One of the more recently introduced control systems at Deere is the company’s revolutionary neural net program. James Hall, an engineering Ph. D and manager of Deere’s investment analysis, began using a neural network to manage a $100 million equity portfolio in December 1992. Since its development, the net driven portfolio has consistently beaten the Standard & Poor’s 500 index by at least 3 percent a year, in spite of the high transaction costs associated with an annual stock turnover of 300 to 400 percent. Moreover, the network has yet to make a poor investment. In fact, the one time Hall’s superior decided to reject the network’s investment advice it turned out that the network had been right.
Neural networks are constructed by using software or silicon to imitate the structure of brain cells and the three dimensional lattice of connections among them. These artificial neurons enable the network to assimilate and correlate hundreds of variables at a single time. More importantly, the neural network has the capacity to learn from its past experience. This ability to learn is what initially attracted Deere to the notion of neural nets. Hall was working at the Deere Technical Center to develop farm machines that learn to run themselves, he recalled. We invented a combine that learns to run itself by watching the driver. Eventually the driver gets out and the machine just takes off and does the job on its own. The combine has been running successfully for several years now. What we have done in the investment department is adapt that same software to learn about stocks by watching the market.
The net’s ability to assimilate hundreds of variables at once made it ideal for stock market analysis. At Deere, the network examines a carefully screened selection of approximately 1,200 large to mid cap stocks to determine the most popular investment style (value, growth, cyclical etc). The portfolio controlled by the network tends to contain 80 to 100 of this style. The network then attempts to weight the portfolio toward the hot style. In addition, it examines the future earnings of the stocks it handles according to an analysis of trading styles. Drawing upon 40 indicators fed to it by Deere managers, the net ranks 1,000 equities in terms of future returns. The network then sells off the stock it owns and buys up all the stock with predicted returns above a certain cutoff. Hall and the network have improved Deere’s pension earnings by more than $3 million a year.
The power of the neural net is not unchecked, however. It takes time and money to develop a network. We trained 5,000 neural nets before we found a system that worked well enough to use with real money, hall explained. The total effort is about five person years of programming, with about a year of that devoted to the investment side. In addition neural net works cannot function properly without high quality and abundant data. The more data a net has, the more accurate its advice. At Deere, data comes in the form of stock information and training from financial experts and seasonal traders. The success of the net is thus primarily determined by how well the trainer defines the task at hand and how much he or she knows about stock. Regardless, singular or unexpected events can completely stump a network. For instance, Deere’s neural network would not have known how to respond to the outbreak of the Gulf War without prior training. In the end, the neural net is only a tool – a tool that can learn to help managers control and a tool that must be controlled.