2011 Fourth International Conference on Business Intelligence and Financial Engineering

Business Cycle Index Forecasting of Grey Model Optimized by Genetic Algorithm

Huayun Yu Dabin Zhang

Computer Science College of Yangtze University Information Management Department of Huazhong
Hubei Jinzhou, Country Normal University
[email protected] Hubei Wuhan, China
[email protected]

Abstract—Business cycle forecasting is the premise and squares need to obtain the optimal unbiased estimate based
foundation of strategy-making, plan-making and decision- on four assumptions on the basis of the premise, often need
making. According to business cycle index forecasting method, to use the second derivative to verify whether the request
due to small sample data of leading indicator, it is difficult to parameter values to make the objective function reaches its
determine changes in trends of business cycle fluctuation. maximum or minimum. As these defects of least squares
Leading index is forecasted to extend its sample by the gray method, fitness of GM (1,1) model is not always ideal.
model. In order to overcome using least squares method to Therefore, we propose a GAGM (1,1) model to improve the
determine Development Coefficient and Endogenous Grey model parameters based on genetic algorithm. It can make
Action of GM(1,1), an optimization grey model based on
more accurate forecasts for the economy to provide a
Genetic Algorithm is proposed. For the propose of verification
scientific basis for analyzing trends.
the validity of the method, the proposed method is applied to
forecast business cycle index of China macro economy, and II. GM (1,1) MODEL
compared with GM(1,1). The experimental result shows the
feasibility and effectiveness of the method.
A. Grey System
Keywords- business cycle; genetic algorithm; grey model; The Grey System theory, established by Julong Deng in
forecasting 1982, is a new methodology that focused on the study of
problems involving small sample and poor information. It
I. INTRODUCTION deals with uncertain systems with partially known
information through generating, excavating, and extracting
Typically, the behavior of business cycle is described by
constructing a composite index with different economic and useful information from what is available. So, system’
financial indicators. The index includes leading index, the operational behaviors and their laws of evolution can be
coincident index and the lagging index, and to predict the correctly described and effectively monitored. In the natural
peak or turning point in the economy [1,2]. For example, in world, uncertain systems with small samples and poor
the regression models, such as dynamic factor model, these information exist commonly. That fact determines the wide
indicators or indices used to estimate the turning point. In the range of applicability of Grey System theory [7,8]..
probability model, these indicators or indices may be used to
B. GM˄1ˈ1˅Forecasting Model
estimate the probability of a recession [3,4]. The premise of
these models application is that the leading indicators must Generally, grey forecasting is referred to using a single
have sufficient sample, or the model can not computer peak variable first-order grey model, GM(1,1), to carry out
and valley of leading index. Especially, due to different data quantitative prediction. Grey forecasting has three basic
length of composite indicators, it can only select the shortest operations: (1) accumulated generating, (2) inverse
time interval to be composited. We don’t use this index to accumulated generating, and (3) grey modeling. Grey
determine the future economic trends. If we are able to forecasting uses the operation of accumulated generating to
predict the leading index data to make an appropriate build differential equations. The procedures of deriving a
extension of sample length, we can well compensate for this GM(1,1) grey model are as follows:
defect. For leading index prediction with less sample data, Step 1: Suppose an initial sequence with n entries is
some methods, for example, smoothing, ARIMA and other
forecasting methods could not do, however, gray model in X (0) = ( x (0) (1), x (0) (2)," , x (0) (i )," , x (0) (n)) (1)
dealing with "small samples, poor information" data has
obvious advantages [5]. For these reasons, we select the gray
(0)
model GM (1,1) to forecast business cycle index. Where the ith entry, x (i ) , is the time series data at
In GM (1,1) model forecasting, we need to computer time i.
development coefficient and endogenous grey action of (0)
Step2: Based on the initial sequence x , a new
model. Currently the main method is least squares, but recent (1)
studies show that when the design matrix is close to sequence x cab be generated by one time accumulated
degradation, the least squares method is often not good generating operation , which is
enough performance [6]. In addition, the method of least

978-0-7695-4527-1/11 $26.00 © 2011 IEEE 42

41
DOI 10.1109/BIFE.2011.30
 X (1) = ( x (1) (1), x (1) (2)," x (1) (i)," x (1) (n)) (2) x '(0) (k + 1) = x (1) (k + 1) − x (1) ( k )
b (5)
i = (1 − ea )( x (0) (1) − )e− ak
x (1) (i ) = ¦ x (1) (k ), i = 1, 2," , n a
k =1
where, . New data
where,
k = 1, 2," , n .
generated as a monotonous growth curve, increased the
regularity of the original data columns, and weakened the
volatility. III. GAGM(1,1)
Step3: Because the exponentially increasing nature of the
(1) (1)
A. Genetic Algorithm
entry, x (i ) , in the sequence x is similar to the
Genetic Algorithms (GA) is based on the principle of
exponentially increasing nature of the solution for a forest- natural selection and genetic genetics optimization search
order differential equation which is accordingly used to method. Genetic algorithm on a computer simulation of
(1)
express the entry, x (i ) , i.e., biological evolution and genetic operations, does not require
specific knowledge of the object, and continuously
differentiable search space of the object. GA has the ability
dx (1) of global optimization. Some of the conventional
+ ax (1) = b optimization algorithms effectively solve the problem,
dt (3) optimization technologies based on genetic algorithms can
often get a better result, so GA often are used in many
Discrete solution of the equation is: practical problems, such as function optimization, automatic
control, image recognition, machine learning, etc.. General
x (1) (k + 1) = [ x (0) (1) − b / a] ⋅ e − ak + b / a (4) genetic algorithm includes three basic operations:
reproduction, crossover and mutation [9].
Eq.(3) is called a GM (1,1) model, where a, b are model B. GAGM(1,1) Model
parameters. Parameter a is called development coefficient, In the GM (1,1) prediction model, a core step is to solve
which reflects the system's development trend. According to the development coefficient a and the endogenous grey
the gray theory, when a is negative, the greater its absolute action b. When the design matrix is close to the degradation,
value, the faster system development; the contrary, the more the least squares performance is often not good enough. To
slowly. Parameter b is called endogenous grey action, it solve this problem, GAGM (1,1) model based on genetic
changes the relationship between the response data. algorithm is proposed in this paper. Optimization parameters
Step 4: Parameters solution a and b values of GM(1,1) model can be obtained based on
Parameters solution is the key to the gray model. The genetic algorithms. First, we use the least square method to
model parameters can be determined using the least squares computer initial parameters a and b values, and then use
regression genetic algorithm to adjust a and b values within a certain
aˆ = ( BT B )−1 BT Y range. It will not only correct parameters a and b values, but
where, also to avoid defect of time cost when genetic algorithms is
used directly to obtain a and b values.
ª 1 (1) 1º
« − 2 ( x (1) + x (2))
(1) In the genetic algorithm iterative process, each individual
» is represented by the parameters a and b values, fitness
« » function is
« − 1 ( x (1) (2) + x (1) (3)) 1» n
B=« 2 » obj = ¦ ( xi(0) − xi'(0) ) / xi
« " » i =1
« »
« 1 (1) » where xi (0) is original sequence, xi′ (0) is forecasting
« − ( x (n − 1) + x (n)) 1 »
(1)
sequence.
¬ 2 ¼
C. Algorithm Process
Y = [ x (0) (2), x (0) (3)," x (0) (n)]T ( wher e, x(1) (1) = x (0) (1))
Algorithm process of proposed GAGM(1,1) is as follows:
Step 4: Grey Forecasting Step 1:
Using the operation of one time inverse accumulated Initial sequence
generating operation, the predicted vale of the entry at time
X (0) = ( x (0) (1), x (0) (2)," x (0) (n)) can be generated to
k+1, x ( k + 1) , of the original sequence x can be by
(0) (0)

Eq.(5) X (1) = ( x (1) (1), x (1) (2)," x (1) (n)) by one time
accumulated generating operation.
Step 2: creating a differential equation.

42
43
 TABLE I.
dx (1) INITIAL INDEX SEQUENCE
+ ax (1) = b Time 09,9 09,10 09,11 09,12 10,1
dt
a
Step 3: The model parameters initial solution 0 and 0
b Index 98.9 97.6 97.4 98 98.6
can be determined using the least square regression. Time 10,2 10,3 10,4 10,5 10,6
Step 4: Generate initial population of parameters. The Index 99.1 100.1 101.0 102.0 102.6
range of a and b in one individual is near to initial solution Select previous seven data of table to build grey model
a0 and b0 . This method can improve optimization speed and forecast. The data is as following:
X = ( x(1), x (2)," x(n)) = (98.9,97.6,97.4,98,98.6,99.1,100.1)
based on Genetic Algorithm. For example,
a ∈ (0.8a0 ,1.2a0 )( when, a0 ≥ 0 ) or A. Testing of class ratio, feasibility analysis of modeling
a ∈ (1.2a0 , 0.8a0 )( when, a0 < 0 ) ˈ x(k )
Definition class ratio: σ (k ) = , k = 2,3," , n
b ∈ (0.8b0 ,1.2b0 )( when, b0 ≥ 0) or x(k − 1)
where,
b ∈ (1.2b0 , 0.8b0 )( when, b0 < 0) .
Step 5: set goal function and calculate the fitness value.
σ = (σ (0) (2), σ (0) (3), σ (0) (4), σ (0) (5), σ (0) (6), σ (0) (7))
(0)

In order to evaluate the fitness of the individual population, = (1.0133,1.0020, 0.99387, 0.99391, 0.9949, 0.9900)
in this paer, object function is set to distance between initial
when n=7, σ (0) ∈ [0.778800783,1.284025417] , it
sequence and forecasting sequence. The object function is as
following: meets the requirements of direct modeling.
n
Min : obj = ¦ ( xi(0) − xi'(0) ) / xi
B. Modeling forcast
i =1
We create gray prediction model using data sequence,
and use proposed GAGAM (1,1) model to forecast th8, th9
where is original sequence, is forecasting
xi (0) xi′ (0) and th10 values. In the parameter adjustment process, with
the predicted value and actual value of distance decreases,
sequence. the objective function value becomes smaller. The changes
Step 6: Selection, crossover and mutation. Genetic of solution and population mean are shown in Fig.1. Here,
algorithm is a heuristic global optimization methods, so for we create the individual using the binary encoding, code
each different problem, genetic algorithms need to set length Lind is 40, population size of is 40, we select
different parameters. Initial population over many individual using a random sample of the way, the probability
generations of evolution process, individuals with low fitness of one-point crossover Pc is 0.7, and default probability Pm
was generated by the crossover and mutation replaced by is 0.7/Lind. We select discrete variation. Forecasting results
new individuals, and adapt to the high individual was are shown in Table 2.
retained.
Step 7: Set termination conditions. When the genetic TABLE II. GAGM˄1ˈ1˅FORECASTING SEQUENCE
algorithm to achieve a specified algebraic loop algebra, or
the objective function reaches its minimum 0, then terminate Time 09,9 09,10 09,11 09,12 10,1
the genetic manipulation and output parameters a, b on better Index 98.9 96.82 97.41 98 98.59
value; otherwise go to Step 5.
Step 8: grey forecasting. Time 10,2 10,3 10,4 10,5 10,6
Forecasting values are obtained with substitution of a Index 99.19 99.79 100.4 100.81 101.23
and b of
equation
x (1) (k + 1) = [ x (0) (1) − b / a] ⋅ e − ak + b / a , and C. Residual test
ε (k )
Definition 1: Relative error
reduction of equation
x (0) (k + 1) = x (1) (k + 1) − x (1) (k ) .  (0)
x (k ) − x (k )
(0)

IV. BUSINESS CYCLE FORECASTING

ε (k ) = *100% ,
x (0) (k )
In order to verify the effectiveness of the proposed model, 
where, x ( 0 ) ( k ) is a fitted value of grey model based
we select leading index data of macroeconomic from
September 2009 to June 2010 monthly, as shown in Table 1. on x ( 0 ) .
Forecasting performance base on GM (1,1) and GAGM (1,1)
Definition 2: Average relative error ε (avg )
is compared.
1 n
ε (avg ) = ¦ ε (k )
n − 1 k =2

43
44
 where, ε (k ) < 10% ˈ Average error 0.0158 0.0104
0 From these results, we found that GAGM(1,1) is better
Definition 3: accuracy p fitting degree than GM(1,1) for previous seven months
p = (1 − ε (avg )) *100%
0
historical data. More importantly, proposed forecasting
method in April 2010, May, June index of prediction
where, p 0 > 80% . accuracy has improved significantly, which verifies the
ε (k ) = (0, 0.0079, −0.0001, −0.00001, 0.00004, validity of the method. Meanwhile, in this method, initial
population generated on the basis of the results obtained by
−0.0009, −0.0031, 0.0059, 0.0118, 0.0135) least square method, therefore, the optimal parameter based
Therefore, ε ( avg ) = 0.00481 , p 0 = 99.519% ,it meets
on genetic algorithm can converge quickly. It is higher
efficiency than parameter solution based on the direct use of
the residual test. genetic algorithm.
VI. CONCLUSION
Forecasting of Business Cycle based on grey model can
be a good solution to the sample less, and the synthesis of
indicators data inconsistency problems. In the grey
prediction model, determining the development coefficient
and the endogenous grey amount is a key point, which is
greater impact to model prediction accuracy. Traditional GM
(1,1) parameters are obtained by least square method, this
method based on some basic assumption is not always the
most suitable, so the traditional GM (1,1) forecasting
accuracy is limited. Parameter optimization based on Genetic
algorithm in the local context. Proposed GAGM(1,1) model
can greatly improve fit of data, and improve prediction
accuracy for the economy's analysis.
ACKNOWLEDGMENT
This work was partially supported by the National Natural
Science Foundation of China (Grant No.70971052) and The
Central University Basic Scientific Research Foundation of
Figure 1. GAGM˄1ˈ1˅solution process Central China Normal University (No. CCNU09B01003).
REFERENCES
V. COMPARISON BETWEEN GAGM(1,1) AND GM(1,1) [1] -.Y.Yao and Z.B. Zhou, Research on Business Cycle and natural gas
In order to analyze performance of proposed market , Beijing: China Statistics Press㧘2006 (in Chinese)
GAGM(1,1)model., we compare proposed model and [2] D.B. Zhang, L.A.Yu, S.Y. Wanget, al., Neural network methods for
forecasting turning points in economic time series: an asymmetric
traditional GM(1,1) model. The result of comparison is as verification to business cycles, Frontiers of Computer Science in
shown in tableċ. China, vol 4, Feb. 2010, pp.254-262
[3] S.Kaufmann, Measuring business cycle with a dynamic Markov
TABLE III. FORECASTING RESULT OF TWO METNODS switching factor model: An assessment using Bayesian simulation
methods, Econometrics Journal, vol 3 Jan. 2000, pp. 39–65
Time Real GM˄1ˈ1˅ GAGM˄1ˈ1˅ [4] I.Robert, J.Jan, R.Ward, Business cycle indexes: Does a heap of data
data forecast error forecast error help?, Journal of Business Cycle Measurement and Analysis, vol 1,
09,9 98.9 98.9 0 98.9 0 Mar. 2004, pp. 309–336
09,10 97.6 97.16 0.0045 96.82 0.008 [5] Y.F.Wang, Predicting stock price using fuzzy grey prediction system
, Expert System Application, vol 22, Jan. 2002,pp.33-39
09,11 97.4 97.68 0.0029 97.41 0.0001
[6] ;5&KHQ The least squares linear model ,Shanghai: Shanghai
09,12 98 98.2 0.0020 98 0 Science and Technology Press㧘2003(in Chinese)
10,1 98.6 98.72 0.0012 98.59 0.0001 [7] R.H.Liang, Application of grey relation analysis to hydroelectric
10,2 99.1 99.24 0.0014 99.19 0.0009 generation scheduling, Electric Power Energy System, vol 21, May
10,3 100.1 99.77 0.0033 99.79 0.0031 1999,pp.357-364
Average error 0.00255 0.00203 [8] C.T Lin and L.H.Chen, A grey analysis of bank re-decreasing the
10,4 101 100.3 0.0069 100.4 0.0059 required reserve ratio .Journal Grey System, vol 11,1999,pp.119-132
10.5 102 100.42 0.0156 100.81 0.0118 [9] '%=KDQJ *HQHWLFDOJRULWKPVDQG IX]]\ORJLFGHVLJQ RI
WKHLQWHJUDWLRQ+XEHL Hubei Science and Technology Press㧘
10.6 102.6 100.54 0.0205 101.23 0.0135 2010(in Chinese)

44
45