Abstract
A clear motivation for this paper is the investigation of a correlation filter to improve the return/risk performance of spread trading models. A further motivation for this paper is the extension of trading futures spreads beyond the ‘Fair Value’ type of model used by Butterworth and Holmes (Citation2002). The trading models tested are the following: the cointegration ‘fair value’ approach; reverse moving average (of which the results of the 20-day model are shown here); traditional regression techniques; and Neural Network Regression. Also shown is the effectiveness of two types of filter: a standard filter and a correlation filter on the trading rule returns. Results show that the best model for trading the WTI–Brent spread is the MACD model, which proved to be profitable, both in- and out-of-sample. This is evidenced by out-of-sample annualised returns of 26.35% for the standard filter and 26.15% for the correlation filter (inclusive of transactions costs).
Notes
1 See for example Mackinlay and Ramaswamy (Citation1988), Yadav and Pope (Citation1990), and Chung (Citation1991), among others.
2 Notable exceptions include Billingsley and Chance (Citation1988), Board and Sutcliffe (Citation1996) and Butterworth and Holmes (Citation2003).
3 The cost of carry is the different between the cash and futures price. This is determined by the cost of buying the underlying in the cash market now and holding until futures expiry. Since the cost of storage of both underlying is identical, they will exactly offset each other.
4 This is US$25 as a percentage of the contract price on 01/01/1995.
5 This figure consists of a 0.085% bid ask spread per barrel and a commission fee of 0.03% per lot (1000 barrels), making a ‘round trip’ transactions cost of 0.37%. Both of these figures have been taken from http://www.sucden.co.uk
6 For I(0) variables a better model could be derived from the bounds testing approach of Pesaran et al . (Citation2001).
7 The optimizing parameter for both filters was the net Sharpe Ratio. That is the earning after costs, divided by the standard deviation.
8 This is simply an ARMA(8, 8) with the 3rd auto regressive term and the 4th, 5th and 8th moving average terms removed.
9 The formula of the model was maintained at each forecast horizon rather than re-estimating repeatedly, see Dunis and Laws (Citation2003).