| Abstract: | This paper discusses a framework for refining an initial object-level rule base with a rule induction to learn meta-level rules which find a data set applicable to an object-level rule. A rule induction process such as ID3 tries to learn meta-level rules and classifies given training data sets into positive data sets and negative ones. The rule refinement process tries to refine an initial object-level rule base on classified data sets by using four refinement strategies. Unifying these two processes, one can obtain a refined object-level rule base with high performance where a meta-level rule selects a data set applicable to it. In order to evaluate the framework, an experiment on real Japanese stock price data shows that a refined object-level rule base, which comes from the initial object-level rule base for representing Granville's Law, has a performance beyond that of the average stock price. The performance is difficult for human technical analysts in a stock market to achieve. The result implies that the framework could create an anomaly from Granville's Law in a stock market technical analysis. |
