An Improved Multi-Objective Genetic Model for Stock Selection with Domain Knowledge - Technologies and Applications of Artificial Intelligence - page 67

paper, we reset the scores of these portfolios to zero to prevent these solutions from
being selected. In addition, although the traditional crowding distance method choos-
es significant solutions on the same front, these solutions typically lack meaningful
advantage in the sense of investment. In this study we thus propose to assign higher
weights to the portfolios of higher return per unit of risk and come up with the
MOGA method to construct more profitable portfolios.
With the domain knowledge, we thus improve the score calculation of stocks. Our
experimental results show that our method provides significant improvement over the
previous one in terms of the return and risk on investment. The goal of this study is
to shed more light on the complex characteristics of stock selection in order to ad-
vance the current state of computational finance using AI-based methodologies.

2 Related Works

The class of genetic algorithms was first proposed by Holland [6] according to the
principle of natural selection and has been a widely used approach for optimization
that can be used to improve investment decisions. Pereira [15] presented a discussion
of the advantages of using the GA in the complex optimization problems that arise in
financial markets, and showed that the performance can be improved when using
binary encoding in the algorithm. Chiang [3] applied some fundamental analysis indi-
cators and the GA to evaluate the quality of stocks and showed his proposed model is
effective in financial markets. Kim and Han [10] proposed a different GA-based ap-
proach for dimension reduction and the determination of connection weights for
ANNs, in order to predict the price movements of stock index. Orito and Yamazaki
[14] calculated the contribution rate from the portfolio, and combined this with the
weight of appropriate risk to set the fitness value. Lai et al . [11] used a double-stage
GA to select stocks from the Shanghai stock exchange from 2001 to 2004, and found
that this approach was more suitable for financial applications than fuzzy or artificial
neural networks. More recently, Huang et al . [7, 8, 9] proposed two hybrid versions of
the SVM and fuzzy-based GA models for stock selection. Using certain statistical
tests, Huang et al . provided evidence that their models considerably outperformed the
relevant benchmark.
As to the multi-objective genetic approaches, NSGA-II, a fast and elitist multi-
objective genetic algorithm proposed by Deb et al . [4], is one of the most well-known.
This method combines non-dominated sorting and crowding distance. For portfolio
optimization, Hoklie and Zuhal [5] used Markowitz's method [13] and the two objec-
tives of return and risk to solve the problem of multi-objective portfolio construction.
Bermudez et al . [1] presented a fuzzy ranking procedure for the portfolio selection
problem. They used the trapezoidal fuzzy numbers to evaluate return and risk, and
then used non-dominated sorting as well as the risk aversion to compose the optimal
portfolios. Li and Xu [12] proposed a multi-objective portfolio selection model with
fuzzy random returns and quantified the return, risk, and liquidity of a portfolio to
select the best solution.