Many attempts have been made to use KDD (Knowledge Discovery in Databases) techniques to predict stock performance. Most of these techniques focus on past documented price history, but do not apply many of the more detailed economic fundamentals. Moreover, we are unaware of published work on feature selection from the hundreds of available features. Finally, most work focuses merely on trying to predict the stock price. We take a cue from a recent article [AB00] in which only the level of performance of the stock is estimated (e.g.; high performing or low performing). We believe this simplifies the problem and has enabled us to use features from a database of firm fundamental information for the years 1988 to 2001 as input, identify a set of features, and build models, which have an accuracy of over 80%. None of the techniques we have used here are new; however, we are unaware of published work on the use of either the basic feature selection techniques used or the models that have been built. The fact that we are able to build better-than-random models leads us to believe new algorithms could be developed that will continue to improve effectiveness.
展开▼
