The significance of mapping data sets when considering commodity time series and their use in algorithmically-traded portfolios

Abstract

Many econometric analyses of commodity futures over the years have been performed using spot or front month contract prices. Using such daily prices without the consideration of the associated contract traded volumes is slightly erroneous because, in reality, traders will typically trade the ‘most liquid’ contract, that is, the contract with the largest average daily volume (ADV). The reason for this is in order to gain the best price when buying or selling. If this ‘true’ time series is to be considered, a mapping procedure is required to account for the price jumps at the time when a trader trades out of the expiring contract and enters the new front month contract. A key finding was that this effect was significant, irrespective of the size of the price jump, sometimes referred to as basis or roll and also due to the accumulated roll over a number of years corresponding to multiple contracts. It was also found that the mapping procedure has a significant effect on the time series and should hence always be employed if the realistic traded time series is to be considered. Given this phenomenon, algorithmically-traded commodities futures must necessarily employ such time series when creating metrics or considering an econometric analysis. The key findings include the importance of diversification in algorithmically-traded portfolios, utilising the AOM and PSI metrics. The mapping of data sets to create realistic ‘live-traded’ time series was found to be significant, while the optimal day of roll over prior to contract expiry was found to be related to the trading volumes for certain commodities. Other key findings include the causalities and spillovers within the metals sector where various relationships are evident once the results were processed and analysed, both pre and post mapping. Interestingly, the key relationships including bidirectional volatility and shock spillovers between the four key metals existed when the unmapped data was used however, many of the feedbacks within these relationships was lost when the mapped data sets were considered. A significant finding was therefore the consistent differences in findings between mapped and unmapped data sets attributed to the optimisation or favourability of the models (whether econometric or algorithmic). This is due to the unmapped data including roll or basis (which the models are fitted to) taking into account the roll or basis and utilising them in finding relationships between data sets. In the mapped data set (the time series seen by traders) the roll or basis is accounted for and hence the relationships found stand in real-time trading situations. The differences in the results show how the effect of mapping can be significant with unmapped data sets displaying results which will not exist in a real time traded time series.