Development of dynamic stock trading system based on fuzzy decision tree

Abstract

In recent years many attempts have been made to predict the behavior of bonds, currencies, stock, stock markets, and other economic markets. These attempts could not build an accurate and efficient stock trading system for prediction of large fuzzy and missing time series databases. In this paper, fuzzy decision tree classifiers (model) have been proposed for dynamic stock market databases. First, we use Bayesian Belief Networks (BBNs) for data cleaning of time series stock market databases. In our algorithm, we present an extraction of both graphical structure and conditional probabilities of a BBN from possible incomplete and unclear information in stock market databases. It is based on a bound and collapse (BC) method. After data cleaning, new Fuzzy Decision Tree (FDT) has been constructed by using weighted fuzzy production rules (WFPR). In WFPR, weight parameter has been assign to each proposition in the antecedent of a fuzzy production rule (FPR) and certainty factor assign (CF) to each rule. Some important variables have been calculated by using certainty factors (e.g. effect of other companies, effect of other stock exchanges, effect of overall world situation, effect of political situation etc) in dynamic stock market. Our proposed new FDT may not be the best generalization due to over-fitting. Therefore, for decision tree pruning, we have introduced Neuro- Pruning by using back-propagation neural networks. In Neuro-Pruning method instead of absolutely removing nodes, we employ a back-propagation neural network to give weights to nodes according to their significance. Our proposed approach will be able to predict dynamic stock market, improve computational efficiency of data mining approaches, and outperforms error-based pruning.