Library
URL http://www.phd.au.edu/proposal/proposal_somsong.html

Ph.D. Programs
Thesis Proposal

Assumption University
Bangkok, Thailand
November 1996

GENETIC FORECASTING ALGORITHM
by
Somsong Chiraphadhanakul


Submitted to
Qualifying Examination Committee
Graduate School of Computer Information Systems
Assumption University, Bangkok, Thailand.
In partial fulfillment of the requirement for

THE DOCTORAL DEGREE

IN

COMPUTER INFORMATION SYSTEM


Approval Page

RESEARCH TITLE	:	GENETIC FORECASTING ALGORITHM
CANDIDATE NAME	:	Ms. Somsong Chiraphadhanakul
ADVISOR NAME	:	Dr. Vichat Avatchanakorn
ACADEMIC YEAR	:	1996


GENETIC FORECASTING ALGORITHMS

Abstract

Recently, genetic algorithms (GAs) are more frequently used in business, scientific, and engineering applications. The increasing popularity of GAs is due to their adaptability and simplicity. In this research, a new forecasting algorithm using features of GAs is proposed. With GAsÒ blind search, the algorithm takes into account every related components that assigned to them. The algorithms combines the powerful search of GAs and their capability to learn about the relationship patterns of past data in order to forecast future values.

The genetic forecasting algorithms consists of two steps : forecasting and learning steps. GAs are used in forecasting step to estimate parameters of the problem domain. Patterns learning is taken into account in the algorithm to captures the patterns relationship of learning data. The effectiveness of the algorithms is examined by applying to two problem domains : commercial banks deposit and bankruptcy prediction. The investigation is performed by computer simulation.


INTRODUCTION

Genetic algorithms (GAs) are search procedures based on natural genetic evolution. Since theoretical developments by J. Holland [1] in 1975, the procedure can be viewed as general purpose optimization methods. Moreover, they are employed in a difficult search, optimization, and machine-learning problems. The application of these search procedures are increasing in a wide range. There are some reasons that why GAs are attractive in application work [2] : GAs can solve hard problems quickly and reliably, easy to interface to existing simulations models, extensible and easy to hybridize. The basic idea of GAs is the genetic program works toward finding better solutions to problems, just like species adapt to the better related environments. Recently, GAs are more frequently used in business, scientific, and engineering applications [3]. The increasing popularity of GAs is due to their adaptability and simplicity.
Forecasting is an important function for planning in various areas such as : control engineering, electric power systems, economics, business strategic development, production, finance, etc. There are various techniques in forecasting based on statistical approach such as : moving average, exponential smoothing, time series, regression, and economic modeling. These traditional techniques some are limited in their scope, and results are frequently low accuracy of prediction [4]. Forecaster always looks for improving the forecasts, that means some alternative forecasting techniques are being examined. In the last decade, an artificial intelligence ( AI ) concepts which includes : knowledge engineering and expert systems, fuzzy systems, neural networks and genetic algorithms are introduced to a task of forecasting.
In this research, a forecasting method using genetic algorithms (GAs) is introduced. With GAsÒ blind search, the algorithm takes into account every component. The algorithm combines the powerful search of GAs and their capacity to learn about the relationship patterns of past data in order to forecast future values. Conceptually, the genetic forecasting algorithm composes of two loops, the genetic forecasting loop, and the pattern learning loop. The genetic forecasting loop aims at minimizing any error between the actual value and the forecast value as well as minimizing average patterns error. The pattern learning loop serves as a means of pattern relationship learning which enhances the predictive power of the algorithm.

PROBLEM DEFINITION

Forecasting is the most important role in doing business. Top management needs forecasts for planning and implementing long-term strategic. Financial management in both private and public organizations are operated under uncertainty or risk, therefore, forecasting is basically needed. The traditional forecasting techniques result in some shortcomings or problems :
1. In multiple regressions method, there are more than one independent variables in the equation , these variables may have high correlation between themselves. Multicollinearity occurs when these variables interfere with each other. The variables with highly correlated will be left out. This may creates a marginal error since the missing components may have significant information for the forecasting process.
2. In time series method, a large data set is required to extrapolate the past behavior and the causal relationships between other factors are not revealed. Thus, in this method, some important indicator may be left out.
3. In artificial neural networks, the problem is that the entrapment in a local minimum which often occurs with a gradient descent algorithm like backpropagation or an energy minimization of the Hopfield network [5] . In addition, neural networks are models on human brain and learn by themselves from the given patterns for their neural models. Consequently, the resulting models can not explain the physical characteristics of the problem domain under study.
To overcome this drawback, a forecasting method using genetic algorithms is introduced. With the use of GAsÒ, search process can avoid the differentiation and can escape from the local minimum. In GAsÒ blind search, the algorithm takes into account every related components . The algorithm combines the powerful search of GAs and their capability to learn about the relationship patterns of past data in order to forecast future values. The objective of this research is to propose a new algorithm in forecasting that can be a guideline for employing various types of models to simulate different agency environments.

FUNDAMENTAL OF SIMPLE GAs

Genetic algorithms (GAs) are search procedures based on natural genetic evolution. The principle behind GAs is essentially Darwinian principles of biological evolution in the sense that within a population, those elements of the population which are stronger or fitter have a better ability to reproduce another generation of elements.
The basic idea of GAs is to maintain a constant-size set of candidate solutions to the current problem. Candidates in the population compete with one another many iterations to gain the best solution to the objective problem, with the control of GAs operators. Theoretical analyses have shown that GAs exploit the knowledge accumulated during a search in a way that they efficiently balance the need to explore new areas of the search space with the need to focus on high performance regions of that space [6].
In simple GAs, a reproduction operator maintain a size of population by a guide of selection mechanism. The search always has new information to explore with the assist from a crossover operator and a mutation operator. Crossover drives information exchange between two random selected parents chromosomes to form a new offspring, which mutation ensures against a premature loss of important information by simply altering the bit at randomly chosen site.

LITERATURE REVIEW

There are various applications using GAs in order to search, optimization, and machine-learning to the problem domains in a wide rang area. Here are some examples which emphasized on using GAs in forecasting task.
Financial forecasting and portfolios management : GAs are used in financial forecasting and money market at a large number. Karjalainen R.E., [7] used GAs to solve optimization problem and find technical trading rules. In addition, several applications presented that GAs can improve a trading system [8, 9]. In the paper of Wittkemper H.G. and Steiner M. [10], they used GAs to optimize and forecasting the systematic risk of stocks. In portfolio management, GAs are used to manage financial portfolios and to optimize portfolio performance [11, 12, 13]. Schwartz E.I. [14] used neural networks trained by GAs to help traders forecast market patterns which was able to earn 25% returns per year investment in the currency market.
Economics : Arifovic J. [15, 16] used GAs as machine-learning to learn correct decision rules in overlapping generation economics, and examine the behavior of the exchange rate and experimental economies .
Production planning : In a dynamic manufacturing environment, GAs are used to find the optimal parameters to predict the requirements of machine and workload assignment and scheduling [17, 18].

GENETIC FORECASTING ALGORITHM

Fig. 1 illustrates a simplified idea of the proposed genetic forecasting algorithm. Conceptually, the algorithm composes of two loops: the genetic forecasting loop, and the pattern learning loop. The genetic forecasting loop aims to minimize an error between the actual value and the forecast value as well as to minimize average patterns error. The pattern learning loop serves as patterns relationship learning to add more predictive power to the algorithm.
The genetic forecasting algorithm uses the following mathematical model to represent the pattern of the problem domain :
y(k+1) = b0 + b1x1(k) + b2x2(k) + ................. + bnxn(k) + we(k), (1)
where y is the output, x1, x2, ....xn are components of the inputs, e is an error, and k is a step.
At any instant step k, the input components x1(k), x2(k),...., and error, e(k) are used to calculate a forecast of the output of the next step, y(k+1), according to the following equation:
y(k+1) = b0 + b1x1(k) + b2x2(k) + , ................... + bnxn(k) + we(k), (2)
where b0, b1,Å. and w, are parameters estimated by using the input data at step kÖ1 and the output data y at step k. The parameters are optimally selected by genetic forecasting loop from a set of chromosomes in the population.
The genetic forecasting loop, the main forecasting function in the proposed algorithm, is illustrated in detail in Fig. 2. The aim of the genetic forecasting loop is to determine parameters b0, b1,.... and w, which are unknown. Genetic forecasting uses chromosomes to represent these parameters at step k. Each chromosome is composed of n+2 orders of parameters concatenating to form a string. A parameter is represented by ten bits. The first bit represents positive/negative sign, while the remaining nine bits represent the parameter value. To decode the nine binary bits to real values, the decoding factor a is used. The range of parameter is limited by a. These decoded parameter values are part of search spaces that GAs have to go around to find an optimum parameter.
The genetic forecasting loop selects the best fit parameter set based on the selection mechanism (to be stated next) of which the fitness function is derived by taking into account a number of patterns in the pattern learning loop. The genetic forecasting loop is deemed satisfactory when the forecasting error is less than a predefined value s.

The pattern learning loop is incorporated into the proposed algorithm to increase the effectiveness of forecasting the result. A set of data (x1, x2, x3, x4, x5 )k, and yk+1, k=1,...m (m is a number of patterns to be learned), are sequentially presented to the genetic forecasting loop to capture the patterns relationship between x at step k, and y at the next step k+1. All patterns are recursively presented until the genetic forecasting loop has learned the m patterns for a predefined number of rounds, R. At this point the optimal parameters are identified. The identified parameters at step m and x at step m, are then used to forecast the

output y for step m+1. The genetic forecasting component determines the optimal yopt(m+1) by the following equation:

yopt(m+1) = b0 + b1x1(m) + b2x2(m) + ÅÅÅ... + bnxn(m) + we(m) (3)
The selection mechanism of the genetic forecasting is based on the evaluation of the fitness function derived from two categories of errors: (i) the percentage error, % error, between the actual and the forecast values at any present pattern k : error = | y(Actual) Ö y(forecast) | (4)
(ii) the average error of all learning patterns : m
pattern [error] = å errorj / m, (5) j=1
where m is a number of patterns being presented to the genetic forecasting loop.

The selection mechanism chooses optimal parameters at step k from a set of chromosomes in the population. It minimizes the following fitness function:

fitness = error + pattern [error] (6)

The optimum identified parameters together with the inputs at step k are used to calculate the forecasting value at step k+1.

CONCLUSION AND RECOMMENDATION

The research proposes the genetic forecasting algorithm that consists of two loops: the genetic forecasting loop and the pattern learning loop. The fitness function of the selection mechanism is defined by minimizing the fitness function in which the errors from both the genetic forecasting and pattern learning loops are taken into account. The effectiveness of the algorithm is examined by applying to two problem domains : commercial banks deposit and bankruptcy prediction. The investigation is be performed by computer simulation.

REFERENCES

[1] J. Holland, Adaptation in Natural and Artificial Systems, Ann Arbor: University of Michigan Press, 1975.

[2] D.E. Goldberg, " Genetic and Evolutionary Algorithms Come of Age" , Communication of ACM, vol. 37,no. 3, March 1994, pp. 113-119.

[3] D.E. Goldberg, Genetic Algorithms in Search Optimization, and Machine Learning, Addison-Wesley, 1989.

[4] Jae K. Shim, Joel G. Siegel, Handbook of Financial Analysis, Forecasting & Modeling, Prentice-Hall, Inc., 1988 .

[5] LiMin Fu, Neural Networks in Computer Intelligence, Mc-Graw-Hill, Inc., 1994

[6] John J. Grefenstette, "Genetic Algorithms", IEEE Expert, vol. 8, no. 5, October 1993, pp. 6-8.

[7] R.E. Karjalainen, " Using Genetic algorithms To Find Technical Trading Rules In Financial Markets" , PhD. Dissertation, University of Pennsylvania, 1994.

[8] Bob Johnstone, " Research & Innovation : Remaking Markets" , Far Eastern Economic Review, vol.156, iss.12, Mar 25, 1993, pp.50.

[9] Murray A Jr Ruggiero, "Breeding a Super Trader (Part 1)" , Futures-Cedar Falls[CMM], vol.26, iss.1, Jan. 1997, pp. 52-54.

[10] Hans-Georg Wittkemper; Manfred Steiner, " Using Neural network to Forecast The Systematic Risk of Stocks ", European Journal of Operational Research, vol.90, iss.3, May 10,1996, pp. 577-588 .

[11] Joe Mullich, " Software Evolves as Ñ Thinking Ò Tool ", Advertising AgeÒs Business Marketing, vol.81, iss.2, Mar. 1996, pp. 10-12 .

[12] Rebecca lewin , " Maximizing Money Management " , Bank Systems & Technology , vol.32, iss.10, Oct. 1995, pp.38-43 .

[13] Karen Corcella, " Market Prediction Turns The Tables On Random Walk ", Wall Street & Technology, vol. 12, iss.13, May 1995, pp.36-42 .

[14] E.I. Schwartz , " Where Neural Networks Are Already At Work ", Business Week, Nov. 2, 1992, pp.136-137 .

[15] Jasmina Arifovic, " The Behavior of The Exchange Rate in The Genetic Algorithm and Experimental Economies", Journal of Political Economy, vol.104, iss3, Jun 1996, pp.510-541.

[16] Jasmina Arifovic, " Genetic Algorithms and Inflationary Economies ", Journal of Monetary Economics, vol.36,iss.1, Aug. 1995, pp.219-243 .

[17] B.Porter; K.L.Mak; Y.S.Wong, " Machine Requirements Planning and Workload Assignment Using Genetic Algorithms", IEEE Proceeding of ICNNÒ 1995, Australia, pp.711-715.

[18] Christian Bierwirth; Herbert Kopfer; Dirk C Mattfeld; Ivo Rixen, " Genetic Algorithm Based Scheduling in a Dynamic Manufacturing Environment ", IEEE Proceeding of ICNNÒ 1995, Australia, pp. 439-443 .

To the begining