Prediction of Market Demand Based on
AdaBoost_BP Neural Network
Song Li
School of Management,
Hebei University
Baoding, Hebei, China
E-mail: lees3432@sina.com
Jing Wang
School of Management,
Hebei University
Baoding, Hebei, China
E-mail: jingle9496@163.com
Bo Liu
Hebei Software Institute
Baoding, Hebei, China
E-mail: 273383529@qq.com
Abstract—Aiming at the disadvantages of prediction model of
single BP neural network, a prediction model was presented by
combining AdaBoost algorithm and BP neural network for
improving the forecasting accuracy of single BP neural network.
The ensemble BP network based on AdaBoost is used as
intelligent algorithm. Overcoming the instability of single BP
neural network, the proposed models can give more accurate and
stable prediction for the novel conditions. The main influence
factors for prediction of market demand of refrigerator are
analyzing detailed and used as the inputs of proposed prediction
model. The efficiency of the proposed prediction model was
tested by simulation of the market demand statistical data of a
refrigerator enterprise in China. The simulation results have
shown that the higher accuracy is expressed in this proposed
model, and it is applicable to practice.
Keywords-prediction of market demand; BP neural networ;
AdaBoost algorithm
I. INTRODUCTION
The market demand forecasting is one of important
elements in market forecasting, also is the basis for enterprise
to make decision on management. The enterprise market
demand is varying all the time, because of its influenced by
internal environment, marketing environment, substitutes and
consumer preferences, and it is a random variable that the
change is bigger. The market demand forecasting includes
annual forecast and monthly forecast. Under pulling
management of demand-oriented, the marketing plan, logistic
plan and financial planning have been directly influenced by
the prediction of enterprise market demand. The monthly
forecast with higher accuracy can help administers make
correct decisions and work out rational production plan and
inventory control programs, and it can meet the requirement of
economic development to a maximum but also increase
economic benefit.
For now, the prediction methods are the most usually
used method to predict market demand which is the regression
analysis prediction [1], gray prediction [2], Markov prediction
[3] and seasonal index number [4], etc. These prediction
methods have become mature, and they are very useful in
prediction of market demand. But these traditional forecasting
methods are mostly based on mathematical statistics. Their
common feature is to establish a subjective model of history
data sequences, then the market demand are calculated and
predicted by these models. Its forecasting accuracy has poor
performance and can not meet the actual application demand.
The forecasting model has not strong robustness because of
any capacity of adaptation and self-learning.
In recent years, the neural network was introduced into
the prediction of market demand by some scholars, and they
use BP neutral network to predict the market demand [5, 6].
However, BP neural network bears the limitations of local
optimization, slow convergence and low precision. The BP
neural network is not very satisfactory in prediction of small
sample size and more random noise.
The AdaBosst algorithm [7] can enhance the forecasting
accuracy of any weak predictor, and it has been successfully
applied in many machine learning problems. In order to
improve the forecasting accuracy of BP neural network,
overcome the limitations of weights initialization of BP neural
network and the subjective factor of training samples, a
prediction model of AdaBoost_BP neural network was
presented by combining AdaBoost algorithm and BP neural
network in this paper. The efficiency of the proposed
prediction model was tested by simulation of the refrigerators
market demand.
II. T
HE PREDEICTION MODEL OF BP NEURAL NETWORK
The BP neural network is a multi-layer feed forward neural
network which has the ability to process nonlinear and complex
system problems. Its main characteristics are the signal forward
pass and the error back propagation.
Set
,1 , 2 ,
,,,
T
iii ik
Xxx x is a system input of a nonlinear
discrete dynamical system, its system output is
i
y ,
1, 2, ,im . According to Kolmogorov theorem, a three
layers BP neural network can accomplish the mapping
relationships from n-dimensional space to m-dimensional
space. Select the BP neural network with a typical three-layer
structure
k
-p-1. Then a mapping completed by BP neural
network is
1
:
k
fR Ro
, the input of the hidden layer neuron is
as follow
1
k
ij i j
i
Swx
T
¦
, 1, 2, ,jp . (1)
2013 International Conference on Computer Sciences and Applications
978-0-7695-5125-8/13 $26.00 © 2013 IEEE
DOI 10.1109/CSA.2013.77
305