International Conference on Mechanical, Industrial and Energy Engineering 2018

23-24December, 2018, Khulna, BANGLADESH

ICMIEE18-299
Forecasting Demand by GMDH Predictor, a Case Study
Syed Misbah Uddin*, Aminur Rahman, Emtiaz Uddin Ansari
Department of Industrial and Production Engineering, Shah Jalal University of Science and Technology, Sylhet 3114,
BANGLADESH

ABSTRACT
Forecasting demand is very important for manufacturing industry and also needed for all type of business and business
suppliers for distribution of finish goods to the consumer on time. Forecasting with high accuracy is required to prevent
wasting and system failure to fulfil market demand. This study is concerned with the determination of accurate models for
forecasting cement demand. In this connection this paper presents results obtained by using a self-organizing model and
compares them with those obtained by usual statistical techniques. A nonlinear modelling technique based on Group Method of
Data Handling (GMDH) is considered here to derive forecasts. Various time series smoothing techniques such as exponential
smoothing, double exponential smoothing, weightage moving average and moving average method are used for forecasting the
demand. For this purpose, Monthly sales data of a typical cement ranging from January, 2007 to February, 2016 was collected.
The mean absolute percentage error (MAPE) and mean sum square error (MSE) are also calculated for comparing the
forecasting accuracy. The comparison of modelling results shows that the GMDH model perform better than other models
based on terms of mean absolute percentage error (MAPE) and mean square error (MSE).

Keywords: Forecast, GMDH algorithm, Time series, MAPE, MSE.

1. Introduction Synthesis of Polynomial Network (ASPN) algorithm to

“Better predictions remain the foundation of all improve the GMDH algorithm.
science. . . ”[1]. Forecast accuracy has been a critical J.A.Muller and Frank Lemke developed and improved
issue in the areas of financial, economic and scientific self-organizing data mining algorithms on the basis of the
modeling, and has motivated the growth of a vast body of above results in 1990s. Further enhancements of the
literature on the development and empirical application of GMDH algorithm have been realized in the
forecasting models [2].More accurately forecasting “KnowledgeMiner” software. The GMDH algorithm has
demand would facilitate for assisting managerial, gradually become an effective tool for modeling,
operational and tactical decision making. Therefore the forecasting, and decision support and pattern recognition
selection of forecasting model is the important criteria of complex systems. There are processes for which it is
that will influence to the forecasting accuracy [3]. needed to know their future or to analyze inter-relations.
GMDH method was originated in 1968 by Prof. Alexey The GMDH algorithm was successfully used to deal with
G. Ivakhnenko in the Institute of Cybernetics in uncertainty, linear or nonlinearity of systems in a wide
Kiev (USSR). This approach from the very beginning range of disciplines such as ecology, economy, medical
was a computer-based method so, a set of computer diagnostics, signal processing, power plant, electric
programs and algorithms were the primary practical power industry and control systems [5-9]. The revised
results achieved at the base of the new theoretical GMDH algorithms [10, 11]), have been introduced to
principles. The method was quickly settled in the large model dynamic systems in flood forecast and petroleum
number of scientific laboratories worldwide due to open resource prediction with some success.
code sharing. At that time code sharing was quite a The purpose of the study is to investigate how good is
physical action since the internet is at least 5 years GMDH predictor as a forecasting tool by comparing the
younger than GMDH. Despite this fact the first results of a self-organizing model with those obtained by
investigation of GMDH outside the Soviet Union had usual statistical techniques. This paper organize as
been made soon by R.Shankar in 1972. Later on different follows: section 2 describes the methodology applied for
GMDH variants were published by Japanese and Polish forecasting the demand. Data collection and analysis are
scientists. presented in section 3. Section 4 provides the results and
The main idea of GMDH is the use of feed-forward discussion while fifth section offers some conclusion.
networks based on short-term polynomial transfer
functions whose coefficients are obtained using 2. Methodology
regression combined with emulation of the self- This is a case study research based on time series data
organizing activity behind NN structural learning [4]. To of cement industry. The data used in this case study are
improve the performance of the GMDH algorithm, monthly sales data of cement. The data span the period
Barron gave a comprehensive overview of some early from January 2007 to February 2016. The dataset
developments of network, and introduced the polynomial consists of 110 months’ time series data. Data were
network training algorithm (PNETTR). Elder proposed analyzed by using various time series model such as
moving average, weighted moving average, single

* Corresponding author. Tel.: +88-01770347090

E-mail addresses: misbah-ipe@sust.edu
exponential smoothing, double exponential smoothing error (MSE) were also calculated to assess forecasting
and least square method of simple linear regression. performance of different models.
In this study, we use the value of α 0.3 and 0.5 for
single exponential smoothing method. Simple 3.1 Analysis by GMDH algorithm
exponential smoothing does not do well when there is GMDH algorithm consists of set of steps that are
a trend in the data. In such situations, several methods described below:
were devised under the name "double exponential Step 1: First N observations of regression-type data
smoothing" or "second-order exponential are taken. The collected load data are first normalized
smoothing. The basic idea behind double exponential with respect to their individual base value in order to
smoothing is to introduce a term to take into account the restrict the variation of data within the same level.
possibility of a series exhibiting some form of trend. Those normalized data are denoted by
This slope component is itself updated via exponential ( 1 ) where M is the total
smoothing.One method sometimes referred to as "Holt- number of input. The original data is separated into the
Winters double exponential smoothing are followed training and test sets [14]. In this study total 110 data
here. One of twosmoothing factor is α which is called were separated into training (58) and test (52) sets. The
data smoothing factor and it’s value, 0 < α < 1, and the 58 data is used for the estimation of the partial
other one β is the trend smoothing factor, 0 < β < 1. descriptions which describe the partial characteristics of
We also used the GMDH predictor version GMDH Data the nonlinear system. The 52 data is used for organizing
Science 3. 5. 9 to derive the forecast. Out of 110 data the complete description which describes the complete
58 months data are used for the training set and rest of characteristic of the nonlinear system.
the data are used for evaluation in checking set. In order Step 2: Select ( ) ( − 1) new input variables
to evaluate the forecasting accuracy of different according to all possibilities of connection by each pair
techniques various central tendency measures as the loss of inputs in the layer. Construct the regression
function were also calculated with the help of following polynomial for this layer by forming the quadratic
formula. expression which approximates the output y in equation
1 𝑛 (1).
MAD = ∑𝑛=1 |( Actual − Forecast)| Step 3: Identify the single best input variable out of
n
n = the number of periods [12]. these( ) input variables, according to the value of mean
square error (MSE). The input of variables that give the
𝑛
∑ {Actual−Forecast }2 best results in the first layer, are allowed to form second
MSE = 𝑘=0
n layer candidate model of the equation (1). Set the new
Where: input ( 1 ) and ( 1)
n = the number of periods [12]. Models of the second layer are evaluated for compliance
𝑛 by using MSE, and again the input variables that give
1 |Actual−Forecast|
MAPE = ∑ ∗ 100% best results will proceed to form third layer candidate
n 𝑛=1 Actual
models. This procedure is carried out as long as the
MSE for the test data set decrease compared with the
3. Data Collection and Analysis value obtained at the previous one as shown in Fig. 2.
Data Collection is a significant aspect of any type of After the best models of each layer have been selected,
research study. The data used in this case study are the output model is selected by the MSE. The model
monthly sales data of cement. The data span the period with the minimum value of the MSE is selected as the
from January 2007 to February 2016. The time series output model [15].
plot is given Fig. 1.
30000
28000

26000

24000
Sales Volume

MSE

22000

20000

18000

16000

14000

12000 0 1 2 3 4 5 6 7 8
0 10 20 30 40 50 60 70 80 90 100 110
Iterations
Month
Fig.2 Stopping criteria of GMDH algorithm
Fig.1 Monthly sales data (Jan 2007 to Feb 2016)
3.2 Analysis by Statistical method
After collecting sales data GMDH algorithm and Various time series smoothing techniques such as
various statistical forecasting techniques were used to exponential smoothing, double exponential smoothing,
forecast. The mean absolute deviation (MAD), mean moving average and regression method were used for
absolute percentage error (MAPE) and mean square forecasting the load demand.Absolute deviations were

ICMIEE18-299- 2
also calculated. The mean absolute deviations (MADs) forecasting techniques are plotted in Fig. 3. It is seen
found from these calculations are listed in table 1. that GMDH algorithm gives lowest value of
MADwhich is best suit.
Table 1MAD of different forecasting methods 3000
Method MAD 2500
3 month Moving Average 2306
2000

MAD
6 month Moving Average 2791
1500
12 month Moving Average 2230
1000
Weightage Moving Average 2056
Regression 2459 500

GMDH Method 704 0

Exponential α=0.3 2286
Exponential α= 0.5 2053
Double Exponential α= 0.3, β= 0.5 2861
Forecasting methods
From Table 1 it is seen that the value of MAD due to
Fig.3Comparison of MAD of different techniques
forecasting by GMDH algorithm is 704. On the other
hand all the statistical method gives four digits MAD. The mean absolute percentage error (MAPE) and mean
The mean absolute percentage error (MAPE) and mean square error (MSE) are plotted in Fig.4 and Fig.5
square error (MSE) were also calculated and reported in
respectively. The comparison of modelling results
Table 2 and Table 3 respectively. It is observed that the
shows that the GMDH model perform better than other
GMDH forecast with only 4% MAPE and nearest value
models based on terms of mean absolute percentage
is 9% which is done by exponential smoothing
error (MAPE) and mean square error (MSE).
technique(α= 0.5). Form Table 3 it is clear that the
model with the minimum value of the MSE is the 14%
GMDH model. 12%
10%
MAPE

8%
Table 2MAPE of different forecasting methods 6%
Method MAPE 4%
3 month Moving Average 11% 2%
0%
6 month Moving Average 14%
12 month Moving Average 11%
Weightage Moving Average 10%
Regression 12%
GMDH Method 4%
Forecasting methods
Exponential α=0.3 11%
Fig.4Comparison of MAPE of different techniques
Exponential α= 0.5 9%
Double Exponential α= 0.3, β= 0.5 13% 12000000
10000000
Table 3MSE of different forecasting methods
8000000
MSE

Method MSE
6000000
3 Month Moving Average 7994519
6 Month Moving Average 10355301 4000000
12 Month Moving Average 7710194 2000000
Weightage Moving Average 6291543 0
Regression 9177720
GMDH Method 824882
Exponential α= 0.3 7619269
Exponential α= 0.5 6220179
Double Exponential α= 0.3, β= 0.5 11913465 Forecasting methods
Fig.5Comparison of MSE of different techniques
4. Results and Discussion To assess the performance of GMDH modelling, last 52
After completing data analysis we have come out with months demand were forecasted and compared with the
some informative results. The calculated Mean absolute test set. The results of that model along with forecasting
deviations (MADs) of forecasted data by different precision are shown in table 4. Normalize mean

ICMIEE18-299- 3
absolute error is found to be 4.65% whereasnormalize [5] A.G. Ivakheneko, G. A. Ivakheneko, A Review of
RMS is 6%. The fitting accuracy of GMDH model Problems Solved by Algorithms of the GMDH,
algorithm is also very good as the value of R2 is 0.90. Pattern Recognition and Image Analysis, Vol. 5,
pp 527-535 (1995).
Table 4Summary results of GMDH modelling [6] G. C. Onwubolu, P. Buryan, F. Lemke, Modeling
Metrics Output / Value Tool Wear in End-Miling Using Enhanced
Post processed result Model fit GMDHLearning Networks, International Journal
of Advance Manufacture Technology, Vol.39, pp
Number of observations 52 1080–1092 (2008).
Normalize mean absolute error 4.65% [7] T. Kondo, A. S. Pandya, A. S. Nagashino, GMDH-
(NMAE) root mean square error
Normalize 6% Type Neural Network Algorithm with a Feedback
(NRMSE)
Standard deviation of residuals 5.8% Loop for Structural Identification of RBF Neural
Network, International Journal of Knowledge-
oefficient of determination 0.90 Based and Intelligent Engineering Systems, Vol. 11,
Correlation coefficient 0.95 pp 157-168 (2007).
[8] V.Puig, M. Witczak, F. Nejjari, J. Quevedo, J.
Our findings have several important implications. Korbicz, A GMDH Neural Network-Based
Useless input variables are eliminated and useful input Approach to Passive Robust Fault Detection Using
variables are selected automatically, the structure a Constraint Satisfaction Backward Test,
parameters and the optimum GMDH architecture can be Engineering Applications of Artificial Intelligence,
organized automatically. The case study on the cement Vol 20, pp 886-897 (2007).
time series data testing demonstrated that the GMDH [9] F. Li, B. R. Upadhyaya,L. A. offey, “Model-
model is robust in the forecasting of nonlinear time Based Monitoring and Fault Diagnosis of Fossil
series. Power Plant Process Units Using Group Method of
5. Conclusion Data Handling, ISA Transactions, Vol. 2, pp 213-
This paper examined the forecasting accuracy of 219 (2009).
different statistical techniques as well as GMDH [10] T. Kondo, Nonlinear Pattern Identification by
predictor. For that purposes ten years secondary sales Multi-Layered GMDH-Type Neural Network Self-
data of a cement were collected. There was low seasonal Selecting Optimum Neural Network Architecture.
variation in their sales. Demand forecasting was International Conference on Neural Information
performed using extrapolative time series methods, such Processing,pp 882-891 (2007)
as exponential smoothing with level, trend, and seasonal [11] F. Chang, Y. Hwang,A Self-Organization
components. Besides that moving average, weighted Algorithm for Real-Time Flood
moving average and regression method were also used Forecast,Hydrological processes, Vol. 13, pp 123-
for forecasting the demand. A nonlinear self-organizing 138 (1999).
model based on Group Method of Data Handling [12] P. K. Sahu, R. Kumar, Demand Forecasting for
(GMDH) was also applied here to derive forecasts.We Sales of Milk Product (Paneer) in Chhattisgarh,
applied the GMDH predictor version GMDH Data International Journal of Inventive Engineering
Science 3. 5. 9. and Sciences, Vol. 1, pp 10-13 (2013).
In order to evaluate the accuracy of prediction, various [13] O. S. Ezennaya, O. E. Isaac, U. O. Okolie,
performance measures such as MAD. MAPE and MSE Analysis of Nigeria‘s National Electricity Demand
were calculated. It is found that there is no result near to Forecast (2013-2030), International Journal Of
the GMDH predictor. GMDH algorithm forecast with Scientific & Technology Research, Vol. 3, pp 333-
only 0.0367 or 4% error which is substantially more 340 (2014).
accurate than statistical method. [14] R. Samsudin,P. Saad, Combination of Forecasting
Using Modified GMDH and Genetic
REFERENCES Algorithm,International Journal of Computer
[1] S.Makridakis, M. Hibon, The M3-Competition: Information Systems and Industrial Management
results, conclusions and implications,International Applications, Vol. 1, pp 170-176(2009).
[15] T. Kondo, J. Ueno, Revised GMDH-type Neural
Journal of Forecasting,Vol. 16,pp451–476 (2000).
Network Algorithm with a Feedback Loop
[2] J. Gooijer, R. Hyndman, 25 Years of Time Series
Identifying Sigmoid Function Neural Network,
Forecasting, International Journal of Forecasting,
International Journal of Innovative Computing,
Vol. 22, pp 443– 473 (2006).
Information and Control, Vol. 2(5), pp 985-996
[3] R. Samsudin, P. Saad, A. Shabri, Hybridizing
(2006).
GMDH and Least quares SVM support vector
machine for forecasting tourism demand, IJRRAS,
Vol. 3, pp 274-279 (2010).
[4] S. J. Farlow, The GMDH Algorithm of
Ivakhnenko, The American Statistician, Vol. 35,
pp 210-215 (1981).

ICMIEE18-299- 4