Stock Data Mining through Fuzzy Genetic Algorithms

Longbing Cao, Chao Luo, Jiarui Ni, Dan Luo, Chengqi Zhang

Abstract

This paper introduces a Fuzzy Genetic Algorithm (FGA) framework designed to uncover hidden relationships between stocks, trading rules, and markets. By blending the global search power of Genetic Algorithms with the nuanced reasoning of Fuzzy Logic, the authors identified high-performing, non-obvious stock pairs. Backtesting on Australian Securities Exchange (ASX) data demonstrated returns exceeding 40%, highlighting the system's ability to find actionable trading opportunities that traditional methods often overlook.

Introduction

In quantitative finance, pairs trading is a cornerstone strategy: identifying two stocks that move together and betting on their convergence when they diverge. However, finding these pairs is increasingly difficult due to high dimensionality and the rigidity of traditional metrics. Standard Genetic Algorithms (GAs) excel at searching large spaces but struggle with the "fuzzy" nature of business requirements, such as liquidity and volatility constraints. This research bridges that gap by creating an FGA system that understands linguistic preferences (e.g., "high return," "low risk") while navigating massive search spaces.

Methodology

The proposed FGA incorporates several novel components to handle financial complexity:

  • Fuzzy Representation: Instead of raw numbers, the system uses the Sharpe ratio and converts it into ten linguistic levels (from 1st to 10th).
  • Fuzzy Aggregation: To find a "rule-stock pair," the system matches a trading rule to a specific stock. As illustrated in the aggregation logic, if a rule is ranked $m$-th and a stock is ranked $n$-th, the resulting pair is ranked $(m+n-1)$-th.
Fuzzy aggregation table for rule-stock pairs.
  • Moment Defuzzification ($\rho$): To resolve overlapping fuzzy categories, the authors use moment defuzzification to calculate a ranking coefficient $\rho$, turning fuzzy overlaps into a crisp number for final selection.
  • Pairs Trading Strategy: A specific workflow calculates correlation and evaluates "actionability" by considering market sectors and liquidity, optimizing distance factors to trigger signals.

Experimental Setup

The study utilized real-world financial data from the ASX "Top 32" stocks. The dataset included historical intraday orderbook transactions from January 1997 to June 2002. An out-of-sample test was conducted on data from January 1, 2002, to June 18, 2002, to validate the algorithm's predictive power in a live-market simulation.

Results and Discussion

The FGA successfully identified 13 "unexpected" pairs that crossed traditional sector boundaries—findings that were surprising to experienced traders who typically focus on intra-industry correlations. In the six-month test period, the strategies achieved a return of over 40% before transaction costs.

Performance metrics including return (%) and signal frequency over testing points.

The system's strength lies in prioritizing "actionability" over pure mathematical correlation, incorporating market dynamics like volatility into the fitness function. The linguistic robustness of the fuzzy logic also helped the algorithm remain stable despite the noise inherent in high-frequency stock data. As shown in the performance plots, the return percentages remained consistently high across various testing points.

Conclusion

While the FGA shows significant promise, real-world mining remains computationally intensive. Future work will focus on implementing parallel genetic algorithms to handle larger datasets and exploring more complex fuzzy-logic combinations. Additionally, future models must account for transaction costs, which can impact the net profitability of high-frequency strategies.

Summary updated

Abstract

Stock data mining such as financial pairs mining is useful for trading supports and market surveillance. Financial pairs mining targets mining pair relationships between financial entities such as stocks and markets. This paper introduces a fuzzy genetic algorithm framework and strategies for discovering pair relationship in stock data such as in high...
BibTeX
@inproceedings{cao2006stock,
  author = {Longbing Cao and Chao Luo and Jiarui Ni and Dan Luo and Chengqi Zhang},
  title = {{Stock Data Mining through Fuzzy Genetic Algorithms}},
  booktitle = {JCIS},
  year = {2006},
  doi = {10.2991/jcis.2006.129},
}
Figure