Australian wool is a big deal—or more precisely, it’s a big deal made up of lots of little deals. With an annual value of AUD 3 billion, Australia’s wool makes up 70 percent of the world’s raw wool used in clothing. It is still marketed in lots by the traditional mode, and each year more than 450,000 farm lots are sold at open cry auction. It’s a very risky marketplace for both the farmers and the buyers who contract in advance to deliver wool to processors.
Different Wools, Different Prices
Although the industry sets a market indicator of the price for different types of wool, there is a need for a numerical model to estimate probable price for each individual lot using its own measurements. Models based on regression analysis have worked satisfactorily for very common types of wool, but can’t be applied to the many other types that come on the market. Recently Kimbal Curtis, a wool industry specialist with the Western Australia Department of Agriculture, began to use NeuralTools to predict market prices for farm lots of less abundant wool types.
It was an ideal challenge for NeuralTools. Because of detailed market recording, the number of records was large but there were also missing data; prices were dynamic; and the relationship of price to wool characteristics was nonlinear and interactive, as well as being dynamic. Curtis began his model by training NeuralTools on a set of nearly 6,000 records from a six-month period. For this purpose he established independent category variables for such factors as place of sale, date, and qualitative aspects of the wool that affect price. He set as independent numerical variables the measurable physical characteristics of the wool. He then used the NeuralTools feature Best Net Search to determine that the best computational mode was Generalized Regression Neural Networks.
Western Australia Department of Agriculture
Finding Critical Factors and Predicting with Accuracy
After testing his neural net on over 1,500 cases, Curtis evaluated the predictive capability of his emerging model by comparing the model results with real-world prices and by using the Variable Impact Analysis function in NeuralTools to determine which variables exerted the most influence on the model’s accuracy. He was able to pinpoint the diameter of the wool fiber as pivotal and simplify his model by discarding non-influential variables. Curtis then refined his model by using the Live Prediction feature – which updates neural network predictions in real-time as input data changes – to investigate the relationships among price, length of wool staple, and strength of wool staple for various diameters of fiber. A few more similar refinement steps and the result is a spreadsheet model that not only produces reliable predictions of wool prices but allows buyers and sellers to explore the price implications of the independent variables.
“NeuralTools ably dealt with the complexities of the problem,” Curtis reported, “freeing me to concentrate on the relationships it found and to compare these with our experience of the wool market.” And he had high praise for the software’s user-friendliness. “The thing I really value in an analytical package is the ability to use it to solve real problems without the process itself becoming a problem. Once I understood the analytical options and chose the appropriate set for my purpose, NeuralTools delivered.”