Essays in Global Commodity Prices and Realised Volatility

This thesis consists of three substantive chapters and an Introduction and a Conclusion. The first substantive chapter (Chapter 1) examines in whether high frequency financial and speculative variables convey information that improves the monthly predictions of an aggregate measure of commodity prices (S&PGSCI) by comparing their Root Mean Squared Error (RMSE) to that from the usual benchmark AR (1). The Mixed Data Sampling models (MIDAS) allow us to obtain forecasts by keeping variables at their original frequencies and therefore to explore the richness of high frequency data. The evidence suggests that MIDAS models estimated recursively, and their analogous monthly version seem to capture some predictive information contained in the speculative variables described by the agricultural managed money spread positions. The most interesting finding – larger RMSE reductions during the crisis period - is an improvement in prediction accuracy from use of speculative positions. This suggests speculation contains information that helps in forecasting commodity prices. The second substantive chapter (Chapter 2) focuses on the ability to forecast the daily Realised Volatility of the Bloomberg Commodity Index Excess return (BCOM) using an Heterogeneous Autoregressive model (HAR) and competing models that include an Implied Volatility (IV) measure either from the Commodity or US Stock Market. The former uses the IV for at the money call options of the Dow Jones-UBS Commodity Index published by DataStream while the latter uses the US Stock Market VIX. The Realised Volatility is measured by three different proxies, absolute returns and two range-based estimators, one based on Parkinson (1980) and the Rogers and the other on Satchell (1991). Both are constructed with open, close, high and low daily prices. In-sample results for the 28/07/2011 to 31/10/17 period show that the IV measure estimates are small but statistically significant, suggesting the IV is a biased estimator of future Realised Volatility. The models used to obtain the one-day-ahead out of sample forecasts from 03/03/16 to 31/10/17 were estimated dynamically following a rolling window. To compare the forecasting accuracy of the models, their respective Root Mean Squared Error (RMSE) were computed. These show that the HAR specification does a good job in forecasting the Realised Volatility by offering better forecast in comparison with the IV measures and popular benchmark models such as GARCH (1,1), E-GARCH (1,1). The third substantive (Chapter 3) investigates the linear Granger causal relationship between a popular speculative proxy of 'excess speculation' (Working’sT index) and the weekly log realised volatility and log returns of wheat futures prices. It also examines the impact of managed money spreading positions as a novel measure of speculation on wheat futures causality. Following Granger and Vector Autoregressive (VAR) methodology, I estimate bivariate VAR regressions. The findings show there is a statistically significant unidirectional linear causality between speculative measures and both wheat log returns and the log realised volatility proxy - the Rogers and Satchell’s range-price estimators. Interestingly, the direction of causality runs from managed money spreading positions to log volatility and log returns but in the opposite direction for the Working’s T index.