Note
The following documentation closely follows the work of Dunis et al. (2015)
Warning
In order to use this module, you should additionally install TensorFlow v2.8.0. and Keras v2.3.1. For more details, please visit our ArbitrageLab installation guide.
Spread Modeling
Introduction
In this module, we are following the works of Dr Christian Dunis and co-authors relating to the efficient modeling of a few major commodity spreads. The main attributes of trading spreads are:
Advantages
Less likely to suffer from information shocks
It is important to note that spreads are less likely to suffer from information shocks, as the movements of the two legs will offset each other.
Less likely to be subject to speculative bubbles
(Sweeney 1988) notes that speculative bubbles are a big source of market inefficiency. This effect is less likely to happen in spread markets because any bubble effect will be replicated in the opposing leg of the spread (assuming the two legs are sufficiently correlated). Therefore, the effect of the bubble is largely offset.
Disadvantages
Limited Return Potential
The spreads are hard to trade since they offer a limited potential return because of the muted effect of market inefficiencies. This is also exacerbated by the fact that two sets of transaction costs have to be covered in order to trade a spread.
Transaction costs for multiple legs
A point made by (Butterworth and Holmes 2002) that ‘the overall profitability of the strategy is seriously impaired by the difficulty, which traders face, in liquidating their positions’ indicates a definite need for more discerning trade selection which is solved using an assortment of filters.
In the literature, the initial motivation of the works was to model and forecast the spread as accurately as possible. As time progressed, the literature started focusing more on seeking models/methods that discriminated between large/small moves, so that transaction costs would be minimized.
Papers used in this module
Modelling and trading the gasoline crack spread: A non-linear story (Dunis et al. 2005)
Spread being modelled : Crack Spread
Benchmark Model : Fair Value Non Linear Cointegration Model
Novel Models : MLP, RNN, HONN
Filters : Threshold Filter, Asymmetric Threshold Filter, Correlation Filter
Best Results Out of Sample : HONN with standard threshold filter
Volatility filters for asset management: An application to managed futures (Dunis et al. 2005)
Filters : Time Varying RiskMetrics volatility model
Strategies : No Trade Strategy, Reverse Strategy
Modelling and Trading The Soybean-Oil Crush Spread with Recurrent and Higher Order Networks: A Comparative Analysis (Dunis et al. 2006)
Spread being modelled : Soy Crush Spread
Benchmark Model : Fair Value Cointegration Model
Novel Models : MLP, RNN, HONN
Benchmark Filter : Traditional Threshold Filter
Filters : Correlation Filter
Best Results In sample : MLP with correlation filter
Best Results Out of sample : MLP with correlation filter
Comparison Method : Risk Adjusted Return
Modelling and Trading the EUR/USD Exchange Rate at the ECB Fixing (Dunis et al. 2008)
Spread being modelled : EUR USD Exchange Rate
Benchmark Models : ARMA MACD, Naive
Novel Models : MLP, HONN, Pi Sigma, RNN
Filters : Threshold Filter
Trading and hedging the corn/ethanol crush spread using time-varying leverage and nonlinear models (Dunis et al. 2013)
Spread being modelled : Corn Crush Spread
Novel Modes : MLP, HONN, GPA
Filters : Time Varying Volatility Filter
Research Notebooks
The following research notebooks can be used to better understand the components of the framework described above.
Crack Spread Modeling - showcases the use of the filters and networks on the crack spread.
Fair Value Modeling - showcases the use of the TAR model on the crack spread.