Forecasting Stock Trend Based on the Constructed Anomaly-Patterns Based Decision Tree

Abstract

Forecasting and modeling the trend of stock price is a crucial task in the field of financial market. Along with the advantage of more abundant and transparent data in recent years, how to find useful features to predict stock trend is important. Many works predicted stock price through technical indicators, and some of the works combined technical and chip analysis to forecast stock trend. In other words, there may have patterns that have influence on the stock trend. Because before the uptrend or downtrend happening, there may exist anomaly patterns. In this paper, we firstly define the anomaly patterns, and then propose an approach for constructing a classifier using decision tree based on the defined anomaly patterns for stock trend prediction. It first locates the stock trend periods from the given stock prices series. Then, for every stock trend period, it will try to identify the anomaly patterns from a given specific period before the stock trend period. The discovered anomaly patterns associated with the label of the stock trend period is formed an instance. Then, the generated instances are used to construct the anomaly-patterns based decision tree. Experiments one the real datasets were made to reveal the effectiveness of the proposed approach.