Modelling, forecasting and trading of commodity spreads



Middleton, Peter
(2014) Modelling, forecasting and trading of commodity spreads. Doctor of Philosophy thesis, University of Liverpool.

[img] PDF
MiddletonPet_Mar2014_18835.pdf - Author Accepted Manuscript
Available under License Creative Commons Attribution No Derivatives.

Download (5MB)

Abstract

Historically, econometric models have been developed to model financial instruments and markets however the vast majority of these ‘traditional’ models have one thing in common, linearity. While this is convenient and sometimes intuitive many linear models fail to fully capture the dynamic and complex nature of financial instruments and markets. More recently, ‘sophisticated’ methodologies have been evolved to accurately capture ‘non-linear’ relationships that exist between financial time series. This rapidly advancing field in quantitative finance is known as Artifical Intelligence. The earliest forms of artificial intelligence are Neural Networks however these have since been developed using more accurate learning algoirthms. Neural networks are also of particular use because of their capability of being able to continually learn as new information is fed into the network. In this research new data is introduced using both fixed and sliding window approaches for training each of the networks. Futhermore, Genetic Programming Algorithms are also highly regarded in the financial industry and have been increasingly applied as an optimisation technique. Therefore, each of the non-linear models are supported by existing research and as a result these methodologies have become practical tools for optimising existing models and predicting future movements in financial assets. In the absence of computational algorithms to rationalise large amounts of data, investors are confronted with a difficult and seemingly impossible task of trying to comprehend large datasets of information. Nevertheless, advancements in computing technology have enabled market participants to benefit from the use of neural networks (NN) and genetic programming (GP) algorithms in order to optimise and identify patterns and trends between explanatory variables and target outputs. This is of particular importance in the agricultural market such as grains, precious metals and other commodities are informationally rich with large amounts of data being readily available to evaluate. Among the first to use neural networks for financial analysis were Rumelhart and McClelland (1986), Lippman (1987), and Medsker et al. (1993). More recently, neural networks and genetic programming algorithms have been extensively applied to the foreign exchange market (Hornik et al., 1989; Lawrenz and Westerhoff, 2003), for credit analysis (Tam and Kiang, 1992), volatility forecasting (Ormoneit and Neuneier, 1996; Donaldson and Kamstra, 1997), option pricing, (Hutchinson et al.,1994), portfolio optimisation (Chang et al., 2000; Lin et al., 2001), to both developed (Swales and Yoon, 1992) and emerging (Kimoto et al., 1990) stock markets, and for optimisation of technical trading rules (Tsai et al.,1999; Neely et al., 2003). The application of non-linear methodologies to futures contracts and inparticular, commodity spread trading, is limited. Trippi and DeSieno (1992) and Kaastra and Boyd (1995), however were among the first to explore and apply neural networks to forecast futures markets. Financial markets and assets are influenced by an array of factors including but not limited to; human behaviour, economic variables, and many other systematic and non-systematic factors . As a result, many academics and practioners have devised numerous approaches and models to explain financial time series such as fundamental analysis, technical analysis and behavioural finance. The purpose of this research however is to identify, forecast and trade daily changes in commodity spreads using a combination of novel nonlinear modeling techniques and performance enhancing trading filters. During the research process, non-linear models such as neural networks and genetic algorithms are used to identify trends in complex and expansive commodity datasets. Each of the methodologies are used to produce predictions for future time periods. In this research forecasts for t+1 horizons are examined. Progressively, each chapter presents an evolution of research in the area of non-linear forecasting to address inefficiencies associated with more traditional neural architectures. In total a collection of five non-linear methodologies are proposed and analysed to trade commodity ‘spreads’. These non-linear methodologies are benchmarked against linear models which include Naïve strategies, Moving Average Convergence Divergence (MACD) strategies, buy and hold strategies, Autoregressive Moving Average (ARMA) models, and Cointegration models. In the final chapter of the research a mixed model approach is employed to include linear outputs from benchmark models as inputs during the training of each neural network. The research includes various adaptations of existing non-linear methodologies such as neural networks and genetic programming. Through historical data input, each non-linear methodology is trained to construct ‘optimal’ trading models. Models are selected to trade commodity spreads using data from Exchange Traded Funds (ETFs) and Futures contracts. In all cases the reader is presented with results from both unfiltered and filtered trading simulations. The aim of this thesis is to benefit both hedgers and speculators who are interested in applying non-linear methodologies to the task of forecasting changes in commodity spreads. By allowing market participants to input numerous explanatory variables, non-linear methodologies such as neural networks and genetic programming algorithms can become a valuable tool for predicting changes in commodity spreads. Empirical evidence reveals that non-linear methodologies are statistically superior compared to existing linear models and they also produce higher risk adjusted returns. Moreover, by including output from linear models in the input dataset to train non-linear models, market participants are also able benefit from a ‘synergy’ of information using a ‘mixed model’ approach. In order to improve trading results the research also offers examples of numerous trading filters which can also be of use to hedgers and speculators. On the whole the research contributes a wealth of knowledge to academic studies as it offers conclusive evidence to support the widespread integration and use of non-linear modelling in the form of artificial intelligence. Empirical results are evaluated by statistical measures as well as financial performance measures which are widely used by financial institutions.

Item Type: Thesis (Doctor of Philosophy)
Additional Information: Date: 2014-03 (completed)
Subjects: ?? HF ??
Divisions: Faculty of Humanities and Social Sciences > School of Management
Depositing User: Symplectic Admin
Date Deposited: 31 Jul 2014 10:22
Last Modified: 16 Dec 2022 04:42
DOI: 10.17638/00018835
Supervisors:
URI: https://livrepository.liverpool.ac.uk/id/eprint/18835