Syntactic pattern based inductive learning and primitive discovery for chart analysis

Abstract

Charts are used for the investment decisions by visualizing the stock prices and trading volumes. From the charts, investors attempt to detect patterns which can help the prediction of future stock prices. Although the charts are useful the human``s interpretation of charts tends to be subject, incomplete, inconsistent, and time consuming. To overcome such difficulties, we have developed a structural pattern recognition scheme. A pattern is represented by pattern primitives and compositional operators. A pattern primitive has several attributes that describe its characteristics. To determine pattern primitives, we have approached it in two ways, one is sensitivity analysis of the duration of the primitive and the other is using the genetic adaptive algorithms. We have set two compositional operators: CONC (concurrent composition) and SEQ (sequent composition). They synthesize primitives to constitute patterns on the time horizon. To synthesize reliable patterns, we have developed an inductive learning algorithm called SYNPLE which generates a set of patterns from the set of primitives with the highest performance within the constraints of syntactical grammars. Performance of a pattern is evaluated from the instances of the pattern. We used three charts in the experiment:the trend line of stock price, the moving average curve of stock price and the moving average curve of trading volume in daily base. The performance of rules generated by the inductive learning are evaluated and we have showed how much the performance is increased after the learning algorithm is applied to primitives. The difference of performance according to what primitive discovery methods - sensitivity of duration and genetic adaptive algorithm - are used is also evaluated.