Stock Selection Using Rule Induction

• Authors: George H. John;Peter Miller;Randy Kerber
• Venue: IEEE Expert: Intelligent Systems and Their Applications
• Year: 1996

Quantitative investors -- or quants, as they are known on Wall Street--try to make investment decisions using mathematical models that describe the behavior of a stock's future price in terms of presently observable variables. This differs from the traditional style of equity investment, which involves personal scrutiny of available data on a company, including subjective assessments of the company's management and competitive situation. Ultimately, the goal of both types of investors is to make money, and both broadly agree on how to go about doing it. The key difference is that quantitative analysts focus their efforts on defining a precise method by which objective information about a company should be processed to predict future performance, whereas traditional analysts repeatedly process this information themselves for each stock they track.We address quantitative stock selection from the perspective of knowledge discovery in databases and data mining. One class of problems that data mining addresses, of which stock selection is a member, is the supervised learning or prediction problem: given a database of some arbitrary number of records (stocks, in this case), and some distinguished fields that we'd like to be able to predict given the other fields, discover some useful patterns in the database and express these patterns in some language. The language must be sufficiently expressive to let a computer fill in missing values of the distinguished field in future databases of the same type. The language should also be understandable, so that an expert or analyst can verify that the discovered patterns make sense and so that important new knowledge discovered by the system can be put to use by the analyst. A wide range of problems fit into this framework; stock selection, bond rating, credit scoring, and targeted marketing are just a few examples.In the specific stock-selection problem we discuss in this article, our database contains quarterly information on 1,160 to 1,480 companies during the period October 1, 1987, to June 30, 1993. For each company and quarter, we have nearly one hundred different fields of information, such as market capitalization, price-to-earnings ratio, and trend information. The data also includes each stock's return on investment over the following three-month period. To predict these future returns, we employed the Recon system, developed at Lockheed Martin's Artificial Intelligence Center, which can induce classification rules to model the data it is given. Because we are interested in predicting whether or not a stock will exhibit exceptional return, we defined a target concept for Recon to learn, the exceptional concept: we marked stocks with returns in the top 20% in a given quarter as exceptional and the rest as unexceptional. Recon's job was to analyze a historical database and produce rules that would classify present stocks as exceptional or unexceptional future performers.To evaluate Recon's stock-selection performance, we simulated trading a portfolio of the 50 stocks ranked highest by Recon throughout four and a half years of historical data. When we took into account trading costs, Recon's portfolio had a total return of 223% over a four-year period, significantly outperforming the benchmark, which returned 95% over the same period. (As with human fund managers, past performance is no guarantee of future returns.) The performance was not attributable to growth/value or size effects alone. We concluded that rule induction is a valuable tool for stock selection.