Mathematics and Statistics 7(4A): 29-40, 2019 http://www.hrpub.

org
DOI: 10.13189/ms.2019.070705

Application of ARIMAX Model to Forecast Weekly Cocoa

Black Pod Disease Incidence
Ling, A. S. C.1,*, Darmesah, G.2, Chong, K. P.2, Ho, C. M.2

1
Malaysian Cocoa Board, Wisma SEDCO, Locked Bag 211, 88999 Kota Kinabalu, Sabah, Malaysia
2
Faculty of Science and Natural Resources, Universiti Malaysia Sabah, Jalan UMS, 88400 Kota Kinabalu, Sabah, Malaysia

Received July 1, 2019; Revised September 8, 2019; Accepted September 23, 2019
Copyright©2019 by authors, all rights reserved. Authors agree that this article remains permanently open access under the
terms of the Creative Commons Attribution License 4.0 International License

Abstract The losses caused by cocoa black pod parameters such as rainfall, temperature and high humidity
disease around the world exceeded $400 million due to as reported by Thorold [2,3], Dakwa [4], Wood [5] and
inaccurate forecasting of cocoa black pod disease Mpika [6]. Both rainfall and relative humidity has strong
incidence which leads to inappropriate spraying timing. correlation which encourages pathogen sporulation by
The weekly cocoa black pod disease incidence is affected reproduction of zoospores to infect the cocoa pods while
by external factors, such as climatic variables. In order to the optimum temperature is propitious to black pod
overcome this inaccuracy of spraying timing, the symptom development [2-6]. It is important to quantify the
forecasting disease incidence should consider the black pod disease fluctuations due to the real effect of
influencing external factors such as temperature, rainfall climatic parameters. Understanding the effect of the
and relative humidity. The objective of this study is to climatic variables on the cocoa black pod incidence can
develop a Autoregressive Integrated Moving Average with identifying the suitable management options such as
external variables (ARIMAX) model which tries to fungicide spraying and culture practices to control the
account the effects due to the climatic influencing factors, disease incidence under projected climate change scenario.
to forecast the weekly cocoa black pod disease incidence. The time-fractional partial differential equation was
With respect to performance measures, it is found that the widely used in study the mathematical biology including
proposed ARIMAX model improves the traditional the description of plant disease epidemic [7] but need the
Autoregressive Integrated Moving Average (ARIMA) numerical approach to solve the equation. Another
model. The results of this forecasting can provide benefits approach was developing the time series model especially
especially for the development of decision support system auto-regressive moving average (ARIMA) described by
in determine the right timing of action to be taken in Box and Jenkins [8] that widely used in forecasting where
controlling the cocoa black pod disease. it used the historical sequences of observations to do the
forecasting. In agriculture, ARIMA model used to forecast
Keywords ARIMAX, Black Pod, Climate, Cocoa, the annual production of several crop in countries that
Forecasting relied the crop for daily life or economy, for example
production of rice [9], wheat [10], coffee [11] and cocoa
[12].
As for the disease monitoring program, ARIMA model
was used to predict Botrytis cinerea spore concentrations
1. Introduction that caused grey mould in Spain and assist in deciding
Black pod or Phytophthora pod rot is the most number of treatments needed [13]. In cocoa, it is very
economically important and widespread disease of cocoa, useful in forecasting the cocoa black pod incidence to
Theobroma cacao L. in Malaysia. The losses due to understand the effect of previous incidence on the current
Phytophthora exceed $400 million worldwide [1]. Among incidence. However, ARIMA model along can’t quantify
the Phytophthora spesies which attacked the cocoa, the effect of climate variables on the cocoa black pod
Phytophthora palmivora is the most widely distributed in disease incidence and help in decision making process. The
the world and it caused global yield loss of 20-30% and tree key problem is how to incorporate the climate information
deaths of 10% annually [1]. The intensity of black pod into the forecasting process and subsequently into the
disease in cocoa was influenced by numerous climatic decision making process. When an ARIMA model

CITE THIS PAPER

[1] Ling, A. S. C. , Darmesah, G. , Chong, K. P. , Ho, C. M. , "Application of ARIMAX Model to Forecast Weekly Cocoa Black Pod
Disease Incidence," Mathematics and Statistics, Vol. 7, No. 4A, pp. 29 - 40, 2019. DOI: 10.13189/ms.2019.070705.
30 Application of ARIMAX Model to Forecast Weekly Cocoa Black Pod Disease Incidence

includes other time series as input variables, the model is [8]. The protocol used in identifying what autoregressive or
referred to as an ARIMAX [14]. The application of moving average models terms are given below;
ARIMAX model has been used in agriculture, economics a. An ACF with steadily decline and a PACF will cut off
and engineering as ARIMAX model provides better suddenly after p lags indicates an autoregressive
forecast performance than ARIMA model [15-17]. process at lag p, AR(p).
Therefore, the present study was undertaken with the b. An ACF cuts off suddenly after q lags and a PACF
following objectives (i) Developing the univariate with steadily dkecline indicates a moving average
ARIMA cocoa black pod disease forecasting models and process at lag q, MA(q).
(ii) Fitting ARIMAX (climate parameters as regressors)
The ACF and the PACF both exhibiting large spikes that
models and testing the validity of the developed models.
gradually die out indicates that there are both
autoregressive and moving averages processes.
2. Methodology 2.1.3. Estimation of Model Parameters
Given a tentative ARIMA (p, d, q) model as follows;
2.1. Theory of Methods Yt = α1Yt – 1 + α2Yt – 2 + … + αpYt – p + ε (3)
2.1.1. ARIMA Models + β1 εt – 1 + β2 εt – 2 + . . . + βq εt – q
ARIMA models are also known as Box-Jenkins models which can also be written as
which require historical time series data of underlying
α(L)Yt = β(L)εt (4)
variables. There are three stages involved in time series
2
approach, namely process of model identification, where L is the lag operator, LYt = Yt – 1 and L Yt = Yt – 2
parameter estimation and model checking. In the stage of
α(L) = 1 – α1L – α2L2 – . . . αpLp (5)
model identification, the data series is determined whether
the series is stationary before the Box-Jenkins model or β(L) = 1 – β1L – β2L – . . . βqL
2 q
(6)
ARIMA model is developed. A stationary series in The degree of homogeneity d is determined in the
Box-Jenkins model will have constant mean, constant identification, and Yt is a dependent variable representing
variance and constant autocorrelation. For a nonstationary stationary series. This stage is to estimate all unkown
series, a differencing on the non-stationary series one or population parameters, which are α1, α2, αp, and β1, β2, ., βq
more times will be carried out to achieve stationary. For in equation (3). The least squares method is used in
example, when we applied a first difference ∆yt = yt – yt – 1 estimating the model parameters of α1, α2, . . ., αp, and β1,
on non-stationary series to achieved stationary, so y is β2, . . ., βq. The estimation procedure used in the least
called as integrated of order 1 or y ~ (1). squares method involved minimizing the sum of squared
residuals,
2.1.2. Test of Stationary Series of Data
In order to test for non-stationary, the Augmented ∑ t =1ε t2
n where εt =
Dickey-Fuller (ADF) test [18] is used where it test for a β(L)–1 α(L)Yt.
unit root in a time series sample. Given
2.1.4. Diagnostic Checking
yt = a + byt – 1 + μt (1)
Diagnostic checking in time series model is similar to
where μt would be a non-stationary random walk if b = 1. the regression analysis which included testing the
To test whether y has a unit root or non-stationary, we parameters and residual tests. Parameters testing by using
regress the t-test is to check and retain only those estimate
∆yt = a + cyt – 1 + μt (2) parameters αˆ ( L) α �(L) and ˆ whose t – ratios are
β ( L)
where c = b–1 and test hypothesis that c = 0 against c < 0. significantly greater than a predetermined critical value
For non-stationary series data, the differencing function (that is, | t | > 2 at 5% significance level). Then, the residual
will be applied to the data sequence to transform into tests are carried out using the Akaike Information Criterion
stationary series. (AIC) test and the Ljung-Box test or also known as Q
statistics.
2.1.3. Model Identification The AIC test formula is calculated as:
When the stationarity has been addressed, then we need
AIC(p,q) = n log(RSS/n) + 2(p + q) (7)
to identify the order (the p and q) of the autoregressive and
moving average terms. The tools used are the where n is the number of data points (observations) and
autocorrelation function (ACF) and the partial RSS is the residual sums of squares. The model with
autocorrelation function (PACF). The formula for both smaller AIC values is considered to be the best model.
ACF and PACF are available in most of time series book Meanwhile the Ljung–Box test the magnitudes of the
 Mathematics and Statistics 7(4A): 29-40, 2019 31

residuals autocorrelation for significance: where xt represents exogenous variables, β their

H0: The data are independently distributed with no serial coefficients, ϕ(L)yt is an AR model (ϕ1yt – 1 + ϕ2yt – 2 + ϕ3yt – 3
correlation + . . . + ϕpyt – p) and θ(L)εt is the MA model (θ1εt – 1 + θ2εt – 2
Ha: The data are not independently distributed which + θ3εt – 3 + . . . + θqεt – q).
they exhibit serial correlation. The approach used in ARIMAX model was based on the
steps described by Andrews [19] with flowchart shown in
The test statistic is:
Fig. 1. The approach involved two phases where the first
n
ρˆ k2 phase deals with a linear regression model and the second
Q n ( n + 2) ∑
= (8) phase deals with integration of AR and MA terms into a
k =1
n−k
multiple regression model. In the first phase, the linear
regression is used to identify the exogenous variables that
where n is the sample size, ρˆ k2 ρ�2k is the sample are significant. Meanwhile, the second phase deals with
iterative searching process to search the order of ARIMA
autocorrelation at lag k, and h is the number of lags being part of the model. When there is large number of
tested. Under H0, the statistic Q follows a chi-square
exogenous variables to be screened, the stepwise
distribution, χ2 (k – p – q). For significance level α, the
regression process is used in selecting and introducing a
critical region for rejection of the hypothesis of
randomness is Q > χ2 ( α, k – p – q). new exogenous variable into the ARIMAX model. This is
followed by the iteration process of finding newly AR or
2.1.5. ARIMAX Model MA or both terms to re-establish the random pattern of
An ARIMAX model is viewed as a multiple regression residuals in the ARIMAX models.
model with one or more autoregressive (AR) terms and/or
one or more moving average (MA) terms. Autoregressive 2.2. Data Collection and Statistical Analysis
terms for a dependent variable are merely lagged values of
that dependent variable that have a statistically significant The study is conducted in the cocoa mature areas in the
relationship with its most recent value. Moving average Cocoa Research and Development Madai, Sabah with the
terms is nothing more than residuals (i.e., lagged errors) coordinates of latitude and longitude as 4o 47’ 10’’ N and
resulting from previously made estimates. The general 117o 57’ 54” E. The plot’s size of 1 hectare was planted
ARIMAX models are as follows: with cocoa clone of PBC 123. The study plot was divided
Autogressive model with exogenous variables (ARX): into two treatments, i.e. first treatment was the combination
of fungicide, pruning and phytosanitary of diseased pods
yt = ϕ(L)yt + βxt + εt (9)
removed from the field while second treatment was the
Moving average model with exogenous variables combination between fungicide and phytosanitary of
(MAX): diseased pods removed from the field. Each treatment has
yt = βxt + θ(L)εt (10) three replications with 30 trees per replication. The random
Autogressive Moving Average with exogenous sampling method was used to select 10 trees in each
replication for the black pod incidence studied meanwhile
variables model (ARMAX):
the assessment of the black pod incidence was adapted
ϕ(L)yt = βxt + θ(L)εt (11) from Ndoumbe-Nkeng [20] and Ngoh Dooh [21] as
follows:
32 Application of ARIMAX Model to Forecast Weekly Cocoa Black Pod Disease Incidence

Figure 1. ARIMAX model building procedure

 Mathematics and Statistics 7(4A): 29-40, 2019 33

Black pod rate in treatment n (BPTRTn) was calculated validation purpose is 23 observations (or weeks) that
each week i, covered observations from July 2016 to December 2016.
BP *100 BPi *100
BPTRTn i =(YP i+AP )(BPi × 100) / (YPi + APi)
i i (YPi +APi )
(12) 3. Results and Discussion
BPi *100where BPi is the
number of black pods observed over 30 selected trees in 3.1. Descriptive Statistics
week i, YPi is the number of young pods over 10 selected
From Fig. 1 to 4, the black pod incidence and weather
trees in week i and APi is the number of adult pods over 30
data have shown volatile pattern, showing ups and downs
selected trees in week i.
The daily weather variables (rainfall, relative humidity over a period of time except rainfall shows less volatile
and temperature) and the weekly black pod incidence data pattern as less rain occurred during that period. The high
from the period of February 2015 to June 2016 collected. black pod incidence is recorded to be high especially
The weather data were recorded using WatchDog data during the rainfall and relative humidity were high during
loggers (1000 Series Micro Station, Spectrum earlier days which is known as lag effect. The high relative
Technologies, Aurora, IL, USA). The statistical tool used humidity was also associated with a high frequency of
in this paper is EViews version 6.0 to analyse the time rainfall often favorable for the occurrence of the disease as
series data and building the ARIMA and ARIMAX model. the presence of moisture on cocoa trees from rain or dew
The forecasting performance of the best-fitting between provided ideal conditions for P. palmivora infection [4].
ARIMA and ARIMAX models as plant disease forecasting The lag effect of the climate change could be due to the
model are compared using three criteria, namely mean incubation period which is the period from fruit infection to
squared error (MSE), root mean squared error (RMSE) and the first disease symptoms [22].
mean absolute error (MAE). The dataset for model
.14
Treatment 1
.12 Treatment 2
Cocoa black pod incidence rate

.10

.08

.06

.04

.02

.00
5 10 15 20 25 30 35 40 45 50 55 60 65

Sample

Figure 1. The black pod disease incidence

34 Application of ARIMAX Model to Forecast Weekly Cocoa Black Pod Disease Incidence

50

40

Daily rainfall (mm) 30

20

10

0
5 10 15 20 25 30 35 40 45 50 55 60 65

Sample

Figure 2. The daily rainfall (mm) recorded

92

88
Relative Humidity (%)

84

80

76

72

68
5 10 15 20 25 30 35 40 45 50 55 60 65

Sample

Figure 3. The relative humidity (%) recorded

40

36
Temperature (Degree Celcius)

32

28

24

20
5 10 15 20 25 30 35 40 45 50 55 60 65

Sample

Maximum Average Minimum

Figure 4. The temperature (oC) recorded

 Mathematics and Statistics 7(4A): 29-40, 2019 35

3.2. Testing the Stationarity of Black Pod Incidence 3.3. ARIMA Modeling
The cocoa black pod incidence data in both treatments Fig. 5 shows the ACF and PACF plots were cut-off after
as shown in Fig. 1were found to be stationary for both one lag in treatment 1. This suggested the presence of
treatments based on the ADF test or unit root test given in autoregressive models AR(1) through PACF plot and also
Table 1. ADF test successfully rejected hypothesis null of presence of moving average models MA(1) through ACF
unit root at the 5% significant level in both treatments. plot. So, possible ARIMA models to be considered are
ARIMA(1,0,0), ARIMA(1,0,1) and ARIMA(0,0,1). Fig. 6
Table 1. Results of ADF Test on disease incidence data shows the ACF and PACF plots with the partial
Treatment 1 Treatment 2 autocorrelation cofficients cut-off after one lag and the
Test equation autocorrelation coefficients cut-off after two lags in
tADF p value tADF p value
Without
treatment 2.
intercept and -2.255 0.024 -1.943 0.050 This suggested the presence of autoregressive model
trend AR(1) through PACF plot and also presence of moving
Intercept
-3.427 0.013 -5.572 <0.001
average models MA(1) and MA(2) through ACF plot. So,
without Trend possible ARIMA models to be considered are
Intercept With
-3.738 0.027 -5.614 <0.001 ARIMA(1,0,2), ARIMA(1,0,1), ARIMA(0,0,1),
Trend
ARIMA(0,0,2), and ARIMA(1,0,0).

Figure 5. The correlogram of black pod disease incidence in treatment 1

36 Application of ARIMAX Model to Forecast Weekly Cocoa Black Pod Disease Incidence

Figure 6. The correlogram of black pod disease incidence in treatment 2

Table 2. Estimated parameters of ARIMA models in Treatment 1 five ARIMA models that have potentially fitted to dataset
Variable Coefficient Standard Error t p value in treatment 2. Both treatments have two ARIMA models,
Constant 0.00726 0.00448 1.619 0.111
i.e. ARIMA(1,0,0) and ARIMA(0,0,1) with all the
parameters significant at 5% level and need validated for
AR(1) 0.48072 0.32276 1.489 0.141
the best fitted model.
MA(1) 0.08530 0.31566 0.270 0.788 The Ljung-Box Q test in treatment 1 gave the Ljung-Box
Constant 0.00632 0.00247 2.558 0.013 Q statistics of ARIMA(1,0,0) and ARIMA(0,0,1) at lag 24
AR(1) 0.55861 0.12880 4.337 <0.001 as 10.750 and 10.248 with p values greater than 5%
Constant 0.01321 0.00288 4.590 <0.001
significant level. This proved residuals of fitted
ARIMA(1,0,0) and ARIMA(0,0,1) in treatment 1 are
MA(1) 0.48622 0.10938 4.445 <0.001
independently distributed with no serial correlation (Table
Table 3. Estimated parameters of ARIMA models in Treatment 2 4). Meanwhile Ljung-Box Q test in treatment 2 gave the
Ljung-Box Q statistics of ARIMA(1,0,0) and
Variable Coefficient Standard Error t p value ARIMA(0,0,1) at lag 24 as 14.091 and 17.504 with p
Constant 0.00606 0.00516 1.174 0.245 values greater than 5% significant level to show there was
AR(1) 0.82492 0.15791 5.224 <0.001 no autocorrelation among the residuals of fitted
MA(1) -0.58850 0.20200 -2.914 0.005 ARIMA(1,0,0) and ARIMA(0,0,1) (Table 4).
MA(2) -0.18836 0.13235 -1.423 0.160 Table 4. Results of the Ljung-Box Q test on residuals on ARIMA
models
Constant 0.00941 0.00627 1.502 0.138
AR(1) 0.71973 0.19560 3.680 <0.001 Treatment Time series model Ljung-Box Q statistic p value

MA(1) -0.51132 0.24698 -2.070 0.043 ARIMA(1,0,0) 10.750 0.986

1
Constant 0.02101 0.0048 4.340 <0.001 ARIMA(0,0,1) 10.248 0.990

AR(1) 0.33293 0.11972 2.781 0.007 ARIMA(1,0,0) 14.091 0.924

2
Constant 0.02052 0.00560 3.665 0.001 ARIMA(0,0,1) 17.504 0.784

AR(1) 0.30074 0.12818 2.346 0.022 The best fitted ARIMA model based on the AIC test and
AR(2) 0.06131 0.12918 0.474 0.637 BIC test in Table 5 show ARIMA(1,0,0) in treatment 1 has
Constant 0.03055 0.00418 7.301 <0.001 the smallest value of AIC (–5.440) and BIC (–5.373)
compared to ARIMA(0,0,1) with AIC value of –5.430 and
MA(1) 0.28092 0.12070 2.327 0.023
BIC value of –5.363. Meanwhile ARIMA(1,0,0) in
Table 2 shows the parameters estimation for three treatment 2 has the smallest value of AIC (–4.409) and BIC
ARIMA models that have potentially fitted to the dataset in (–4.342) compared to ARIMA(0,0,1) with AIC value of –
treatment 1. Table 3 shows the parameters estimation for 4.385 and BIC value of –4.3193 (Table 5).
 Mathematics and Statistics 7(4A): 29-40, 2019 37

Table 5. Results of the AIC test and BIC test average temperature at time t – 3 (Tmeant–3), maximum
Treatment
Time series AIC BIC
Decision
temperature at time t (Tmaxt) and t – 2 (Tmaxt–2). There
model value value were nine climate variables in treatment 2 have correct
ARIMA(1,0,0) –5.440 –5.373 Selected
1
correlation at 5% significant level (Table 8). The climate
ARIMA(0,0,1) –5.430 –5.363 variables were the total rainfall at time t – 4 (TRainFt – 4)
ARIMA(1,0,0) –4.409 –4.342 Selected and t – 5 (TRainFt–5), average relative humidity at time t – 3
2 (MeanRHt – 3), t – 4 (MeanRHt–4), t – 5 (MeanRHt–5), t – 6
ARIMA(0,0,1) –4.385 –4.319
(MeanRHt–6) and t – 7 (MeanRHt–7), maximum
The selected ARIMA(1,0,0) model for both treatments temperature at time t (Tmaxt) and t – 5 (Tmaxt–5).
explained that the forecast cocoa black pod disease
incidence at time t was depend on the incidence at time t – 1. Table 7. Correlation analysis on climate variables in treatment 1
This was because the cocoa black pod disease incidence Climate variable Correlation coefficient p value
existed in treatments 1 and 2 mostly caused by the
MeanRHt – 6 0.293 0.017
environment as study by Guest [23] that showed under
favorable condition will caused the sporangia from Tmeant – 3 0.243 0.050
diseased pod to develop within 48 hours of infection. Tmaxt -0.392 0.001
Tmaxt – 2 -0.256 0.038
3.4. ARIMAX Modeling
Table 8. Correlation analysis on climate variables in treatment 2
The twelve steps given in the flowchart shown in Fig. 1
was used to build the ARIMA model. Part of the tests have Climate variable Correlation coefficient p value
been carried out in ARIMA modeling. The pair-wise TRainFt – 4 0.265 0.031
Granger Causality tests in Table 6 for treatments 1 and 2 TRainFt– 5 0.283 0.021
showed three null hypotheses of Granger Causality tests
MeanRHt – 3 0.285 0.020
being rejected at 5% significant level to indicate one-way
causal relationships from the exogenous variables to a MeanRHt – 4 0.374 0.002
dependent variable. The exogenous variables in treatment 1 MeanRHt– 5 0.416 <0.001
were the relative humidity at time t (MeanRHt), average MeanRHt – 6 0.438 <0.001
temperature at time t (Tmeant) and maximum temperature MeanRHt– 7 0.355 0.004
at time t (Tmaxt). As for treatment 2, the exogenous
Tmaxt -0.268 0.660
variables were the total rainfall at time t (TRainFt), average
relative humidity at time t (MeanRHt) and maximum Tmaxt – 5 -0.266 0.069
temperature at time t (Tmaxt).
The stepwise regression that develop a linear regression
Table 6. Granger Causality test involved climate variables model between the climate variables and the cocoa black
p value
pod disease incidence showed both Tmeant – 3 and Tmaxt
Null Hypothesis Treatment Treatment
variables have significant F values at 5% level in treatment
1 2 1 while only MeanRHt – 4 in treatment 2 has significant F
TRainFt does not granger cause BPTRTt 0.062 0.007 values at 5% level (Table 9).
BPTRTt does not granger cause TRainFt 0.346 0.407 Table 9. Stepwise regression analysis
MeanRHt does not granger cause BPTRTt 0.050 0.026
Parameter Standard p R
Variable F
BPTRTt does not granger cause MeanRHt 0.944 0.917 Estimate Error value Squared
Tmeant does not granger cause BPTRTt 0.042 0.702 Treatment 1
Constant 0.07897 0.07109 1.23 0.271
BPTRT1t does not granger cause Tmeant 0.345 0.578
Tmeant – 3 0.00492 0.00232 4.50 0.038 0.2892
Tmint does not granger cause BPTRTt 0.474 0.142
Tmaxt -0.00588 0.00123 22.93 <0.001
BPTRTt does not granger cause Tmint 0.094 0.446
Tmaxt does not granger cause BPTRTt <0.001 0.021 Treatment 2
Constant -0.16575 0.05086 10.62 0.002
BPTRTt does not granger cause Tmaxt 0.686 0.832
0.1892
MeanRHt – 4 0.00245 0.00063 14.94 <0.001
It is important to ensure the exogenous variables have
correct correlation sign to the disease incidence. The Table 10 shows the residuals from stepwise regression
Spearman correlation analysis in treatment 1 determined model in treatments1 and 2 rejected the hypothesis null of
four climate variables have correct correlation at 5% unit root at the 5% significant level in ADF test. This
significant level (Table 7). The climate variables were the indicated the residuals from the stepwise regression model
average relative humidity at time t – 6 (MeanRHt–6), in treatments 1 and 2 were stationary.
38 Application of ARIMAX Model to Forecast Weekly Cocoa Black Pod Disease Incidence

Table 10. Results of ADF Test on residuals from stepwise regression proved that the difference in maximum temperature
Treatment 1 Treatment 2 measured outside and inside canopy layers of cocoa trees
Test equation p p caused the maximum temperature to be an important factor
tADF tADF
value value in effecting the cocoa production and also the disease
Without intercept and infection [25-26]. Meanwhile the selection of average
0.656 0.855 -6.120 <0.001
trend relative humidity as one of the predictors in ARIMAX
Intercept without Trend -5.317 <0.001 -6.070 <0.001 model in the treatments 2 due to the control measure
Intercept With Trend -5.756 <0.001 -6.039 <0.001 applied to the cocoa trees were free from pruning with
fungicide application only. The cocoa tree without pruning
Both the Ljung-Box Q statistics in treatments 1 and 2 have dense canopy and caused little sunshine penetration
(Table 11), 12.289 (lag 5) and 4.218 (lag 1) have p value that will create highly relative humidity to favor the
greater than 0.05 which indicated autocorrelation exist in Phytophthora’s inoculums to infect the cocoa pods [5, 27].
the residuals at 5% significant level. This provides an
indication that AR and/or MA terms must be added into the
model to remove the serial correlation. 3.5. Model Validation

Table 11. Results of the Ljung-Box Q test on residuals from stepwise The best cocoa black pod disease forecasting model
regression between ARIMA and ARIMAX model was selected based
Treatment Lags Ljung-Box Q statistic p value
on its forecasting performance using three criteria, namely
mean squared error (MSE), root mean squared error
1 5 12.289 0.0153
(RMSE) and mean absolute error (MAE). Table 13 shows
2 1 4.218 0.0400 ARIMAX model in treatment 1 has the lowest values of
MSE (i.e. 0.00057), RMSE (i.e. 0.02389) and MAE (i.e.
Table 12 shows the ARIMAX model in treatment 1 was
0.01730) as compared to ARIMA model with MSE value
established using the AR(1) term and variable Tmaxt as
of 0.00061, RMSE value of 0.02467 and MAE value of
both were significant at 5% level while ARIMAX model in
0.01773. Meanwhile ARIMAX model in treatment 2 has
treatment 2 involved the AR(1) term and variable the lowest values of MSE k(i.e. 0.00955), RMSE (i.e.
MeanRHt – 4 as both were significant at 5% level. The sign 0.09773) and MAE (i.e. 0.07765) as compared to ARIMA
check on the coefficient of the exogenous variable as given model with MSE value of 0.01003, RMSE value of
in Table 12 correctly correlated with the disease incidence. 0.10013 and MAE value of 0.08139. This showed
Table 12. Fitted ARIMAX models ARIMAX model has performed better compared to
ARIMA model and therefore, ARIMAX model was
Standard p
Treatment Variable Coefficient
Error
t
value selected as the best disease forecasting model in this study.
Constant 0.14900 0.04316 3.452 0.0010
The gap between the actual and ARIMAX predicted values
seen to be far especially on sample with higher black pod
1 AR(1) 0.40777 0.12807 3.184 0.0023
incidence rates as seen in Fig.7 and Fig.8. The reason
Tmaxt -0.00418 0.00126 -3.310 0.0016 probable due to nonlinear effects in cocoa black pod
Constant -0.1563 0.04948 -3.159 0.0024 incidence rate and also influencing external factors that not
AR(1) 0.29427 0.11029 2.668 0.0097 being considered in the study, i.e. duration of pods being
2
MeanRHt exposed to wetness as direct impact to the disease.
0.00222 0.00062 3.598 0.0006
–4
Table 13. Actual versus forecasted values
The selection of maximum temperature as one of the Treatment 1 Treatment 2
predictors in ARIMAX model in treatment 1 was due to its Validation test
ARIMA ARIMAX ARIMA ARIMAX
treatment applied the pruning as one of its control measure
in managing the cocoa black pod disease incidence. The MSE 0.00061 0.00057 0.01003 0.00955
pruning provided good ventilation around and within the RMSE 0.02467 0.02389 0.10013 0.09773
cocoa tree which can reduced the relative humidity to less MAE 0.01773 0.01730 0.08139 0.07765
volatile compared to temperature [24]. There were study
 Mathematics and Statistics 7(4A): 29-40, 2019 39

Black pod incidence rate

.12
Residual
Actual .10
Fitted ARIMAX .08
.06
.04
.02

.06 .00

.04
.02
Residual

.00
-.02
-.04
-.06
10 20 30 40 50 60 70 80

Sample

Figure 7. Plot of forecasted value of ARIMAX model in treatment 1

.25

Black pod incidence rate

Residual
Actual .20
Fitted
.15

.10
.15
.05
.10
.00
Residual

.05 -.05

.00

-.05

-.10
10 20 30 40 50 60 70 80

Sample

Figure 8. Plot of forecasted value of ARIMAX model in treatment 2

4. Conclusions black pod disease.

This study shows that there were an effect of climate

variables to the cocoa black pod incidence. Key findings Acknowledgements
indicate that maximum temperature and relative humidity
have significant correlation with black pod incidence and The author also would like to thank the Director-General,
suggested as indicator in forecasting the cocoa black pod Malaysian Cocoa Board (MCB) and Director of Upstream
incidence. The results in this paper has found that ARIMA Technology Cocoa for permission to publish and
model with climate variables as input series i.e. ARIMAX reviewing this paper.
consistently showed the superiority over ARIMA models
in capturing the percent relative deviations pertaining to
cocoa black pod incidence forecasts in two different
treatment plots. The ARIMAX models performed well REFERENCES
with lower error as compared to the ARIMA models. The
plant disease forecasting models developed from [1] International Cocoa Organization, Online available from
http://www.icco.org/about-cocoa/pest-a-diseases.html
ARIMAX model able to provide benefits especially for the
development of decision support system in determine the [2] C. A. Thorold. The epiphytes of Theobroma cacao in Nigeria
right timing of action to be taken in controlling the cocoa in relation to the incidence of black pod disease
40 Application of ARIMAX Model to Forecast Weekly Cocoa Black Pod Disease Incidence

(Phytophthora palmivora), Journal of Ecology, Vol.40, with Considering Temperature Effects: An Improved
No.1, 125-142, 1952. ARIMAX Model, Mathematical Problems in Engineering
Vol.2015, 2015.
[3] C. A. Thorold. Observations on black pod disease
(Phytophthora palmivora) of cacao in Nigeria, Transactions [18] E. Graham, J. R. Thomas, H. S. James. Efficient tests for an
of the British Mycological Society, Vol.38, 435-452, 1955. autoregressive unit root, Econometrica, Vol.64, No.4,
813-836, 1996.
[4] J. T. Dakwa. The relationship between black pod incidence
and the weather in Ghana, Ghana Journal of Agricultural [19] B. Andrews, M. Dean, R. Swain, C. Cole Building ARIMA
Science, Vol.6, 93-102, 1973. and ARIMAX models for predicting long-term disability
benefit application rates in the public/private sectors,
[5] G. A. R. Wokod. Black pod: Meteorological factors, In: Society of Actuaries, University of Southern Maine, 2013.
Gregory PH (ed) Phytophthora diseases of cocoa, Longman,
London, 1974. [20] M. Ndoumbe-Nkenga, C. Cilasb, E. Nyemba, S. Nyassea, D.
Bieysseb, A. Florib, I. Sachec. Impact of removing diseased
[6] J. Mpika, E. I. Auguste, B. K. Ismael, K. Brou, L. K. Jean, A. pods on cocoa black pod caused by Phytophthora
Kouassi, A. Z. Nicodeme, A. Severin. Effects of climatic megakarya and on cocoa production in Cameroon, Crop
parameters on the expression of the black pod disease on Protection, Vol.23, 415-424, 2004.
Theobroma cacao in Cote d’Ivoire, Journal of Applied
Biosciences, Vol.20, 1183-1193, 2009. [21] J. P. Ngoh Dooh, B. N. Zachée Ambang, W. N. Kuate
Tueguem, H. Alain, N. G. Ntsomboh. Development of
[7] O. A. Arqub, Z. Odibat, M. Al-Smadi. Numerical solutions Cocoa Black Pod Disease (Caused by Phytophthora
of time-fractional partial integrodifferential equations of megakarya) in Cameroon when Treated with Extracts of
Robin functions types in Hibert space with error bounds and Thevetia peruviana or Ridomil, International Journal of
error estimates, International Journal of Numerical Methods Current Research in Biosciences and Plant Biology, Vol.2,
for Heat and Fluid Flow, Vol.28, No.4, 828 – 856, 2018. No.3, 47-59, 2015.
[8] G. E. P. Box, G. M. Jenkins. Time Series Analysis: [22] M. Ndoumbè-Nkeng, M. I. B. Efombagn, S. Nyassé, E.
Forecasting and Control, Holden-Day, San Francisco, 1976. Nyemb, I. Sache, C. Cilas Relationships between cocoa
Phytophthora pod rot disease and climatic variables in
[9] M. Hemavathi, K. Prabakaran. ARIMA Model for Cameroon, Canadian Journal of Plant Pathology, Vol.3,
Forecasting of Area, Production and Productivity of Rice 309-320, 2009.
and Its Growth Status in Thanjavur District of Tamil Nadu,
India, International Journal of Current Microbiology and [23] D. I. Guest. Black pod: diverse pathogens and the greatest
Applied Sciences, Vol.7, No.2, 149-156, 2018. global impact on cocoa yield, Phytopathology, Vol.97,
No.12, 1650-1653, 2007.
[10] M. Amin, M. Amanullah, A. Akbar. Time series modeling
for forecasting wheat production of Pakistan, Journal of [24] C. H. Lee. Planting cocoa challenges and reality in Malaysia,
Animal and Plant Sciences, Vol.24, No.5, 1444-1451, 2014. Online available from http://www.iipm.com.my/ipicex201
4 / docs/2012/oral/PLANTING%20COCOA-%20CHALLE
[11] E. Harris, A. R. Abdul-Aziz, R.K. Avuglah. Modeling NGES %20AND%20REALITY%20IN%20MALAYSIA.p
annual Coffee production in Ghana using ARIMA time df.
series Model, International Journal of Business, Social and
Scientific Research, Vol.2, No.7, 175-186, 2012. [25] R. A. C. Miranda, L. C. E. Milde, A. L. Bichara, S. Cornell
Daily characterisation of air temperature and relative
[12] S. Ankrah, K. A. Nyantakyi, E. Dadey. Modeling the Causal humidity profiles in a cocoa plantation, Pesquisa
Effect of World Cocoa Price on Production of Cocoa in Agropecuária Brasileira, Vol.29, 345-353, 1994.
Ghana, Universal Journal of Agricultural Research, Vol.2,
No.7, 264-271, 2014. [26] A. J. Daymond, P. Hadley. The effects of temperature and
light integral on early vegetative growth and chloroplyll
[13] F. J. Rodríguez-Rajo, V. Jato, M. Fernández-González, M. J. fluorescence of four contrasting genotypes of cacao
Aira. The use of aerobiological methods for forecasting (Theobroma cacao), Annals of Applied Biology, Vol.145,
Botrytis spore concentrations in a vineyard. Grana, Vol.49, 257-262, 2004.
No.1, 56-65, 2010.
[27] J. Flood, D. Guest, K. A. Holmes, P. Keane, B. Padi, E.
[14] A. Pankratz. Forecasting with Dynamic Regression Models, Sulistyowati. Cocoa under attack, In: J. Flood, R. Murphy
Wiley-Interscience, New York, 1991. (eds) Cocoa futures. Chinchina, CO: CABIFEDERACAFE,
[15] B. Yogarajah, C. Elankumaran, R. Vigneswaran. 2004.
Application of ARIMAX Model for Forecasting Paddy
Production in Trincomalee District in Sri Lanka, In: 3rd
International Symposium, South Eastern University of Sri
Lanka, 21-25, 2013.
[16] C. Kongcharoen, T. Kruangpradit. Autoregressive Integrated
Moving Average with Explanatory Variable (ARIMAX)
Model for Thailand Export, In: 33rd International
Symposium on Forecasting, South Korea, South Korea, 1-8,
2013.
[17] H. Cui, X. Peng. Short-Term City Electric Load Forecasting