International Journal of Database Theory and Application

Vol.9, No.1 (2016), pp.67-76

http://dx.doi.org/10.14257/ijdta.2016.9.1.06

The Combination Forecasting Model of Auto Sales Based on

Seasonal Index and RBF Neural Network

Lihua Yang and Baolin Li*

School of Economics and Management, Hubei University of Automotive
Technology, Shiyan, 442002,China
yanglh168@126.com;libaolin108@163.com

Abstract
To effectively predict auto sales and improve the competitiveness of automotive
enterprise, the characteristics of actual auto sales were analyzed, owing to the seasonal
fluctuations and the nonlinearity of monthly sales, the combination forecasting model
based on seasonal Index and RBF neural network was proposed. The weights of the two
single models were computed using mean absolute percentage error and the sum of
square error respectively, the result shows that mean absolute percentage error is more
effective. Finally, the prediction accuracy of different models was compared based on the
criteria of MAPE and RMSE, and the effectiveness of the method was proved, the
proposed model can take advantage of the strengths of the two single models, the results
indicate that the combination forecasting model suitable for auto sales has high
prediction accuracy, which can provide a certain reference to auto sales forecasting.

Keywords: Combination Forecasting, auto Sales, seasonal Index, RBF

1. Introduction
With the rapid development of automotive industry, the automobile market is playing
more and more important role on the national economic development. The accuracy of
market information is the core competitiveness of auto enterprises, the auto sales
forecasting can affect the scientific of decision-making of managers, so it is very
important to enterprise decision. Therefore, how to improve the sales forecasting accuracy
has become a hot spot for automotive industry, which can help company to improve the
competitiveness of marketing.
In recent years, many scholars have researched the sales forecasting methods. In
generally, the forecasting methods can be generally divided into two categories, namely,
the qualitative forecasting methods and quantitative forecasting methods. The latter
methods include some traditional statistics methods, such as moving average, exponential
smoothing and multiple regression analysis. In fact, auto sales are affected by many
factors, such as economic situation, state policy, the income of the family, and so on.
These complicated factors cause the remarkable fluctuation and non-linear characteristics
of the historic sales data, so some data don’t have trends and display high fluctuation, to
solve the problem, some data mining algorithms are applied to sales forecasting due to the
complexity of sales data.
ANN models have been revealed to be more efficient and effective than many
traditional statistical forecasting models, Despite the reported advantages, it is relatively
more time-consuming for ANN to perform forecasting, a new model which employs both
the extreme learning machine and the traditional statistical methods is proposed, a
comparison with other traditional methods has shown that this ELM fast forecasting
model is quick and effective [1]. To improve the accuracy of fashion retail forecasting, a
two-stage dynamic forecasting model is proposed, which is combined with both long-term
and short-term predictions [2]. Time series forecasting, as an important tool in many

ISSN: 2005-4270 IJDTA

Copyright ⓒ 2016 SERSC
International Journal of Database Theory and Application
Vol.9, No.1 (2016)

decision support systems, has been extensively studied and applied for sales forecasting
over the past few decades, a hybrid forecasting scheme which combines the classic
SARIMA method and wavelet transform is proposed [3]. A new market forecasting
system for tobacco wholesalers in China based on a new developed market demand
forecasting model is presented, this model can help sales company to improve the
forecasting accuracy for annual and monthly market demands significantly [4]. Nonlinear
and non-stationary evolution and dependencies with diverse macroeconomic variables
hinder accurate long-term prediction of the future of automotive sales, a structural
relationship identification methodology is presented to identify the dynamic couplings
among automobile sales and economic indicators, the empirical results suggest that
VECM can significantly improve prediction accuracy of automotive sales for 12-month
ahead prediction compared to the classical time series techniques [5].Using the
combination forecasting method which based on PLS to predict the province’s cigarette
sales of the next year, the results show that the prediction accuracy is good, which could
provide a certain reference to cigarette sales forecasting in a province[6]. Two strategies
for forecasting – using the set of good forecasters in the testing phase versus filtering out
the bad forecasters and using the resulting set of forecasters were compared, experimental
results with over 30 sales series indicate that our heuristics greatly out-perform those such
as the simple mean [7]. An effective multivariate intelligent decision-making model is
developed to provide effective forecasts by integrating a data preparation and
preprocessing module [8]. Support vector regression was applied to the newspaper or
magazines sales forecasting is a superior method [9].
Most researches usually adopt a single model, but the single model has its own
limitation, which can only use effective information of a certain. Recently, combination
forecasting model has gained a lot of attention. The concept of combination model
obtained from different models has been discussed and used in different fields [10].
However, a few studies have been proposed to examine the accuracy of combination
model for auto sales. In order to overcome the inaccuracy of the single forecasting model,
a new combination model is established to increase accuracy of auto sales forecasting, in
which two forecasting models based on seasonal index and radical basis function neural
network are introduced. Actual monthly sales data from January 2010 to December 2013
are used to establish the model, actual monthly sales data from January 2014 to April
2014 are used to test the model. These models are compared according to the MAPE and
RMSE of forecasting results. This new forecasting model is an extension to the
combination prediction method.
The organization of the paper is shown in Figure 1. Firstly, we obtain the related auto
sales data, analyzing the data characteristics; secondly, two single models were
established based on the selection data; thirdly, the weights of combination model were
calculated based on relative error and absolute error respectively, by comparing the
predictive accuracy of the different combination models, the relative error is more
effective in computing the weights; finally, the forecasting results of three models were
compared based on the criteria of MAPE and RMSE, then the optimization model was
proposed.

68 Copyright ⓒ 2016 SERSC

 International Journal of Database Theory and Application
Vol.9, No.1 (2016)

Data selection

Seasonal index model Radical basis function

Forecasting results

Relative error Absolute error

Weight calculation

Combination model

Results comparison based on MAPE and RMSE

Optimization model

Figure 1.The Organization of the Paper

2. Establishment of the Two Single Models

2.1. Data Selection

In this paper, we obtained monthly auto sales data from China association of
automobile manufacturers. To do monthly auto sales forecasting, the models are
constructed based on monthly sales data from January 2010 to December 2013, namely 4
years, a total of 48 samples. The forecasting models are tested using the monthly sales
data, from January 2014 to April 2014, a total of 4 samples. The monthly sales data are
shown in Figure 2.

Figure 2. Monthly Sales Data

Copyright ⓒ 2016 SERSC 69

International Journal of Database Theory and Application
Vol.9, No.1 (2016)

From the above figure, we can see that the series exhibit a remarkable periodic
component. One hand, There is an obvious seasonal increase in January and March each
year, on the other hand, there is a strongly seasonal decrease in February and July. The
time series indicate that seasonal underlying periodicity probably is 12 months. We can
also notice that the seasonal variations appear to grow with the upward series trend, which
implies a multiplicative model is better than an additive model.
We can gains a more quantitative conclusion about the underlying periodicity from the
autocorrelations and partial autocorrelations of the time series, the autocorrelation
function in Figure 3. shows a significant peak at a lag of 12, which suggests the presence
of an annual seasonal component in the data.

Figure 3. Autocorrelation Function

2.2. Seasonal Index Model

Time series can be decomposed into four components: the long term trend, the seasonal
component, the cyclical component and the irregular component, the relations between
the four components are summarized as two models, multiplicative model and additive
model. From the above analysis, we can draw the conclusion that multiplicative seasonal
adjustment model can be applied to auto sales forecasting. The seasonal decomposition
procedure can creates four new variables by analyzing the auto sales original variables,
where SAF is seasonal adjustment factors, representing seasonal variation; SAS is
seasonally adjusted series, representing the original series removing seasonal variations;
STC is smoothed trend-cycle component, which is a smoothed series that shows both
trend and cyclic components; ERR is the residual component of the series.
Using the seasonal decomposition procedure, we have removed the seasonal
component of a periodic time series to produce a series that is more suitable for trend
analysis. Here, the moving average weight is all points equal, the smoothed trend-cycle
series are shown in Figure 4, this series show a clear upward trend in the following figure.
The seasonal factors series are shown in Figure 5, we can see that there are a number of
peaks and troughs.

70 Copyright ⓒ 2016 SERSC

 International Journal of Database Theory and Application
Vol.9, No.1 (2016)

Figure 4. The Smoothed Trend-Cycle Series

Figure 5. Seasonal Factors for Sales

2.3. RBF Model

A radial basis function network is an artificial neural network that uses radial basis
functions as activation functions,which has many advantages, such as excellent fault
tolerance, highly parallel mechanism, and adaptation.
The RBF network is composed of an input layer, hidden layer and an output layer, each
layer contains a number of neurons, the input nodes of RBF neural network send signals
to the hidden layer, and the output node is usually a simple linear function. As input to
output is nonlinear, and the network output is linear to adjustable parameter mapping, the
weight value of the network can be recursively calculated, this structure is known as a
feedforward architecture because the connections in the network flow forward from the
input layer to the output layer without any feedback loops. RBF network is supervised in
the sense that the predicted results can be compared with actual values of the target
variables.
The input layer of the network contains the predictors, and the output layer contains the
responses. From the above analysis, the time series show that seasonal periodicity is 12
months, the variable of nextmonth is selected as the dependent variable, the auto sales of
12 months before nextmonth are selected as independent variables, a numeric variable of
traintest is specified as partitioning variable, monthly sales data from January 2010 to
December 2013 with a positive value on this variable are assigned to the training sample,
monthly sales data from January 2014 to April 2014 with a value of 0 on this variable are
assigned to the testing sample, the best number of hidden units is automatically computed
according to the smallest error in the testing data, the activation function for the hidden

Copyright ⓒ 2016 SERSC 71

International Journal of Database Theory and Application
Vol.9, No.1 (2016)

layer is the normalized radial basis function, activation function for the output layer is
identity function.
The processing results show that 36 cases were assigned to the training data set, 4 cases
to the testing data set. The number of units in the input layer is the number of independent
variables, there are 12 neurons in the input layer, likewise, there is 1 unit in the output
layer, and there are 3 units in the hidden layer.

3. Establishment of combination Forecasting Model

3.1. Calculation of the Weight
The study of combination forecasting has become the hot topic in the forecasting field,
it was first proposed by Bates and Granger in 1969, they advocated the use of weighted
average in the combination forecasting in order to make use of information as much as
possible. Information is different to be used in each single forecasting model, and its
accuracy is also different, the combination model can take advantage of useful
information of different single model, improving forecasting accuracy. How to allocate
appropriate weight to single model is the core step of constructing combination model,
now the combination forecasting model is generally established based on the criteria of
the sum of absolute error, but some researchers point out that the relative error may be
more accurate than the absolute error [9]. Considering the forecasting accuracy of relative
error, here, the MAPE is proposed to compute the weight of the single model, MAPE
represents mean absolute percentage error.
In the combination model, the sum of the weights is constrained to be equal to one, and
the value of the weight cannot be less than zero. The principle of calculating weight is that
the model of the smaller mean absolute percentage error should have a higher weight, and
the model of the higher error of mean absolute percentage should have a smaller weight.
Expressed in a formula, it can be written as follows.
m
Wi  M i1 /  M i1 , i  1,2,, m
i 1

Where m is the number of model,

Wi is the weight of the ith model, M i is the mean
M
absolute percentage error of the ith model. In the traditional inverse variance method, i
represents the sum of square error of the ith model, the formula is as follows.
n
M i   ( yt  yt ) 2
t 1

The combination forecasting model formula is expressed as followings:

Y  W1Y1  W2Y2

Where Y1 is the predictive value of seasonal index model, Y2 is the predictive value of
Y
RBF neural network. is the predictive value of combination forecasting model.

3.2. Combination Forecasting

The procedure of establishing the combination model is shown as follows.
1) Firstly, the seasonal index model and RBF neural network are established
respectively to predict the monthly auto sales data. The predictive results and relative
errors of the two models are shown in Table 1.

72 Copyright ⓒ 2016 SERSC

 International Journal of Database Theory and Application
Vol.9, No.1 (2016)

Table 1. The Forecasting Results of Two Single Models

Actual Seasonal relative RBF relative
Date_
sales index error network error
JAN 2014 215.64 232.58 0.079 200.19 0.072
FEB 2014 159.64 161.96 0.015 200.11 0.253
MAR 2014 216.91 231.93 0.069 200.26 0.077
APR 2014 200.42 203.84 0.017 200.19 0.001

Here, the measuring unit of auto sales is ten thousand, the formula of relative error is as
follows.
yt  yt
RE 
yt
y y
Where t denotes the predicted value, t denotes the actual value.
From the Table 1, we can see that the two kinds of model forecasting results were
different, because the single model can take advantage of different information of the
data.
2) Secondly, after computing the forecasting results of two single models, the weights
of two single models in the combination model can be calculated according to the testing
data set. On the one hand, the mean absolute percentage error is proposed to calculate the
weight of the single model, the weight of the seasonal index model is 0.692, and the
weight of the RBF neural network is 0.308, on the other hand, the sum of square error is
used to compute the weight of the single model, the weight of the seasonal index model is
0.803, and the weight of the RBF neural network is 0.179. The different weights are
shown in Table 2.

Table 2. The Comparison of the Weights

The mean absolute The sum of square
Model
percentage error weight error weight

Seasonal index 0.692 0.803

RBF network 0.308 0.197

3) Thirdly, based on the above weights, we can respectively calculate the forecasting
result of combination model, the first combination model use MAPE to calculate the
weights, the formula is as follows.
Y  0.692Y1  0.308Y2 (5)
The second combination model use the sum of square error to calculate the weights, the
corresponding formula is as follows,
Y  0.803Y1  0.197Y2 (6)
The results of different weights are shown in Table 3.

Table 3. The Forecasting Results of Different Combination Model

Actual Combination relative Combination relative
Date_
sales Model 1 error Model 2 error
JAN
215.64 222.61 0.032 226.19 0.049
2014
FEB
159.64 173.71 0.088 169.49 0.062
2014
MAR
216.91 222.17 0.024 225.68 0.040
2014

Copyright ⓒ 2016 SERSC 73

International Journal of Database Theory and Application
Vol.9, No.1 (2016)

APR
200.42 202.71 0.011 203.12 0.013
2014

By calculation, the MAPE of the combination model 1 is 3.90, and the MAPE of the
combination model 2 is 4.11, the former is lower than the latter, the result indicates that
relative error is prior to absolute error in the criteria of calculating weight, the MAPE is
more effective.
4) Finally, by comparing the forecasting results of the combination model with the two
single models, we can draw the conclusion that the mean absolute percentage error and
the root mean square error of the combination model are lower than the two single
models. In this paper, the MAPE and RMSE are regarded as the criteria for comparing the
forecasting precision of different models,
The related formulae are expressed as followings:
100 n yt  yt
MAPE  
n t 1 yt

1 n
RMSE  
n t 1
( yt  y t ) 2

y
Where t represents the predicted value of the t time period,
yt represents the actual
value of the t time period, n is the total number of observations.
The results are shown in Table 4.

Table 4. The Comparison of Different Models

Criteria Seasonal index RBF network Combination model
MAPE 4.48 10.08 3.90
RMSE 11.51 23.20 8.36

From the Table 4, we can conclude that the precision of the combination model is
higher than that of two models in terms of the MAPE and RMSE. Because the
combination forecasting model can take into account the more factors and make use of
useful information of every single model, so we obtain a more accuracy forecasting result
which provide more scientific decision-making. The value of MAPE and RMSE for
seasonal index is between RBF network model and combination model, so the seasonal
index model has the better accuracy than RBF neural network.

4. Conclusion
In this paper, we summarized the achievements of previous studies on forecasting, a
new combination forecasting model based on seasonal index model and RBF neural
network is proposed, and a new weight calculation method based on mean absolute
percentage error is also introduced. The conclusions are described as follows.
1) Two models of seasonal index model and RBF neural network model are used to
predict the monthly auto sales, and the results indicate that seasonal index model is better
than RBF model in terms of the mean absolute percentage error, different models can
achieve different accuracy for forecasting monthly sales due to using the different
information from the data.
2) The weights of two models are calculated respectively according to the mean
absolute percentage error and the sum of square error. Forecasting results of two different

74 Copyright ⓒ 2016 SERSC

 International Journal of Database Theory and Application
Vol.9, No.1 (2016)

weights combination models are compared, results indicate that the weights of two single
models can be appropriately assigned based on MAPE, the proposed relative error method
is better than absolute error method on the calculation of weights.
3) The results of three models are compared in terms of mean absolute percentage error
and root mean square error, the MAPE and RMSE values of combination model are
smallest, the corresponding values of RBF neural network are largest, so the new
combination model based on seasonal index and RBF neural network can improve the
forecasting accuracy, We believe that the proposed combination model is highly
applicable for forecasting sales in the auto industry.

Acknowledgements
This paper is granted by Special Research Foundation of Young Teachers of Hubei
University of Automotive Technology (No. 2012XQ02).The authors would like to thank
the editors and the anonymous reviewers for their helpful and constructive comments.

References
[1] Y. Yu, T. M. Choi and C. L. Hu, ―An intelligent fast sales forecasting model for fashion products‖,
Expert Systems With Applications, vol. 38, no. 6, (2010), pp. 376-385.
[2] Y. R. Ni and F. Y. Fan, ―A two-stage dynamic sales forecasting model for the fashion retail‖, Expert
Systems With Applications, vol. 38, no. 3, (2010), pp. 1529-1536.
[3] T. M. Choi, Y. Yu and K. F. Au, ―A hybrid SARIMA wavelet transform method for sales forecasting‖,
Decision Support Systems, vol. 51, no. 1, (2011), pp. 130-140.
[4] D. J. Fang and W. B. Weng, ―Sales Forecasting System for Chinese Tobacco Wholesalers‖, in Proc.
International Conference on Environmental Sciences, China, (2011), pp. 380-386.
[5] A. Sa-ngasoongsong, S. T. S. Bukkapatnam and J. Kim, ―Multi-step sales forecasting in automotive
industry based on structural relationship identification‖, International Journal of Production Economics
vol. 140, no. 2, (2012), pp. 875-887.
[6] B. Luo, L. Wang and W. W. Yan, ―A New Type of Combination Forecasting Method Based on PLS—
The Application of It in Cigarette Sales Forecasting‖, American Journal of Operations Research, vol. 2,
no. 3, (2012), pp. 408-416.
[7] V. Murlidharan and B. Menezes, ―Frequent pattern mining-based sales forecasting‖, OPSEARCH, vol.
50, no. 4, (2013), pp. 455-474.
[8] Z. X. Guo, W. K. Wong and M. Li, ―A multivariate intelligent decision-making model for retail sales
forecasting‖, Decision Support Systems, vol. 55, no. 1, (2013), pp. 247-255.
[9] X. D. Yu, Z. Q. Qi and Y. M. Zhao, ―Support Vector Regression for Newspaper/Magazine Sales
Forecasting‖, Proceeding Computer Science, vol. 17, (2013), pp. 1055-1062.
[10] L. Wan, B. Luo, H.M. Ji and W.W. Yan, ―A type of combination forecasting method based on time
series method and PLS‖, American Journal of Operations Research, vol. 2, no.4, (2012), pp.467-472.

Authors

Lihua Yang, She is a Lecturer, School of Economics and

Management, Hubei University of Automotive Technology. She
received her B.S degree from Yangtze University in 2003, and her
M.S. degree from Southwest Jiaotong University in 2006.Her
research interests include data mining and decision support.

Baolin Li, He is a Lecturer, School of Economics and Management,

Hubei University of Automotive Technology. He received his B.S
degree from Lanzhou University in 2005, and his M.S. degree from
Northwestern Polytechnical University in 2012. His research interests
include data analysis and data management.

Copyright ⓒ 2016 SERSC 75

International Journal of Database Theory and Application
Vol.9, No.1 (2016)

76 Copyright ⓒ 2016 SERSC