Modelling the structure of Australian Wool Auction prices

Abstract

The largest wool exporter in the world is Australia, where wool being a major export is worth over AUD $2 billion per year and constitutes about 17 per cent of all agricultural exports. Most Australian wool is sold by auctions in three regional centres. The prices paid in these auction markets are used by the Australian production and service sectors to identify the quality preferences of the international retail markets and the intermediate processors. One ongoing problem faced by wool growers has been the lack of clear market signals on the relative importance of wool attributes with respect to the price they receive at auction. The goal of our research is to model the structure of Australian wool auction prices. We aim to optimise the information that can be extracted and used by the production and service sectors in producing and distributing the raw wool clip.Most of the previous methods of modelling and predicting wool auction prices employed by the industry have involved multiple-linear regressions. These methods have proven to be inadequate because they have too many assumptions and deficiencies. This has prompted alternative approaches such as neural networks and tree-based regression methods. In this thesis we discuss these alternative approaches. We observe that neural network methods offer good prediction accuracy of price but give minimal understanding of the price driving variables. On the other hand, tree-based regression methods offer good interpretability of the price driving characteristics but do not give good prediction accuracy of price. This motivates a hybrid approach that combines the best of the tree-based methods and neural networks, offering both prediction accuracy and interpretability.Additionally, there also exists a wool specifications problem. Industrial sorting of wool during harvest, and at the start of processing, assembles wool in bins according to the required wool specifications. At present this assembly is done by constraining the range of all specifications in each bin, and having either a very large number of bins, or a large variance of characteristics within each bin. Multiple-linear regression on price does not provide additional useful information that would streamline this process, nor does it assist in delineating the specifications of individual bins.In this thesis we will present a hybrid modular approach combining the interpretability of a regression tree with the prediction accuracy of neural networks. Our procedure was inspired by Breiman and Shang’s idea of a “representer tree” (also known as a “born again tree”) but with two main modifications: 1) we use a much more accurate Neural Network in place of a multiple tree method, and 2) we use our own modified smearing method which involves adding Gaussian noise. Our methodology has not previously been used for wool auction data and the accompanying price prediction problem. The numeric predictions from our method are highly competitive with other methods. Our method also provides an unprecedented level of clarity and interpretability of the price driving variables in the form of tree diagrams, and the tabular form of these trees developed in our research. These are extremely useful for wool growers and other casual observers who may not have a higher level understanding of modelling and mathematics. This method is also highly modular and can be continually extended and improved. We will detail this approach and illustrate it with real data.The more accurate modelling and analysis helps wool growers to better understand the market behaviour. If the important factors are identified, then effective strategies can be developed to maximise return to the growers.In Chapter 1 of this thesis, we present a brief overview of the Australian wool auction market. We then discuss the problems faced by the wool growers and their significance, which motivate our research.In Chapter 2, we define the predictive aspect of the modelling problem and present the data that is available to us for our research. We introduce the assumptions that must be made in order to model the auction data and predict the wool prices.Chapter 3 discusses neural networks and their potential in our wool auction problem. Neural networks are known to give good results in many modern applications resolving industrial problems. As a result of the popularity of such methods and the ongoing development of them, our research partner, the Department of Agriculture and Food, Government of Western Australia, performed a preliminary investigation into neural networks and found them to give satisfactory predictions of wool auction prices. In our Chapter 3, we perform an analysis and assessment of neural networks, specifically, the generalised regression neural networks (GRNN). We look at the strengths and weaknesses of GRNN, and apply them to the wool auction problem and comment on their relevance and usability in our wool problem. We detail the problems we face, and why neural networks alone may not be the best approach for the wool auction problem, thus laying the foundation for the development of our hybrid modular approach in Chapter 5. We also use the numerical prediction results from GRNN as the benchmark in our comparisons of different modelling methods in the rest of this thesis.Chapter 4 details the tree-based regression methods, as an alternate approach to neural networks. In analysing the tree-based methods with our wool auction data, we illustrate the tree methods’ advantages over neural networks, as well as the trade-offs, with our auction data. We also demonstrate how powerful and useful a tree diagram can be to the wool auction problem. And in this Chapter, we improve a typical tree diagram further by introducing our own tabular form of the tree, which can be of immerse use to wool growers. In particular, we can use our tabular form to solve the wool specification problem mentioned earlier, and we incorporate this tabular form as part of a new hybrid methodology in Chapter 5. In Chapter 4 we also consider the ensemble methods such as bootstrap aggregating (bagging) and random forests, and discuss their results. We demonstrate that, the ensemble methods provide higher prediction accuracies than ordinary regression trees by introducing many trees into the model. But this is at the expense of losing the simplicity and clarity of having only a single tree. However, the study of assemble methods do end up providing an excellent idea for our hybrid approach in Chapter 5.Chapter 5 details the new hybrid approach we developed as a result of our work in Chapters 3 and 4 using neural networks and tree-based regression methods. Our hybrid approach combines the two methods with their respective strengths. We apply our new approach to the data, compare the results with our earlier work in neural networks and tree-based regression methods, then discuss the results.Finally, we conclude our thesis with Chapter 6, discussing the potential of our new hybrid approach and the directions of possible future works.