Available online at www.sciencedirect.

com

Computers & Industrial Engineering 53 (2007) 610627

www.elsevier.com/locate/dsw

Forecasting Thailands rice export: Statistical techniques

vs. articial neural networks
a,* b
Henry C. Co , Rujirek Boosarawongse
a
Technology & Operations Management, California Polytechnic and State University, 3801 W. Temple Avenue, Pomona, CA 91786, USA
b
Department of Applied Statistics, Faculty of Science, KMITL, Bangkok, Thailand

Received 17 October 2006; received in revised form 14 February 2007; accepted 8 June 2007
Available online 14 June 2007

Abstract

Forecasting the international trade of rice is dicult because demand and supply are aected by many unpredictable
factors (e.g., trade barriers and subsidies, agricultural and environmental factors, meteorological factors, biophysical fac-
tors, changing demographics, etc.) that interact in a complex manner. This paper compares the performance of articial
neural networks (ANNs) with exponential smoothing and ARIMA models in forecasting rice exports from Thailand. To
ascertain that the models can reproduce acceptable results on unseen future, we evaluated various aggregate measures of
forecast error (MAE, MSE, MAPE, and RMSE) during the validation process of the models. The results reveal that while
the HoltWinters and the BoxJenkins models showed satisfactory goodness of t, the models did not perform as well in
predicting unseen data during validation. On the other hand, the ANNs performed relatively well as they were able to track
the dynamic non-linear trend and seasonality, and the interactions between them.
 2007 Elsevier Ltd. All rights reserved.

Keywords: Articial neural network; Time series analysis; HoltWinters; BoxJenkins; Forecasting

1. Introduction

Rice, Oryza sativa L., is the staple food for over half of the worlds population. Nine of the top 10 rice pro-
ducers China, India, Indonesia, Bangladesh, Vietnam, Thailand, Myanmar, the Philippines, and Japan are
in Asia. In these countries, rice cultivation is intertwined with religious observances, festivals, customs, folk-
lore, and traditions. The worlds top three rice producers China, India, and Indonesia consume most of the
rice domestically. As harvests are always subject to whims of Mother Nature, a potential exporter (e.g., China)
may temporarily become a rice importer.
International rice prices are determined by supply and demand, and to a large extent, governmental inter-
ventions (trade barriers and subsidies). Forecasting rice prices has been an area of active research, and
researchers have relied on statistical techniques to predict rice prices. Deetae (1991) applied the decomposition

*
Corresponding author. Tel.: +1 909 869 2465.
E-mail addresses: hco@csupomona.edu (H.C. Co), rujirek@mozart.inet.co.th (R. Boosarawongse).

0360-8352/$ - see front matter  2007 Elsevier Ltd. All rights reserved.
doi:10.1016/j.cie.2007.06.005
 H.C. Co, R. Boosarawongse / Computers & Industrial Engineering 53 (2007) 610627 611

Table 1
Proportion of export by weight
Category % By weight
Jasmine 31
White 39
Parboiled 26
Glutinous 4

method and the BoxJenkins model to investigate rice farm-gate prices. The study revealed that the BoxJen-
kins model performed better. Kerdsomboon (1999) applied various statistical forecasting methods to study
planting area, rice production, and agricultural farm prices. The author found the regression model to be suit-
able for forecasting planting area and rice production, while the BoxJenkins model performed better in pre-
dicting agricultural farm prices. Maliwan (2003) employed the ARIMA model to forecast the domestic prices
of jasmine rice. Likewise, Fansiri (2004) employed the ARIMA model to predict rice export prices. The studies
revealed that the autoregressive (AR) model ts suciently well. Sangpattaranate (2005) compared 4 forecast-
ing techniques HoltWinters additive model, BoxJenkins model, decomposition method and regression
analysis for forecasting the rice prices in Thailand. The decomposition method was found to be most suitable
in forecasting the prices of jasmine rice and main-crop rice with 5%, 15%, 25% broken, while the BoxJenkins
model was best for main-crop rice with 10% broken.
The worlds top three rice exporters are Thailand, India, and Vietnam, respectively. Thailand, with a mar-
ket share of around 25%, has been the worlds top rice exporter for decades. The majority of Thailands pop-
ulation is in the agricultural sector. Over 50% of the countrys farmland is devoted to rice. About 55% of rice is
produced for domestic consumption, and the remaining 45% is for export. Rice export accounts for about 20%
of Thailands total agricultural export earnings (Vanichanont, 2004).
The Rice Exporters Association in Thailand is the center for coordination among its 186 members. The
export target set by the association at the beginning of each year is closely monitored by Thailands Ministry
of Commerce. It is a key economic indicator for the country. The export target is developed by considering
demand from importing countries and the supply of exporting countries. Information such as world popula-
tion growth, world rice stock, production rate, consumption rate, etc. are also taken into account in setting the
target. Unfortunately, the supply and demand of rice are aected by many unpredictable factors (e.g., agricul-
tural and environmental factors, meteorological factors, biophysical factors, changing demographics, etc.)
that interact in a complex manner. Setting the export target has always been a great challenge.
The objectives of this article are twofold: (1) to examine the network architecture of articial neural net-
works (ANNs) in forecasting rice exports from Thailand and (2) to evaluate how various time-series forecast-
ing methods compare in predicting the unseen future.
Thailand exports several categories of rice jasmine rice (Thai Hom Mali rice), white rice, glutinous rice,
and parboiled rice. Table 1 shows the proportion of export quantity of each category in the last decade. Prior
to year 2001, the monthly export data of jasmine rice was combined with white rice. Accordingly, we merged
the two time series into one. Adding the three categories (jasmine + white rice, glutinous, parboiled rice) of
rice gave us a 4th time series, i.e., total export.
The time series data used in this study was provided by the Rice Exporters Association in Thailand. The
data covers the months of January 1996December 2005. We used the time series from January 1996 to
December 2004 for model tting, and the remaining data for model validation. A model may t historical data
perfectly, but fail to predict the future. The purpose of model validation is to ascertain that the tted model
can reproduce acceptable results on unseen future. To obtain an understanding of the underlying forces and
structure that produced the observed time series data, we investigated three alternative models. These are:
exponential smoothing, the BoxJenkins model, and articial neural-network simulation.

2. Exponential smoothing

Exponential smoothing methods originated in the 1950s and 1960s with the work of Brown (1959),
Holt (1957), and Winters (1960). Since then numerous variations on the original methods have been
612 H.C. Co, R. Boosarawongse / Computers & Industrial Engineering 53 (2007) 610627

proposed. Various literatures have addressed the exceptional performance of the exponential smoothing
methods. For example, Satchell and Timmermann (1995) and Chateld et al. (2001) showed that simple
exponential smoothing (SES) is optimal for a wide range of data generating processes. The simple expo-
nential smoothing method was shown to perform better than rst order ARIMA models (Hyndman,
2001).
Hyndman, Koehler, Snyder, and Grose (2002) provides taxonomy of the exponential smoothing methods.
Each method consists of one of ve types of trend (none, additive, damped additive, multiplicative, and
damped multiplicative) and one of three types of seasonality (none, additive, and multiplicative). Among
the 15 dierent exponential smoothing methods, the best known are simple exponential smoothing (no trend,
no seasonality), Holts linear method (additive trend, no seasonality), HoltWinters additive method (additive
trend, additive seasonality) and HoltWinters multiplicative method (additive trend, multiplicative seasonal-
ity) (De Gooijer et al., 2006). Gardner and McKenzie (1988) provides some simple rules based on the variances
of dierenced time series for choosing an appropriate exponential smoothing method. Tashman and Kruk
(1996) compared these rules with others proposed by Collopy and Armstrong (1992) and an approach based
on the Bayes information criterion (BIC). Hyndman et al. (2002) also proposed an information criterion
approach, but using the underlying state space models. Analysis of monthly rice export data indicated season-
ality (see Section 3), hence the HoltWinters additive method is used in this paper.

2.1. Initial conditions

HoltWinters exponential smoothing requires initial estimates of the level, the trend and the seasonal indi-
ces, as well as appropriate values for the smoothing constants a, b, and c. In this paper, we used the Solver tool
in Microsoft Excel to determine the optimal values of these parameters. The target for optimization is the sum
of squared forecast errors. The resulting models showed satisfactory goodness of t.

2.2. Forecast evaluation and accuracy measures

A forecasting method generates a multivariate error distribution, and the purpose of an error measure is to
provide an informative and clear summary of the distribution (Murphy & Winkler, 1992). Varous accuracy
measures have been used to evaluate the performance of forecasting methods (see for instance, Mahmoud,
1984).
The forecast error is the dierence between the forecast value and the actual value for the corresponding
period, i.e., et = Yt  Ft, where et is the forecast error at period t, Yt is the actual value at period t, and Ft
is the forecast for period t. The most common error measures include mean squared error, root mean squared
error, mean absolute error, median absolute error, mean absolute percentage error, median absolute percent-
age error, symmetric mean absolute percentage error, symmetric median absolute percentage error, mean rel-
ative absolute error, median relative absolute error, geometric mean relative absolute error, relative mean
absolute error, relative root mean squared error, log mean squared error ratio, percentage better, percentage
better mean absolute error, and percentage better mean squared error (De Gooijer et al., 2006). The relative
root mean squared error is also known as Theils U statistic or U2 (Theil, 1966). The average ranking of a
method relative to all other methods considered has also been used. The mean absolute error (MAE), median
absolute percentage error (MAPE), mean squared error (MSE) and the root mean squared error (RMSE) are
the most commonly used error measures in business, and are used to evaluate the forecast models in this
paper.

2.3. Model tting

The data used in model tting covered the months of January 1996December 2004. The data from January
2005 to December 2005 was used for validation. Fig. 1 shows the plots of the 4 time series from January 1996
to December 2004.
Fig. 2 shows a screenshot of the Excel worksheet for tting the HoltWinters additive model. The initial
values of the level (cell C22), the trend (cell D22) and the seasonal indices (cells E11:E22), as well as values
 H.C. Co, R. Boosarawongse / Computers & Industrial Engineering 53 (2007) 610627 613

(Thousand) metric tons (Thousands) Metric Ton

1200 900
800
1000
700
800 600
500
600
400
400 300
200
200 100
0
0 0 20 40 60 80 100
0 20 40 60 80 100
Month Month

(a) Total Thai Rice (b) White & Jasmine Rice

(Thousand) metric tons (Thousand) metric tons

60 60

50 50

40 40

30 30

20 20

10 10

0 0
0 10 20 30 40 50 60 70 80 90
0 10 20 30 40 50 60 70 80 90
Month Month

(c) Glutinous Rice (d) Parboiled Rice

Fig. 1. Monthly Rice Export (January 1996December 2004).

for the smoothing constants a, b, and c (cells C6:C8) have been optimized (using Excels Solver tool) to min-
imize the sum of squared errors (cell L24). Notice that minimizing the sum of squared errors (cell L24) min-
imizes the MSE and the RMSE.
The same Excel template (shown in Fig. 2) was used for all categories of rice (total, white and jasmine, glu-
tinous, and parboiled). Using the Solver tool in Microsoft Excel, we were able to minimize the MSE and the
RMSE. Fig. 3 shows plots comparing actual export volumes with the tted values for the 4 time series.
The plots in Fig. 1 suggest that rice export may follow some consistent seasonal pattern. The estimated sea-
sonal indices of the 4 time series are summarized in Table 2.
The aggregate error measures for the 4 time series are shown in Table 3. Many textbooks recommend the
use of the mean absolute percent error (MAPE) for evaluating the accuracy of population forecasts (e.g., Han-
ke & Reitsch, 1995; Bowerman, Oconnell, & Koehler, 2004), and it was the primary measure in Alon, Qi, and
Sadowski (2001) for comparing the performance of articial neural networks and traditional methods in fore-
casting U.S. aggregate retail sales.
Table 3 shows that, except for glutinous rice, the HoltWinters additive model appeared to have tted the
time series data satisfactorily.

2.4. Model validation

The purpose of model validation is to ascertain that the tted model can reproduce acceptable results on
unseen future. Table 4 compares the aggregate error measures for validating the tted models. Recall that we
used data from January 1996 to December 2004 to t the models. The data used in the validation, January
December 2005, are unseen during model-tting.
From Table 4, it is apparent that the tted HoltWinters additive models were not able to reproduce
acceptable results on unseen validation data. The biggest disappointment is the white and jasmine category.
While the MAPE is less than 14% (see Table 3) during model-tting, the MAPE shown in Table 4 is more
than 47%! Although historical data was able to produce a good t, the models failed during validation to pre-
dict unseen data.
614 H.C. Co, R. Boosarawongse / Computers & Industrial Engineering 53 (2007) 610627

Fig. 2. Excel worksheet for HoltWinters additive model.

3. BoxJenkins model for forecasting

The BoxJenkins (Box, Jenkins, & Reinsel, 1994; Wei, 1990) method is a statistical technique for time series
forecasting. This study used the seasonal autoregressive integrated moving average (ARIMA)
(p, d, q) (P, D, Q)s model, with sub-index s denoting the seasonal period. For any stationary time series,
Zt, the ARIMA (p, d, q) (P, D, Q)s of Zt is UP(Bs)/p(B)(1  B)d(1  Bs)DZt = hq(B)HQ(Bs)at, where UP(Bs)
and HQ(Bs) are the regular autoregressive and moving average factors; UP(B) and HQ(Bs) are the seasonal
autoregressive and moving average factors, respectively. The BoxJenkins method consists of the following
steps:

1. Identication. The BoxJenkins model assumes that the time series is stationary. This study used logarith-
mic transformation for non-constant variance, and dierencing for non-constant mean. The sample ACF
and PACF of the stationary series were then computed and examined to identify the order of p and q.
2. Estimation. The least-squares method for non-linear parameters was used to determine the parameters of
the model.
3. Diagnostic checking. The LjungBox statistic (Ljung & Box, 1978) was used to investigate model adequacy,
i.e., to ascertain that the residuals are white noise (zero mean, constant variance, uncorrelated process and
normally distributed).
4. Forecasting. The model giving the least mean squared error (MSE) was chosen for forecasting. The fore-
casting model was then used to compute the tted values and forecasts.
 H.C. Co, R. Boosarawongse / Computers & Industrial Engineering 53 (2007) 610627 615

1,200,000 1,000,000
Actual Actual
1,000,000 Fitted 800,000 Fitted
Error Error

800,000 600,000

600,000 400,000

400,000
200,000

200,000
0

0
-200,000

-200,000
-400,000

-400,000 (a) Total Rice Export (b) White & Jasmine Rice Export

70,000 350,000
Actual
60,000 Actual
Fitted 300,000
Fitted
Error
50,000 Error
250,000
40,000
200,000
30,000
150,000
20,000
100,000
10,000
50,000
0
0
10,000-

20,000- -50,000

30,000- -100,000

40,000- -150,000
(c) Glutinous Rice Export (d) Parboiled Rice Export

Fig. 3. HoltWinters model-tting errors.

Table 2
Seasonal Indices (HoltWinters)
Month Total export White and jasmine Glutinous Parboiled
January 0.96585 0.97303 1.07246 0.92231
February 0.87561 0.90491 1.07719 0.73213
March 0.91500 0.97816 0.98582 0.71602
April 0.87881 0.91277 0.96981 0.74298
May 0.97725 0.95517 0.90979 1.04129
June 0.94920 0.95650 0.75123 0.94006
July 1.01968 1.03834 0.77802 1.02245
August 0.97502 0.94218 0.92591 1.11896
September 0.98337 0.93551 0.80276 1.18277
October 1.16362 1.07499 0.97361 1.46143
November 1.12420 1.10500 1.21992 1.16589
December 1.16980 1.22033 1.52714 0.94696

Table 3
Aggregate error measures of model tting error (HoltWinters)
Category MAE MAPE(%) MSE RMSE
Total rice export 70,265 12.3 7,811,980,594 88,385
White and jasmine 59,891 13.9 5,780,870,803 76,032
Glutinous 5880 33.0 54,766,847 7400
Parboiled 23,689 21.1 977,153,476 31,259
616 H.C. Co, R. Boosarawongse / Computers & Industrial Engineering 53 (2007) 610627

Table 4
Aggregate measures of validation errors (HoltWinters)
Category MAE MAPE (%) MSE RMSE
Total rice export 207,882 35.1 53,678,782,994 231,687
White and jasmine 194,042 47.4 43,461,583,428 208,474
Glutinous 9,792 39.9 149,546,455 12,229
Parboiled 46,512 31.3 2,943,833,536 54,257

The data used in model tting covered the months of January 1996December 2004. The data from January
2005 to December 2005 was used for validation. The plots in Fig. 1 indicate that the time series are non-sta-
tionary in the mean and the variance. Fig. 4 shows the transformed time series. First, the natural logarithm
was taken, and then dierencing was applied. The sample ACF and PACF indicated that there is seasonality
in the series. Accordingly, seasonal dierencing was applied. The transformed series are stationary. The sam-
ple ACF and the sample PACF for the transformed series are plotted in Figs. 5 and 6, respectively.
The plots suggest that the ARIMA model is appropriate. Several models were found, and the best tted
model (least MSE) is presented below. At signicance level 0.05, the estimates were found signicant. The val-
ues under each estimator (in parentheses) below are the standard error: Total Thai rice: ARIMA
(0,1,1) (0,1,1)12
1  B1  B12 ln Z t 1  0:4816B1  0:5258B12 at 1
0:0961 0:1137
ln Z t ln Z t1 ln Z t12  ln Z t13 at  0:4816at1  0:5258at12 0:2532at13

White and Jasmine rice: ARIMA (0,1,1) (0,1,1)12

1  B1  B12 ln Z t 1  0:5331B1  0:8767B12 at 2
0:0915 0:0864
ln Z t ln Z t1 ln Z t12  ln Z t13 at  0:5331at1  0:8767at12 0:4674at13

Glutinous rice: ARIMA (0,1,1) (0,1,1)12

0.8 0.8

0.6 0.6

0.4 0.4

0.2 0.2

0 0

0.2- -0.2

0.4- -0.4

0.6- -0.6

0.8- -0.8

(a) Total Thai Rice (b) White & Jasmine Rice

3
3 2.5
2.5 2
2 1.5
1.5 1
1 0.5
0.5 0
0 -0.5
-0.5 -1
-1 -1.5
-1.5 -2
-2 -2.5
-2.5 (c) Glutinous Rice (d) Parboiled Rice

Fig. 4. The dierenced series of natural logarithms.

 H.C. Co, R. Boosarawongse / Computers & Industrial Engineering 53 (2007) 610627 617

Autocorrelation Function for DDSlnT Autocorrelation Function for DDSLNFW

1.0
0.8 1.0
Autocorrelation

0.8

Autocorrelation
0.6
0.6
0.4
0.4
0.2
0.2
0.0
-0.2 0.0
-0.2
-0.4
-0.4
-0.6
-0.6
-0.8
-0.8
-1.0
-1.0

10 20 30 40 10 20 30 40
Lag Corr T LBQ Lag Corr T LBQ Lag Corr T LBQ Lag Corr T LBQ
Lag Corr T LBQ Lag Corr T LBQ Lag Corr T LBQ Lag Corr T LBQ
1 -0.40 -3.89 15.61 13 0.17 1.19 49.60 25 -0.03 -0.19 59.16 37 0.05 0.28 87.51
1 -0.33 -3.48 12.46 13 0.16 1.21 60.01 25 -0.04 -0.27 81.44 37 0.13 0.80 111.92
2 -0.06 -0.52 15.98 14 0.06 0.41 49.99 26 -0.08 -0.52 59.95 38 0.09 0.55 88.81
2 -0.13 -1.20 14.27 14 0.10 0.73 61.27 26 -0.07 -0.48 82.17 38 0.02 0.10 111.96
3 0.11 0.89 17.09 15 -0.19 -1.31 54.04 27 0.24 1.59 67.58 39 -0.16 -0.98 92.95
3 0.06 0.58 14.71 15 -0.17 -1.21 64.79 27 0.18 1.23 86.94 39 -0.10 -0.65 113.75
4 -0.05 -0.45 17.38 16 0.14 0.99 56.48 28 -0.19 -1.22 72.39 40 0.13 0.76 95.59
4 0.12 1.13 16.41 16 -0.03 -0.22 64.90 28 0.02 0.11 86.98 40 -0.00 -0.01 113.75
5 -0.02 -0.13 17.40 17 0.02 0.16 56.55 29 0.01 0.04 72.39
5 -0.16 -1.46 19.35 17 0.15 1.12 68.06 29 -0.14 -0.94 89.93
6 0.02 0.16 17.44 18 -0.04 -0.26 56.73 30 0.04 0.29 72.68
6 0.02 0.18 19.40 18 -0.02 -0.12 68.09 30 0.08 0.54 90.94
7 0.06 0.51 17.83 19 -0.07 -0.45 57.27 31 -0.09 -0.61 73.95
7 0.18 1.64 23.32 19 -0.12 -0.87 70.06 31 -0.07 -0.44 91.60
8 -0.08 -0.66 18.48 20 0.03 0.17 57.35 32 0.16 1.03 77.72
8 -0.20 -1.72 27.88 20 0.06 0.40 70.50 32 0.06 0.42 92.21
9 0.02 0.17 18.53 21 0.03 0.23 57.49 33 -0.14 -0.88 80.57
9 -0.01 -0.11 27.90 21 0.08 0.58 71.40 33 -0.08 -0.57 93.36
10 -0.04 -0.32 18.69 22 -0.07 -0.49 58.15 34 0.15 0.91 83.76
10 0.12 1.07 29.81 22 -0.19 -1.36 76.52 34 0.24 1.59 102.58
11 0.23 1.89 24.38 23 0.05 0.34 58.48 35 -0.05 -0.32 84.17
11 0.12 1.01 31.58 23 0.18 1.23 80.90 35 -0.19 -1.25 108.62
12 -0.45 -3.57 46.47 24 0.07 0.45 59.05 36 -0.14 -0.86 87.18
12 -0.45 -3.77 56.71 24 -0.05 -0.32 81.22 36 -0.06 -0.40 109.28

(a) Total Thai Rice (c) White & Jasmine Rice

Autocorrelation Function for DDSLNG Autocorrelation Function for DDSLNG

1.0
1.0
Autocorrelation

0.8
0.8

Autocorrelation
0.6
0.4 0.6
0.2 0.4
0.0 0.2
-0.2 0.0
-0.4 -0.2
-0.4
-0.6
-0.8 -0.6
-1.0 -0.8
-1.0

10 20 30 40
10 20 30 40
Lag Corr T LBQ Lag Corr T LBQ Lag Corr T LBQ Lag Corr T LBQ
Lag Corr T LBQ Lag Corr T LBQ Lag Corr T LBQ Lag Corr T LBQ
1 -0.41 -3.99 16.46 13 0.13 0.95 47.56 25 0.01 0.08 88.61 37 -0.12 -0.69 121.14
1 -0.41 -3.99 16.46 13 0.13 0.95 47.56 25 0.01 0.08 88.61 37 -0.12 -0.69 121.14
2 -0.03 -0.28 16.57 14 -0.10 -0.72 48.74 26 0.10 0.61 89.96 38 0.13 0.71 123.72
2 -0.03 -0.28 16.57 14 -0.10 -0.72 48.74 26 0.10 0.61 89.96 38 0.13 0.71 123.72
3 -0.09 -0.76 17.37 15 0.29 2.01 58.17 27 -0.14 -0.82 92.46 39 -0.14 -0.77 126.89
3 -0.09 -0.76 17.37 15 0.29 2.01 58.17 27 -0.14 -0.82 92.46 39 -0.14 -0.77 126.89
4 0.00 0.02 17.37 16 -0.19 -1.26 62.22 28 -0.03 -0.18 92.59 40 0.18 0.98 132.17
4 0.00 0.02 17.37 16 -0.19 -1.26 62.22 28 -0.03 -0.18 92.59 40 0.18 0.98 132.17
5 0.12 1.03 18.92 17 -0.01 -0.04 62.22 29 0.05 0.29 92.92
5 0.12 1.03 18.92 17 -0.01 -0.04 62.22 29 0.05 0.29 92.92
6 -0.00 -0.04 18.92 18 0.19 1.29 66.75 30 -0.13 -0.80 95.47
6 -0.00 -0.04 18.92 18 0.19 1.29 66.75 30 -0.13 -0.80 95.47
7 -0.06 -0.51 19.32 19 -0.31 -2.02 78.39 31 0.25 1.47 104.26
7 -0.06 -0.51 19.32 19 -0.31 -2.02 78.39 31 0.25 1.47 104.26
8 -0.13 -1.04 20.99 20 0.22 1.41 84.61 32 -0.10 -0.57 105.69
8 -0.13 -1.04 20.99 20 0.22 1.41 84.61 32 -0.10 -0.57 105.69
9 0.28 2.28 29.31 21 -0.17 -1.04 88.18 33 0.03 0.19 105.85
9 0.28 2.28 29.31 21 -0.17 -1.04 88.18 33 0.03 0.19 105.85
10 -0.06 -0.46 29.69 22 0.06 0.34 88.58 34 -0.18 -1.05 110.82
10 -0.06 -0.46 29.69 22 0.06 0.34 88.58 34 -0.18 -1.05 110.82
11 0.11 0.81 30.90 23 -0.01 -0.05 88.59 35 0.22 1.29 118.60
11 0.11 0.81 30.90 23 -0.01 -0.05 88.59 35 0.22 1.29 118.60
12 -0.36 -2.79 45.57 24 -0.00 -0.00 88.59 36 -0.03 -0.17 118.74
12 -0.36 -2.79 45.57 24 -0.00 -0.00 88.59 36 -0.03 -0.17 118.74

(c) Glutinous Rice (d) Parboiled Rice

Fig. 5. Sample ACF for the dierenced series of natural logarithms.

1  B1  B12 ln Z t 1  0:7290B1  0:8707B12 at 3

0:0670 0:0814
ln Z t ln Z t1 ln Z t12  ln Z t13 at  0:729at1  0:8707at12 0:6347at13

Parboiled rice: ARIMA (0,1,1) (0,1,1)12

1  0:2938B1  B1  B12 ln Z t 1  0:8304B1  0:8744B12 at 4

0:1422 0:0811 0:0862
ln Z t 1:2938 ln Z t1  0:2938 ln Z t2 ln Z t12  1:2938 ln Z t13
0:2938 ln Z t14 at  0:8304at1  0:8744at12 0:7261at13 :

Fig. 7 shows plots comparing actual export volumes with the tted values for the 4 time series. The aggre-
gate measures of model-tting errors for the 4 time series are shown in Table 5. Except for glutinous rice, the
BoxJenkins models appeared to have tted the data well.
Table 6 shows the aggregate measures of validation errors for the 4 time series. The MAPE ranged from
about 35% to 47%. Evidently, the BoxJenkins models failed to reproduce acceptable results on the unseen
validation data (see MAPE).
In the following section, we explore the use of articial neural network in forecasting.
618 H.C. Co, R. Boosarawongse / Computers & Industrial Engineering 53 (2007) 610627

Partial Autocorrelation Function for DDSlnT Partial Autocorrelation Function for DDSLNFW

Partial Autocorrelation
Partial Autocorrelation

1.0 1.0
0.8 0.8
0.6 0.6
0.4 0.4
0.2 0.2
0.0 0.0
-0.2 -0.2
-0.4 -0.4
-0.6 -0.6
-0.8 -0.8
-1.0 -1.0

10 20 30 40 10 20 30 40

Lag PAC T Lag PAC T Lag PAC T Lag PAC T Lag PAC T Lag PAC T Lag PAC T Lag PAC T
1 -0.40 -3.89 13 -0.14 -1.36 25 0.02 0.16 37 -0.04 -0.38 1 -0.33 -3.48 13 -0.14 -1.46 25 -0.09 -0.93 37 -0.02 -0.24
2 -0.26 -2.56 14 -0.09 -0.88 26 -0.08 -0.75 38 -0.06 -0.62 2 -0.27 -2.79 14 -0.04 -0.44 26 -0.08 -0.86 38 -0.02 -0.22
3 -0.04 -0.38 15 -0.17 -1.67 27 0.13 1.29 39 0.02 0.19 3 -0.10 -1.02 15 -0.14 -1.50 27 0.01 0.07 39 -0.01 -0.13
4 -0.04 -0.40 16 -0.03 -0.31 28 -0.06 -0.57 40 -0.00 -0.00 4 0.09 0.99 16 -0.09 -0.91 28 0.04 0.46 40 -0.02 -0.18
5 -0.04 -0.40 17 0.04 0.38 29 0.05 0.50 5 -0.08 -0.82 17 0.01 0.07 29 -0.01 -0.05
6 -0.02 -0.21 18 0.03 0.31 30 -0.06 -0.63 18 0.05 0.53 30 0.05 0.57
6 -0.04 -0.37
7 0.07 0.72 19 -0.01 -0.13 31 -0.12 -1.12
7 0.16 1.67 19 0.07 0.74 31 -0.07 -0.75
8 -0.02 -0.16 20 -0.11 -1.06 32 0.04 0.36
8 -0.09 -0.97 20 -0.11 -1.10 32 -0.02 -0.17
9 -0.01 -0.09 21 0.03 0.30 33 -0.06 -0.56
9 -0.06 -0.58 21 0.03 0.28 33 -0.03 -0.31
10 -0.07 -0.72 22 -0.15 -1.43 34 0.01 0.13
23 0.11 1.10 10 0.05 0.48 22 -0.10 -1.08 34 0.12 1.27
11 0.25 2.39 35 0.14 1.33
24 -0.09 -0.83 36 -0.19 -1.81 11 0.19 1.93 23 0.21 2.24 35 0.12 1.26
12 -0.34 -3.35
12 -0.33 -3.47 24 -0.23 -2.41 36 -0.28 -2.87

(a) Total Thai Rice (b) White & Jasmine Rice

Partial Autocorrelation Function for DDSLNG Partial Autocorrelation Function for DDSLNG

Partial Autocorrelation
Partial Autocorrelation

1.0 1.0
0.8 0.8
0.6 0.6
0.4 0.4
0.2 0.2
0.0 0.0
-0.2 -0.2
-0.4 -0.4
-0.6 -0.6
-0.8 -0.8
-1.0 -1.0

10 20 30 40 10 20 30 40

Lag PAC T Lag PAC T Lag PAC T Lag PAC T Lag PAC T Lag PAC T Lag PAC T Lag PAC T

1 -0.41 -3.99 13 -0.08 -0.77 25 -0.08 -0.76 37 -0.04 -0.36 1 -0.41 -3.99 13 -0.08 -0.77 25 -0.08 -0.76 37 -0.04 -0.36
2 -0.24 -2.35 14 -0.36 -3.49 26 -0.10 -0.94 38 -0.00 -0.00 2 -0.24 -2.35 14 -0.36 -3.49 26 -0.10 -0.94 38 -0.00 -0.00
3 -0.26 -2.55 15 0.01 0.08 27 -0.15 -1.43 39 0.01 0.13 3 -0.26 -2.55 15 0.01 0.08 27 -0.15 -1.43 39 0.01 0.13
4 -0.23 -2.26 16 -0.23 -2.23 28 -0.20 -1.96 40 -0.09 -0.88 4 -0.23 -2.26 16 -0.23 -2.23 28 -0.20 -1.96 40 -0.09 -0.88
5 -0.04 -0.35 17 -0.00 -0.04 29 0.03 0.29 5 -0.04 -0.35 17 -0.00 -0.04 29 0.03 0.29
6 0.02 0.21 18 0.27 2.59 30 -0.09 -0.88 6 0.02 0.21 18 0.27 2.59 30 -0.09 -0.88
7 -0.03 -0.28 19 -0.14 -1.34 31 0.04 0.43 7 -0.03 -0.28 19 -0.14 -1.34 31 0.04 0.43
8 -0.19 -1.85 20 -0.09 -0.84 32 0.01 0.09 8 -0.19 -1.85 20 -0.09 -0.84 32 0.01 0.09
9 0.17 1.67 21 -0.07 -0.71 33 -0.02 -0.16 9 0.17 1.67 21 -0.07 -0.71 33 -0.02 -0.16
10 0.15 1.49 22 0.02 0.23 34 -0.09 -0.85 10 0.15 1.49 22 0.02 0.23 34 -0.09 -0.85
11 0.30 2.90 23 0.14 1.41 35 0.00 0.01 11 0.30 2.90 23 0.14 1.41 35 0.00 0.01
12 -0.16 -1.59 24 -0.25 -2.40 36 -0.03 -0.26 12 -0.16 -1.59 24 -0.25 -2.40 36 -0.03 -0.26

(c) Glutinous Rice (d) Parboiled Rice

Fig. 6. Sample PACF for the dierenced series of natural logarithms.

4. Neural network model for forecasting

The use of articial neural networks for forecasting and modeling has generated considerable interest in
recent years. The ANN have been found to be able to decode non-linear time series data (Lapedes & Farber,
1987), and have been applied in stock market (Chong & Kyoung, 1992; Freisleben, 1992; Kimoto, Asakawa,
Yoda, & Takeoka, 1990; Schoneburg, 1990; Yao, Lan, & Li, 2000), bond ratings, (Dutta & Shekhar, 1990),
commodity and currency exchange (Bergerson & Wunsch, 1991; Grudnitski & Osburn, 1993; Hann & Steurer,
1996; Kuan & Liu, 1995; Refenes, 1995), and other dicult-to-predict situations.

4.1. Articial neural networks

Articial neural networks are non-linear mapping systems with a structure loosely based on principles
observed in biological nervous systems. Articial neural networks oer many advantages over conventional
statistical methods (Shachmurove, 2002). The ANN uses the data to develop an internal representation of
the relationship between the variables, and does not make assumptions about the nature of the distribution
of the data. Another advantage is that while traditional regression analysis is not adaptive, articial neural
networks readjust their weights as new input data becomes available (Kuo & Reitsch, 1995; Pao, 1989;
Gilbert, Krishnaswamy, & Pashley, 2000). Applications include pattern classication, clustering and
 H.C. Co, R. Boosarawongse / Computers & Industrial Engineering 53 (2007) 610627 619

1200000 1000000
Actual Actual
1000000
Fitted 800000 Fitted
800000 Error
600000 Error
600000
400000
400000

200000
200000

0 0
0 20 40 60 80 100 120 0 20 40 60 80 100 120
-200000 -200000
-400000
-400000
(a) Total Rice Export (b) White & Jasmine Rice Export

70000 350000
Actual
60000 Actual
Fitted 300000
Fitted
50000 Error
250000 Error
40000
200000
30000
150000
20000

10000 100000

0 50000
0 20 40 60 80 100 120
-10000 0
0 20 40 60 80 100 120
-20000
-50000
-30000
-100000
-40000
-150000

(c) Glutinous Rice Export (d) Parboiled Rice Export

Fig. 7. BoxJenkins model-tting errors.

Table 5
Aggregate measures of model tting error (BoxJenkins)
Category MAE MAPE (%) MSE RMSE
Total Rice Export 87,076 14.8 11,629,630,009 107,840
White and jasmine 66,520 14.9 7,403,982,816 86,046
Glutinous 6011 38.4 60,553,081 7782
Parboiled 28,232 24.5 1,344,067,931 36,662

Table 6
Aggregate measures of validation errors (BoxJenkins)
Category MAE MAPE (%) MSE RMSE
Total rice export 268,798 45.3 85,697,961,128 292,742
White and jasmine 192,749 46.9 42,355,763,147 205,805
Glutinous 9982 40.1 164,990,380 12,845
Parboiled 51,137 34.9 3,550,722,984 59,588

categorization, function approximation, prediction and forecasting, optimization, content-addressable

memory, and control of dynamic systems (Jain, Jianchang, & Mohiuddin, 1996). The articial neural
networks have also been found to produce better predictive accuracies than conventional statistical methods
for a variety of problems in industrial engineering, marketing, banking and nance, insurance, and telecom-
munications (Smith & Gupta, 2000).

4.2. A multi-layer perception model

The model we used in this paper is based on a special case of a feed-forward neural network known as a
multi-layer perception (MLP). An MLP neural network is a non-linear, non-parametric regression model
620 H.C. Co, R. Boosarawongse / Computers & Industrial Engineering 53 (2007) 610627

commonly referred to as a function approximator. The articial neural networks used in this study are fully
connected, feed-forward MLP neural networks with three layers: an input layer, a hidden layer and an output
layer.
Fig. 8 illustrates the topology of the neural networks used in this paper. All input nodes (neurons) are
connected to every hidden node and every hidden node is connected to the output nodes. Squares represent
input nodes, and circles depict nodes with sigmoidal non-linearities. The rightmost nodes are the nodes on
the output layer. Triangles represent the bias for each node.

4.3. The input and output layers

The activity of the input nodes represents the raw information that is fed into the network. The perfor-
mance of neural networks is sensitive to the choice of the input nodes. We explored various applications of
neural network in dicult-to-predict situations such as the stock market, bond ratings, commodity and cur-
rency exchange, etc. The literature, particularly on neural-network applications in analyzing the volatility of
stock indices (see for example, Malliaris & Salchenberger (1996) & Freisleben (1992) provided valuable
insights in selecting the input variables).
The basic idea is to model the time series (volume of rice export), in terms of explanatory variables, or
inputs. After experimenting with various network architectures, the following input nodes appeared to per-
form best in both training and validation:

1. Rice export for the current month.

2. Absolute dierence between volume of rice export for current month and that for the previous month.
3. Sign of the dierence between current months volume of rice export and that of the previous month,
encoded as 0.8 for positive dierence, 0.2 for negative and 0 for no change.
4. 3-month moving average of the volume of rice export.
5. Absolute rst dierence of the 3-month (centered) moving average.
6. Encoded sign of the rst dierence of the 3-month (centered) moving average.
7. 12-month moving average of the volume of rice export.
8. Absolute rst dierence of the 12-month (centered) moving average.

Fig. 8. ANN topology.

 H.C. Co, R. Boosarawongse / Computers & Industrial Engineering 53 (2007) 610627 621

9. Encoded sign of the rst dierence of the 12-month (centered) moving average.
10. Seasonal index for the volume of rice export next month.

The rst 9 input nodes track the movement of monthly rice exports. Nodes 4 and 7 are the 3-month and
12-month moving averages, respectively. Nodes 2, 5 and 8 track the absolute value of the dierences, while
nodes 3, 6 and 9 are encoded signs for the direction (increasing or decreasing) of the dierences. The sample
ACF and PACF for the transformed series in Figs. 5 and 6 indicate that rice exports have exhibited some pattern
of seasonality. Input node 10 is the seasonal indices derived from the HoltWinters additive model (see Table 2).
Our neural-network simulation showed that the ANN trained better with the inclusion of Input node 10.
The output layer consists of one node: the volume of rice export the following month.

4.4. The hidden layer

The activity of each hidden node depends on the activities of the input nodes and the weights of the con-
nections (synapses) between the input and the hidden nodes. Likewise, the activity of each output node
depends on the activities of the hidden nodes and the weights of the connections between the hidden and
the output nodes. The weight of each connection represents the strength of the connection.
In this paper, we experimented on various topologies, consisting of one to three hidden layer(s). The liter-
ature claims that a single hidden layer neural network is sucient for a network to be a universal function
approximator (Funahashi, 1989, Hornik, Stinchcombe, & White, 1989, Hornik et al., 1990, Cybenko,
1989). We found this to be true in this paper.
As a universal function approximator, with sucient number of nodes in the hidden layer, any non-linear
function can be approximated by a neural network. However, it is widely known that a large number of hid-
den nodes can lead to overtting of the training data. The ANN tends to memorize the data at hand and per-
formance varies greatly over dierent data sets from an identical population. The ANN may t the training
data perfectly (including the noise) but will fail to predict the test data. On the other hand, ANN will not train
well (large discrepancy between the true target and the expectation of the model output) if the topology is too
simple for the data at hand. A complex topology creates large variance while a simple topology creates large
bias (Geman, Bienenstock, & Doursat, 1992). Variance and bias create classication error.
The connection strengths and the activation function determine when each node is active. Activation
functions commonly used in MLP neural network include linear, logistic sigmoid, hyperbolic tangent, and
Gaussian. Convergence in training depends on the choice of activation function. In MLP neural networks,
the logistic sigmoid function is often used, and is used in this paper. The hidden layer(s) in the network
and the non-linear activation function of nodes allow non-linearity to be introduced into neural networks.
The number of the network layers, hidden nodes and the stopping criteria was determined through trial-
and-errors process because no commonly accepted theory exists for predetermining the optimal number of
nodes in the hidden layer (Tu, 1996). Using 13 hidden nodes, we found our network not capable of learning,
as the high error rates on both testing and training sets indicated. A signicant dierence was noticed when 4
or more nodes were used. Apparently, this is the minimum number of hidden nodes that is required for the
data at hand. All neural networks used in this study consisted of one hidden layer of 5 nodes, as shown in
Fig. 8.

4.5. Training patterns

The ability to learn is a fundamental trait of intelligence. With a suitable number of nodes in the hidden
layer, the MLP can be trained to approximate any non-linear function to any arbitrary degree of accuracy
(White, 1992; Cybenko, 1989).
In a fully connected MLP neural network, all the input nodes are connected to every hidden node and every
hidden node is connected to the output nodes. Training an ANN involves nding the synaptic weight values
such that the output of the model is close to the expected target (i.e., the volume of rice export next month).
Most neural networks contain some form of learning rule to modify the weights of the connections. The
weights are determined to minimize the sum of squared dierences between network outputs and its desired
622 H.C. Co, R. Boosarawongse / Computers & Industrial Engineering 53 (2007) 610627

outputs. In this study, we use error back propagation, which is probably the most used algorithm to train
MLP (Rumelhart, Hinton, & Williams, 1986). It is basically a gradient descent algorithm of the error com-
puted on a suitable learning set.
The monthly export data provided the training and testing patterns for the neural network. For conve-
nience, we transformed the original data by applying the following linear equation:
y  y min
^y t t 5
y max  y min
In (5), yt is the volume of export for month t, and ymin and ymax are the minimum and maximum observed
values for the entire data set. Note that 0 6 ^y t 6 1. Each training pattern consisted of 10 input and 1 output
values. The input and output values are numbers between 0 and 1. The tted value ^y t1 is a non-linear function
of the observed values ^y tj , j = 0, 1, . . . , 11.

4.6. Cross-validation training

Training an ANN involves nding the synaptic weight values such that the output of the model is close to
the expected target (i.e., the volume of rice export next month).
Suppose we are testing the ability of a trained ANN to forecast the volume of export for a partic-
ular month, t + 1. The training data used to update the weights should include data through month t,
but not the data for month t + 1 (the testing pattern). This way, the testing pattern can be used as an
indication of whether or not memorization is taking place. When a neural network memorizes the train-
ing data, it produces acceptable results for the training data, but poor results when tested on unseen
data.
We used cross-validation training to ascertain that the trained ANN can reproduce acceptable results on
unseen future. From the monthly export data for each category of rice, we generated 12 training sets and test-
ing sets. Each set of training data was use to test the output corresponding to each month from January to
December 2005. For example, the rst set of training data consisted of observed values through December
2004, to test the output for January 2005. The second set tests the output for February 2005, using training
data through December 2004 plus the output (predicted value) from the rst training set. Notice that we did
not use the actual observed values in January 2005 because our training period ends in December 2004, and
any observed value beyond this point is unseen by the trained neural network. Similarly, the third set of train-
ing data consisted of observed values through December 2004, plus the output from the rst two training sets,
to test the output for March 2005, etc.

4.7. Stopping criteria

We specied various stopping criteria for terminating the training session. These criteria include (a)
maximum number of epochs to be used for training; one epoch is equivalent to one complete sweep of
the training data set through the network, (b) average root-mean square (RMS) error; the average
RMS error represents the RMS error over all patterns in the training and testing data les. Training will
stop when the total RMS error falls below this value, (c) maximum RMS error, and (d) the percentage of
correctly classied patterns.
In all our training sessions, the percentage of correctly classied patterns was the overriding crite-
rion. A pattern with an RMS training error less than a predetermined error margin is considered
learned. In this study, we set the error margin at 0.125, and the percentage of correctly classied pat-
terns to 95%.

4.8. The training sessions

Fig. 9 shows the plot of the RMS errors during a training session. The training set consisted of data
for total Thai rice export through December 2004, and the testing set is for the month of January 2005.
The darker curve shows the RMS error for the training set. As the ANN learns, the RMS error decreased
 H.C. Co, R. Boosarawongse / Computers & Industrial Engineering 53 (2007) 610627 623

Fig. 9. RMS errors during a training session.

sharply and then tapered o. The training patterns were used to update the weights, until 95% of the pat-
terns in the training set have been classied as learned. The lighter curve shows the RMS error for the
testing set (total export in January 2005). The curve at rst dropped sharply to near 0, bounced up,
and then dropped again. Toward the end of the training session, the RMS error stayed below the
RMS error of the training set.
Not all training sessions were successful. Fig. 10 shows the plot of the RMS errors during another training
session. The testing set is for predicting the export of glutinous rice export for the month of November 2005.
As the ANN learns, the RMS error of the training set decreased sharply and then tapered o. The RMS error
of the testing set dropped initially, but bounced up and continued to rise as the ANN memorized the training
data. It will be shown later that while glutinous rice export accounted for only 4% of rice exports, it is the most
dicult to forecast.

4.9. Results of cross-validation

The cross-validation results for the various categories of rice are shown in Table 7. Except for glutinous
rice, the neural networks appear to perform well. The training sessions for the parboiled category were the
most successful, with an MAPE of about 12%. The other two categories of rice white and jasmine and
total rice export had average forecast errors of 14.4% and 18.8%, respectively.
The worst performance, as noted earlier, was the training of the ANN for glutinous rice. The mean abso-
lute percentage error (MAPE) for the category was about 26%. While the training data was able to produce a
good t, the ANN for the category gave less than satisfactory results when tested on unseen data.
624 H.C. Co, R. Boosarawongse / Computers & Industrial Engineering 53 (2007) 610627

Fig. 10. RMS errors during a training session (ANN memorizing).

Table 7
Aggregate measures of validation errors (neural networks)
Category MAE MAPE (%) MSE RMSE
Total rice export 110,465 18.8 16,514,374,477 128,508
White and jasmine 60,290 14.4 5,101,045,697 71,422
Glutinous 6967 26.2 94,682,202 9730
Parboiled 18,728 12.1 570,637,152 23,888

5. Summary and concluding remarks

We examined the network architecture of articial neural networks in forecasting rice exports from Thai-
land, and compared the performance of articial neural networks with the HoltWinters additive exponential
smoothing model and the BoxJenkins ARIMA model.
The performance of articial neural networks largely depends on the network architecture. Plots of the time
series, and analysis of the sample ACF and PACF of the transformed time series indicated that there is trend
and seasonality. A network topology consisting of 10 input nodes was found to perform well in both training
and validation. One of 10 input nodes tracks the seasonality of the time series, using the seasonality indices
derived from HoltWinters additive exponential smoothing model. The other 9 nodes comprise of three
3-tuples to track the movement of the time series. The idea of using input nodes to encode the direction
(increasing or decreasing) of movement came from a survey of neural-network applications in forecasting,
particularly on stock index prediction.
 H.C. Co, R. Boosarawongse / Computers & Industrial Engineering 53 (2007) 610627 625

Table 8
ANN vs. statistical methods (aggregate validation error measures)
MAE MAPE (%) MSE RMSE
Category: total rice export
Articial neural networks 110,465 18.8 16,514,374,477 128,508
BoxJenkins 268,798 45.3 85,697,961,128 292,742
HoltWinters 207,882 35.1 53,678,782,994 231,687
Category: white and jasmine
Articial neural networks 60,290 14.4 5,101,045,697 71,422
BoxJenkins 192,749 46.9 42,355,763,147 205,805
HoltWinters 194,042 47.4 43,461,583,428 208,474
Category: glutinous
Articial neural networks 6967 26.2 94,682,202 9730
BoxJenkins 9982 40.1 164,990,380 12,845
HoltWinters 9792 39.9 149,546,455 12,229
Category: parboiled
Articial neural networks 18,728 12.1 570,637,152 23,888
BoxJenkins 51,137 34.9 3,550,722,984 59,588
HoltWinters 46,512 31.3 2,943,833,536 54,257

Fig. 11. Comparison of validation MAPE.

The performance of articial neural networks was compared with two traditional time-series forecasting
methods: the HoltWinter additive exponential smoothing model and BoxJenkins model. These two methods
were chosen because of their ability to model trend and seasonal uctuations present in the data. Exponential
smoothing models require estimates of the initial condition. In the HoltWinters additive model, the initial
values of the level, trend, and the 12 monthly seasonal indexes need to be estimated. In this paper, the problem
of determining the initial condition was solved as an optimization problem, with the objective of minimizing
the sum of squared forecast errors. The procedure was carried out on a spreadsheet (MS Excel), and the result-
ing models showed satisfactory goodness of t. Unfortunately, the models did not perform as well in predict-
ing unseen data during validation.
626 H.C. Co, R. Boosarawongse / Computers & Industrial Engineering 53 (2007) 610627

Forecasting is about predicting the unseen future. To produce a meaningful forecast, the forecast errors
must be within acceptable limits. A model may produce good t of historical data, and yet fail to predict
the unseen future. Table 8 summarizes the aggregate validation error measures, and the bar chart in
Fig. 11 compares the validation MAPE of the three forecasting methods. As shown, the articial neural net-
works performed relatively well, with the lowest mean absolute percentage errors in all four categories. The
optimized HoltWinters additive model performed slightly better than BoxJenkins.
The articial neural networks produced better predictive accuracies because they are non-linear mapping
systems. The articial neural networks use the time series data to develop an internal representation of the
relationship between the variables, and do not make assumptions about the nature of the distribution of
the data. Another advantage is that while traditional regression analysis is not adaptive, articial neural net-
works readjust their weights as new input data becomes available. The network architecture similar to the ones
used for predicting rice export may be used to analyze international trade of other commodities such as corn,
coee, etc.
The biggest drawback of the methodology is that articial neural networks are black boxes. It is impos-
sible to gure out how relations in their hidden layers are estimated (Li, 1994; Gilbert et al., 2000).

References

Alon, I., Qi, M., & Sadowski, R. J. (2001). Forecasting aggregate retail sales: A comparison of articial neural networks and traditional
methods. Journal of Retailing and Consumer Services, 8, 147156.
Bergerson, K., & Wunsch, D. C. (1991). A commodity trading model based on a neural network-expert system hybrid. In: Proceedings of
the IEEE international conference on neural networks.
Bowerman, B. L., Oconnell, R. T., & Koehler, A. B. (2004). Forecasting, time series and regression: An applied approach. Belmont, CA:
Thomson Brooks/Cole.
Box, G. E. P., Jenkins, G. M., & Reinsel, G. C. (1994). Time series analysis, forecasting and control (3rd ed.). Englewood Clis, NJ:
Prentice-Hall.
Brown, R. G. (1959). Statistical forecasting for inventory control. New York: McGraw-Hill.
Chateld, C., Koehler, A. B., Ord, J. K., & Snyder, R. D. (2001). A new look at models for exponential smoothing. The Statistician, 50,
147159.
Collopy, F., & Armstrong, J. S. (1992). Rule-based forecasting: Development and validation of an expert systems approach to combining
time series extrapolations. Management Science, 38, 13941414.
Chong, H. L., & Kyoung, C. P. (1992). Prediction of monthly transition of the composition stock price index using recurrent
backpropagation. ICANN92, 2, 16291632, Brighton, UK, 1992.
Cybenko, G. (1989). Approximation by superposition of a sigmoidal function. Mathematics of Control, Signal and Systems, 2, 303314.
Deetae, N. (1991). A comparative study of forecasting techniques on rice, cassava, and mungbean. Farm Gate Prices, Master Thesis.
Thailand: Kasetsart University.
De Gooijer, Jan G., & Hyndman, R.J. (2006). 25 years of time series forecasting. International Journal of Forecasting, Elsevier, 22(3), 443
473.
Dutta, S., & Shekhar, S. (1990). Bond rating: A non-conservative application of neural networks. IEEE International Conference on Neural
Networks.
Fansiri, J. (2004). Forecasting the rice-export price by ARIMA method. Technical report. Thailand: Chiang Mai University.
Freisleben, B. (1992). Stock market prediction with backpropagation networks, industrial engineering applications of articial intelligence
and expert system. 5th International conference, June 1992, Paderborn, Germany, pp. 451460.
Funahashi, K. (1989). On the approximate realization of continuous mappings by neural networks. Neural Networks, 2, 183192.
Gardner, E. S., Jr., & McKenzie, E. (1988). Model identication in exponential smoothing. Journal of the Operational Research Society,
39, 863867.
Geman, S., Bienenstock, E., & Doursat, R. (1992). Neural networks and the bias/variance dilemma. Neural Computation, 4, 158.
Gilbert, E. W., Krishnaswamy, C. R., & Pashley, M. M. (2000). Neural Network Applications in Finance: A Practical Introduction.
Grudnitski, G., & Osburn, L. (1993). Forecasting S and P and gold futures prices: An application of neural networks. The Journal of
Futures Markets, 13(6), 631643.
Hanke, J. E., & Reitsch, A. G. (1995). Business forecasting (5th ed.). Englewood Clis, NJ: Prentice-Hall.
Hann, T. H., & Steurer, E. (1996). Much do about nothing? exchange rate forecasting: Neural networks vs. linear models using monthly
and weekly data. Neurocomputing, 10, 323339.
Holt, C. C. (1957). Forecasting seasonals and trends by exponentially weighted averages. O.N.R. Memorandum 52/1957, Carnegie
Institute of Technology. Reprinted with discussion in 2004. International Journal of Forecasting, 20, 513.
Hornik, K., Stinchcombe, M., & White, H. (1989). Multilayer feed-forward networks. Neural Networks, 2, 359366.
Hornik, K., Stinchcombe, M., & White, H. (1990). Universal approximation of an unknown mapping and its derivatives using multilayer
neural networks. Neural Networks, 3, 551560.
 H.C. Co, R. Boosarawongse / Computers & Industrial Engineering 53 (2007) 610627 627

Hyndman, R. J., Koehler, A. B., Snyder, R. D., & Grose, S. (2002). A state space framework for automatic forecasting using exponential
smoothing methods. International Journal of Forecasting, 18, 439454.
Jain, A. K., Jianchang, M., & Mohiuddin, K. M. (1996). Articial neural networks: A tutorial, 29(3) (March 1996), Special issue: Neural
computing: Companion issue to Spring 1996, IEEE Computational Science & Engineering, pp. 3144, ISSN:0018-9162.
Kerdsomboon, M. (1999). Forecasting of agricultural products and prices. Master Thesis. Thailand: Chulalongkorn University.
Kimoto, T., Asakawa, K., Yoda, M., & Takeoka, M. (1990). Stock market prediction system with modular neural networks. Proceedings
on UCNN I, 16.
Kuan, C.-M., & Liu, T. (1995). Forecasting exchange rates using feed-forward and recurrent neural networks. Journal of Applied
Econometrics, 10, 347364.
Kuo, C., & Reitsch, A. (1995). Neural networks vs. conventional methods of forecasting. Journal of Business Forecasting, Methods and
Systems, 14(4), 1722.
Lapedes, A., & Farber, R. (1987). Nonlinear Signal Processing Neural Information Processing System Natural and Synthetic.
Li, E. Y. (1994). Articial neural networks and their business applications. Information and Management, 1994.
Ljung, G. M., & Box, G. E. P. (1978). On a measure of lack of t in time series models. Biometrika, 65, 297303.
Maliwan, K. (2003). The analysis of price movement of jasmine rice in domestic market. Master Thesis. Thailand: Kasetsart University.
Mahmoud, E. (1984). Accuracy in forecasting: A survey. Journal of Forecasting, 3, 139159.
Malliaris, M., & Salchenberger, L. (1996). Using neural networks to forecast the S&P 100 implied volatility. Neurocomputing, 10(2),
183195.
Murphy, A. H., & Winkler, R. L. (1992). Diagnostic verication of probability forecasts. International Journal of Forecasting, 7, 435455.
Pao, Y. H. (1989). Adaptive pattern recognition and neural networks. Reading, MA: Addison-Wesley Publishing Company, Inc..
Refenes, A. N. (1995). Neural networks in the capital markets. Chichester: John Wiley.
Rumelhart, D. E., Hinton, G. E., & Williams, R. J. (1986). Learning internal representations by error propagation. In D. E. Rumelhart &
J. L. McClelland (Eds.). Parallel Distributed Processing Foundations (Vol. 1, pp. 318362). Cambridge, MA: MIT Press/Bradford
Books.
Satchell, S., & Timmermann, A. (1995). On the optimality of adaptive expectations: Muth revisited. International Journal of Forecasting,
11, 407416.
Sangpattaranate P. (2005). Forecasting of rice prices in Thailand. Master Thesis. Thailand: Kasetsart University.
Schoneburg, E. (1990). Stock price prediction using neural networks: A project report. Neurocomputing, 2, 1727, Seattle, WA, pp. 1289
1293.
Shachmurove, Y. (2002). Applying articial neural networks to business, economics and nance, departments of economics. The City
College of the City University of New York. Available online from http://www.econ.upenn.edu/Centers/CARESS/CARESSpdf/
02-08.pdf.
Smith, K. A. & Gupta, J. N. D. (2000). Neural networks in business: Techniques and applications for the operations researcher. Computers
& Operations Research, 27(1112), 10231044. Available from http://citeseer.ist.psu.edu/smith00neural.html.
Tashman, L., & Kruk, J. M. (1996). The use of protocols to select exponential smoothing procedures: A reconsideration of forecasting
competitions. International Journal of Forecasting, 12, 235253.
Theil, H. (1966). Applied economic forecasting. Amsterdam: North-Holland.
Tu, J. V. (1996). Advantages and disadvantages of using articial neural networks versus logistic regression for predicting medical
outcomes. Journal of Clinical Epidemiology, 49, 12251231. doi:10.1016/S0895-4356(96)00002-9.
Vanichanont, P. (2004). Thai rice: Sustainable life for rice growers. FAO rice conference, Rome, Italy, 1213 February 2004. Available
online from http://www.fao.org/rice2004/en/pdf/pramote.pdf.
Wei, W. S. (1990). Time series analysis. New York: Addison-Wesley.
White, H. (1992). Articial neural networks approximation and learning theory. Cambridge, MA: Blackwell Publishers.
Winters, P. R. (1960). Forecasting sales by exponentially weighted moving averages. Management Science, 6, 324342.
Yao, J. T., Lan, C. L., & Li, Y. L. (2000). Forecasting the KLSE index using neural networks. Omega: International Journal of
Management Science, 28(4), 455466.