Temporal Association Rule Mining: With Application to US Stock Market

Abstract

A modified framework, that applies temporal association rule mining to financial time series, is proposed in this paper. The top four components stocks (stock price time series, in USD) of Dow Jones Industrial Average (DJIA) in terms of highest daily volume and DJIA (index time series, expressed in points) are used to form the time-series database (TSDB) from 1994 to 2007. The main goal is to generate profitable trades by uncovering hidden knowledge from the TSDB. This hidden knowledge refers to temporal association rules, which represent the repeated relationships between events of the financial time series with timeparameter constraints: sliding time windows. Following an approach similar to Knowledge Discovery in Databases (KDD), the basic idea is to use frequent events to discover significant rules. Then, we propose the Multi-level Intensive Subset Learning (MIST) algorithm and use it to unveil the finer rules within the subset of the corresponding significant rules. Hypothesis testing is later applied to remove rules that are deemed to occur by chance. After which, multi-period portfolio optimization is done to demonstrate the practicality of using the rules in the real world.