models of realized volatility are able to outperform both GARCH models and the random walk in terms of in-sample explanation power and out-of-sample prediction power. The volatility models that perform best are adapted from applications for other financial assets and different time horizons. These models are so flexible, however, that they can be adapted to forecast volatilities for time horizons important for tactical asset allocations. The return models developed here include seasonal and macroeconomic factors as explanatory variables for commodity returns and adjust for heteroschedasticity. They show similar explanatory power for past data than comparable return models used for equity returns or bond returns. While these models perform much better than random walk or standard time series models in-sample, the outperformance for one month out-of-sample forecasts is only marginally better than for a random walk model. These findings are similar to other asset classes like equity or FX markets. Since the models nevertheless provide some economic insight to what drives commodity returns, they are preferable to simple random walk or time series models. Further improvements to these models could be made using a multivariate environment but the relatively short history of data compared to the increasing number of free parameters of the models render this almost impossible in practice. Thus, the models described here are close to optimal both in terms of ex-post explanation power and ex-ante predictive power. They are easily estimated with standard statistical procedures, combine the latest findings of academic research and are sufficiently reliable to use as indicators for future asset returns and volatilities. Hence, they should provide added value to the development and implementation of a commodity strategy and a diversified strategy including commodities and other asset classes. When implemented to forecast other asset classes as well, the models provide can provide a unified framework to estimate future asset class and portfolio returns and risks and help in asset allocation decisions