Vegetable Science (2016) 43 (1) : 62-69

Study on onion price forecast using time series methods

SP Bhardwaj, Ranjit Kumar Paul and KN Singh

Received: July 2015 / Accepted: May 2016

Abstract Introduction
Agricultural production is characterized by risks and Price forecasting has been very important in decision
uncertainties arising largely due to uncertain yields and making at all levels and in different sectors of the
relatively low price elasticity of demand, of most economy. Agriculture is characterized by risks and
commodities. Commodity price movements have a major uncertainties largely due to uncertain yields and relatively
impact on overall macroeconomic performance. Hence, low price elasticity of demand, of most commodities.
commodity-price forecasts are a key input to macroeconomic Decision makers require some information about the
policy planning and formulation. The price volatility in case future and the likelihood of the possible future outcomes.
of Onion is considered to be well known in India. This study Price forecasts are critical to market participant for
has been undertaken to forecast Onion prices before the
making production and marketing decisions and to policy
crop arrival and particularly in the lean periods which
makers who administer commodity programs and assess
witnesses high rise in Onion price. The administration may
the impacts of domestic or international markets.
find enough time period to readjust supply position of
Onions in order at avoid high price situation. The study has Therefore, commodity price movements have a major
been illustrated with the time series data on daily Spot price impact on overall macroeconomic performance of
of Onion in Delhi Azadpur Market from 01 January 2009 to commodity markets. Hence, commodity-price forecasts
30September 2012. This study was undertaken to obtain a are a key input to macroeconomic policy planning and
suitable forecast model for forecasting Onion prices. ARIMA formulation. The price volatility in case of Onion is
(1, 1, 2) model gives reasonable and acceptable forecasts; it considered to be notorious one in India.
does not perform well when there existed volatility in the The literature on price forecasting has focused on two
data series. In this study, GARCH (1, 1) has also been used main groups of linear, single-equation, reduced-form
to forecast prices. The model performs better than ARIMA
econometric models as well as Time Series models. The
(0, 1, 1) because of its ability to capture the volatility by the
first group (Financial Models) includes models which
conditional variance of being non-constant throughout the
time. Vector Auto Regressive (VAR) a multivariate model for
are directly inspired by financial economic theory and it
forecast was also attempted but the performance of the model is based on the market efficiency hypothesis (MHE),
was not improved over GARCH model. The GARCH (1,1) while models belonging to the second group (Structural
was concluded to be a better model than others in Models) consider the effects of commodity market
forecasting price of Onion because the values for test agents and real variables on commodity prices. Reza
statistics calculated using this model were smaller than Moghaddasiand et.al (2008) has used annual farm and
those calculated using other model and also both the AIC guaranteed prices of wheat and rice (as a competitive
and SIC values from GARCH model were smaller and the product) and wheat stock for 1966 to 2006 and the
percent deviation in forecast price from actual price was findings revealed the superiority of time series models
comparatively low in GARCH model. Therefore, it showed (unit root and ARIMA (3,2,5)) for forecasting of wheat
that GARCH is a better model than ARIMA for estimating price. ARIMA models outperformed the structural model
daily prices. in predicting the price of wheat for the period 1966-
Keywords: Price forecast, series methods, onion, ARIMA, 2006. Rangsan Nochai et.al (2006) has studied model
GARCH modeal of forecasting oil palm price of Thailand in three types
of prices as farm price, wholesale price and pure oil
price for the period of five years, 2000-2004. The
Indian Agricultural Statistics Research Institute, Library Avenue, objective of the research was to find an appropriate
Pusa Campus, New Delhi-110012
ARIMA Model for forecasting in three types of oil palm
Emails: [email protected], [email protected]
price by considering the minimum of mean absolute
Vegetable Science, Vol. 43, January - June 2016 63

percentage error (MAPE). The MAPE for each model Fuller (ADF) test was taken as evidence of statioinarity.
was found to be very small. The forecasting technique used for a time series analysis
that contains a trend or seasonal or non-stationary data
Chakriya Bowman et.al (2004) assessed the accuracy
was Auto Regressive Integrated Moving Average
of a number of alternate measures of forecast
(ARIMA) which was considered to be most suitable
performance. The analysis indicated that although
model. The minimum mean absolute percentage errors
judgmental forecasts tend to outperform the model-based
(MAPEs) of forecasting values were used in selecting
forecasts over short horizons of one quarter for several
an adequate model.
commodities, models incorporating futures prices
generally yield superior forecasts over horizons of one Statioinarity Test or Unit Root Test: The most widely
year or longer. When evaluating the ex-post effectiveness used tests for unit roots are Dickey and Fuller (1979)
of forecasts, standard statistical measures were test and the Augmented Dickey Fuller (ADF) test. Both
commonly used. This research focused primarily on are used to test the null hypothesis that the series has
RMSE, which gives a measure of the magnitude of the unit root or non stationary. The DF Test is stated as
average forecast error, as an effectiveness measure. K. follows:
Assis et.al (2010) has compared the forecasting
performances of different time series methods for Yt  µ  ρYt 1  e t ……………(1)
forecasting cocoa bean prices. Four different types of Where µ and  are parameters and et is random term.
univariate time series methods or models were Here the null hypothesis is that H0 :  = 1 indicating that
compared, namely the exponential smoothing, the series is non-stationary.
autoregressive integrated moving average (ARIMA),
generalized autoregressive conditional heteroskedasticity ΔYt  µ  γYt 1  e t …………..(2)
(GARCH) and the mixed ARIMA/GARCH models. The
time series data was became stationary after the first Where  = - 1 & Yt = Yt - Yt-1
order of differencing. Based on the results of the ex- The null hypothesis is H0 :  = 0. The test can be carried
post forecasting (starting from January until December out by performing a test on the estimated . The 
2006), the mixed ARIMA/GARCH model outperformed statistics under the null hypothesis of a unit root does
the exponential smoothing, ARIMA and GARCH models. not follow the conventional t distribution. Dickey and
Liew Khim Sen et.al (2007) had taken up time series Fuller (1979) showed that distribution under null
modeling and forecasting of the Sarawak black pepper hypothesis is non standard and simulated critical values
price. Their empirical results showed that for selected sample size. If the error term et is auto-
Autoregressive Moving Average (ARMA) time series correlated, the equation (2) is modified as
models fit the price series well and they have correctly
predicted the future trend of the price series within the m
sample period of study. Guillermo Benavides (2009) Yt = µ +  Yt – 1 + i   yt – 1 + t ……………(3)
examined the volatility accuracy of volatility forecast i 1
models for the case of corn and wheat futures price
returns. The models applied here were a univariate Where m = number of lagged difference terms required
GARCH, a multivariate ARCH (the BEKK model), an so that the error term t is serially independent. The
option implied and a composite forecast model. The null hypothesis is the same as the DF test, i.e., H0 :  =
results showed that the option implied model is superior 0, implying that Yt is non stationary. When DF test is
to the historical models in terms of accuracy and that applied to models like the equation (3), it is called
the composite forecast model was the most accurate Augmented Dickey Fuller (ADF) test.
one (compared to the alternative models) having the Time Series Models: The price forecasts based on
lowest mean-square-errors. these models are only the non-structural-mechanical
forecasts. Autoregressive integrated moving average
Materials and Methods
(ARIMA) models are a class of linear models that are
The study has been illustrated with the time series data capable of representing stationary as well as non-
on daily Spot price of Onion in Delhi Azadpur Market stationary time series. This approach to forecasting is
from 01 January 2009 to 30 September 2012. The time based on Box and Jenkins (1970) popularly known as
series properties of commodity prices were assessed ARIMA model. The methodology refers to the set of
by performing unit root tests. Rejection of the null procedures for identifying, fitting, and checking ARIMA
hypothesis of a unit root under the Augmented Dickey models with time series data.
 64 Bhardwaj et al. : Study on onion price forecast using time series methods

Non-seasonal Box-Jenkins Models for Stationary

Wt  Δ d Yt  (1  B) d Yt , ——————————(8)
Series:
(1) A pth –order autoregressive model: AR(p), which ARIMA (p,d,q) has the general form:
has the general form:
ψ p (B)(1  B)d Yt  µ  θ q (B)ε t or
Yt  α 0  α1Yt 1  α 2 Yt  2  ...............  α p Yt  p  ε t
ψ p (B)Wt  µ  θ q (B)ε t ————(9)
———(4)
Yt = Response (dependent) variable at time t Model Checking: In this step, model must be checked
for adequacy by considering the properties of the
Yt 1 , Yt  2, ........., Yt  p = Response variable at time lags residuals whether the residuals from an ARIMA model
must have the normal distribution and should be random.
t-1, t-2, ……., t-p, respectively.
An overall check of model adequacy is provided by the
α 0 , α1 , α 2 ,.........., α p  Coefficients to be estimated, Ljung-Box Q statistic. The test statistic Q is as follows-
t = Error term at time t. m
rk2 (e)
(2) A qth-order moving average model: MA(q), Q m  n(n  2)  ~ X 2m  r
k 1
n  k
which has the general form:
Where rk(e) = the residual autocorrelation at lag k, n=
Yt  µ  ε t  θ1ε t 1  θ 2 ε t  2  .................  θ q ε t  q
the number of residuals, m= the number of time lags
———————(5) included in the test. If the p-value associated with the Q
Where, Yt = Response (dependent) variable at time t statistic is small (p-value < ), the model is considered
inadequate.
µ = Constant mean of the process, θ 1 , θ 2 ,......... .., θ q  GARCH Method: In econometrics, Auto Regressive
Coefficients to be estimated Conditional Heteroskedasticity (ARCH) models are used
to characterize and model observed time series. They
ε t  Error term at time t., ε t 1 , ε t  2 ,..........., ε t  q  are used whenever there is reason to believe that, at any
Errors in previous time periods that are incorporated in point in a series, the terms will have a characteristic
the response Yt. size, or variance. In particular ARCH models assume
the variance of the current error term to be a function
(3) Autoregressive Moving Average Model:
of the actual sizes of the previous time periods’ error
ARMA(p,q), which has general form:
terms: often the variance is related to the squares of the
Yt  ψ 0  ψ1Yt 1  ψ 2 Yt  2  ....................  ψ p Yt  p  previous innovations. Such models are often
called ARCH models (Engle, 1982), although a variety
ε t  θ1ε t 1  θ 2 ε t  2  ...........  θ q ε t  q —————(6) of other acronyms are applied to particular structures
of model which have a similar basis. ARCH models are
ARIMA model-building: According to equation (5), a employed commonly in modeling financial time
highly useful operator in time-series theory is the lag or series that exhibit time-varying volatility clustering, i.e.
backward linear operator (B) defined by BYt = Yt-1 periods of swings followed by periods of relative calm.
Model for non-seasonal series are called Autoregressive If an autoregressive moving average model (ARMA
integrated moving average model, denoted by ARIMA model) is assumed for the error variance, the model is a
(p, d, q). Here p indicates the order of the autoregressive Generalized Autoregressive Conditional
part, d indicates the amount of differencing, and q Heteroskedasticity (GARCH, Bollerslev (1986)) model.
indicates the order of the moving average part. If the The GARCH (1, 1) Model Specification: To measure
original series is stationary, d = 0 and the ARIMA models the extent of price volatility, GARCH (1, 1) Model has
reduce to the ARMA models. The difference linear been applied in the study
operator (), defined by-
Yt  X t   t ——–––——————————(10)
ΔYt  Yt  Yt 1  Yt  BYt  (1  B)Yt ————(7)
The stationary series Wt obtained as the dth difference Yt     t – 1    2t – 1 —–––––——————(11)

( Δ d ) of Yt The mean equation given in equation (10) is written as

a function of exogenous variables with an error term.
Vegetable Science, Vol. 43, January - June 2016 65

(1,1) process. The autoregressive root which governs

Since  2t is the one-period ahead forecast variance based
the persistence of volatility shocks is the sum of  plus
on past information, it is called the conditional variance. . The ARCH parameters corresponds to  and
The conditional variance equation specified in equation GARCH parameters to . If the sum of ARCH and
(11) is a function of three terms: A constant term: . , GARCH coefficients close to 1, indicating that volatility
the volatility from the previous period, measured as the shocks are quite persistent.
lag of the squared residual from the mean equation:
Vector Autoregressive (VAR) process: Let us
t – 1 (the ARCH term), Last period’s forecast variance: consider a univariate time series yt, t=1,2,…,T arising
from the model
 2t – 1 (the GARCH term).
The (1, 1) in GARCH (1, 1) refers to the presence of a y t    1 y t-1  2 yt-1  .... k y t-k  u t , u t ~ IN (0,)
first-order autoregressive GARCH term (the first term (14)
in parentheses) and a first-order moving average ARCH where, ut is a sequence of uncorrelated error terms and
term (the second term in parentheses). An ordinary
 j, j=1,…,k are the constant parameters. This is a
ARCH model is a special case of a GARCH specification sequentially defined model; yt is generated as a function
in which there are no lagged forecast variances in the of its own past values. This is a standard autoregressive
conditional variance equation-i.e., a GARCH (0, 1).
framework or AR(k), where k is the order of the
This specification is often interpreted in a financial autoregression.
context, where an agent or trader predicts this period’s If a multiple time series yt of n endogenous variables is
variance by forming a weighted average of a long term considered, the extension of (1) will give the VAR(k)
average (the constant), the forecasted variance from model (VAR model of order k), i.e. it is possible to
last period (the GARCH term), and information about specify the following data generating procedure and
volatility observed in the previous period (the ARCH model yt as an unrestricted VAR involving up to k lags
term). If the asset return was unexpectedly large in either of yt,
the upward or the downward direction, then the trader
will increase the estimate of the variance for the next y t    A1 y t -1  .....  A k y t -k  u t , u t ~ IN (0,  )
period. This model is also consistent with the volatility (15)
clustering often seen in financial returns data, where
large changes in returns are likely to be followed by where, yt =(y1t, y2t,…, ynt) is (n×1) random vector,
further large changes. There are two equivalent each of the Ai is an (n×n) matrix of parameters,  is a
representations of the variance equation that may aid fixed (n×1) vector of intercept terms. Finally,
you in interpreting the model: ut=(u1t,u2t,…,unt) is a n-dimensional white noise or
innovation process, i.e., E(u t)=0, E(utu t)= and
(1) If we recursively substitute for the lagged E(utus)=0 for st. The covariance matrix  is assumed
variance on the right hand side of equation(11) we can to be non-singular. Using lag operator (L) (2) can be
express the conditional variance as a weighted average
of all the lagged squared residuals: written as, (I n - A1 L - ... - A k Lk ) y t    u t .

 The process yt is said to be stable if the roots of the


 2t    j –1
2t – j ——————— (12) polynomial, | I n - A1 L - ... - A k Lk |  0 lie outside the
(1 – ) j1
complex unit circle i.e. have modulus greater than one.
We can see that the GARCH (1, 1) variance specification Diagnostic Measures
is analogous to the sample variance, but it down-weights
more distant lagged squared errors. Information criteria: In statistics, the Bayesian
(2) The error in the squared returns is given by information criterion (BIC) or Schwarz criterion
(also SBC, SBIC) is a criterion for model
v t  2t
–  2t . Substituting for the variance in the selection among a finite set of models. It is based, in
variance equation and rearranging terms we can write part, on the likelihood function, and it is closely related
our model in terms of the errors: to Akaike information criterion (AIC). While fitting a
model, it is possible to increase the likelihood by adding
2t    (  ) 2t – 1   t –  v t – 1 —————(13) parameters, but doing so may result in over fitting. The
BIC resolves this problem by introducing a penalty term
Thus, the squared error follow a hetroskedastic ARMA
 66 Bhardwaj et al. : Study on onion price forecast using time series methods

for the number of parameters in the model. The penalty The RMSE is similar to MAE. The MAE and RMSE
term is larger in BIC than in AIC. The BIC was developed depend on the scale of the dependent variable. These
by Gideon E. Schwarz, who gave a Bayesian argument should be used as relative measures to compare forecasts
for adopting it. It is closely related to the Akaike for the same series across different models.
information criterion (AIC). In fact, Akaike was so
The relative mean absolute prediction error (RMAPE)
impressed with Schwarz’s Bayesian formalism that he
is calculated using the following formula
developed his own Bayesian formalism, now often
referred to as the ABIC for “a Bayesian Information T h
Criterion” or more casually “Akaike’s Bayesian ŷ t  y t
Information Criterion”.
RMAPE = 100  yt
/h
t  T 1
The statistical measures of fit called information criteria. The RMAPE calculates the forecast error as a percentage
Let: n = number of observations (e.g. data values, of actual value.
frequencies), k = number of parameters to be estimated
(e.g. the Normal distribution has 2: mu and sigma), Results and Discussion
Lmax = the maximized value of the log-Likelihood for
the estimated model (i.e. fit the parameters by MLE and Unit Root Test: Augmented Dickey Fuller (ADF) test
record the natural log of the Likelihood.) was applied to the Spot price series data to test the null
hypothesis that the series has unit root or non stationary.
SIC (Schwarz information criterion, aka Bayesian The results are given in Table-1. The ‘ -Statistics’
information criterion BIC) obtained for all the price series is significant and greater
SIC = ln[n]k – 2ln [Lmax] than at 1 percent level, the null hypothesis of series has
unit root or non stationary data series cannot be rejected.
AIC (Akaike information criterion)
The alternative hypothesis is true. Thus data series is
 2n  subjected to first differencing to make the data stationary.
A IC    k – 2 ln [L max ] The results of differenced series indicated that the ‘ -
 n – k – 1
Statistic’ obtained for price series is not significant
The aim is to find the model with the lowest value of and less than at 1 percent level, we are bound to reject
the selected information criterion. the null hypothesis and the alternative hypothesis of
Absolute Accuracy Performance Measures of stationary series and no unit root is true. The data series
Forecast: The absolute accuracy analysis is the statistic, became stationary at one differencing and the data is
mean squared error (MSE), defined as: MSE = now ready for further econometric analysis. In Table -
2 Augmented Dickey Fuller Test for Quantity Arrival of
 (ŷ t – y t ) 2 , Where y t
and ŷ t are the actual and Onion Delhi Market showed that the series is stationary
forecast values, respectively. MSE is considered as a at current level.
“non-parametric” statistic that indicates the size of the Estimation equation of ARIMA (1, 1,2): Model for
individual forecast errors from actual values. The square non-seasonal series are called Autoregressive integrated
root of MSE, called the root mean squared error (RMSE) moving average model, denoted by ARIMA ( p, d, q).
represents the mean size of forecast error, measured in
the same units as the actual values Table 1: Augmented Dickey Fuller Test for spot market price
of onion Delhi market
Th
Level Data At First Difference
MSE =  (ŷ t  y t ) 2 t-Statistic Prob.* t-Statistic Prob.*
t  T 1 ADF Test value -3.05347 0.1182 -41.3835 0.00
1% level -3.96613 -3.96613
5% level -3.41377 -3.41377
Th 10% level -3.12895 -3.12895
RMSE =  (ŷ t  y t ) 2 Table 2: Augmented Dickey Fuller Test for quantity arrival
t  T 1
of onion Delhi market
The absolute size of the errors the mean absolute Level Data
forecast error (MAE) is used: t-Statistic Prob.*
ADF Test value -6.2229 0.00
1% level -3.43593
Th
5% level -2.86389
MAE =  ŷ t  y t /h 10% level -2.56807
t  T 1
Vegetable Science, Vol. 43, January - June 2016 67

Here p indicates the order of the autoregressive part, d it takes a long time to change.
indicates the amount of differencing, and q indicates
Parameter estimation of Vector Autoregressive
the order of the moving average part. If the original
(VAR) Model: In Table-5 the coefficient of price
series is stationary, d = 0 and the ARIMA models reduce
variable (-0.20834) and for quantity (-0.13796) both
to the ARMA models. Estimate the parameters for a
the variables used in the model are statistically significant
tentative model has been selected on the basis of
as evident from t value. The lag quantity arrival and lag
significance level of AR and MA terms as given in Table-
prices of onion in the mandi influence the forecasts of
3. In this particular case both moving average term and
onion prices to some extent.
autoregressive terms was found statistically significant.
Evaluation forecast Performances of forecast
Parameter Estimation GARCH (1, 1) Model: In Table
Models
4, the conditional mean equation, the parameter found
is  = -50.5779 and one statistically significant AR term Information criterion: The AIC and SIC values are
(-0.07402). While the conditional variance equation gives obtained from equation estimation from both ARIMA
= 169872.4 1= 0.32953and a high value of 1 = and GARCH models using E-Views and given in Table-
0.563632A which implied that volatility is persistent and 6. We found that both the AIC and SIC values from
GARCH model are smaller than that from ARIMA model.
Table-3 Parameter estimation of ARIMA (1,1,2) Therefore, it shows that GARCH is a better model than
Variable Coefficient Std. Error t-Statistic Prob. other models for estimating daily prices
C -4.94758 24.11265 -0.20519 0.8375
AR(1) -0.22063 0.029654 -7.44018 0 Forecast Performance: In the forecasting stage, we
MA(2) -0.06963 0.03033 -2.29573 0.0219 calculate RMSE, MSE and MAE and RMAPE values
R-squared 0.045312 Mean dependent var -4.8594
Adjusted R-squared 0.04363 S.D. dependent var 1091.123 from different models. These are tabulated in Table-7.
S.E. of regression 1067.055 Akaike info criterion 16.78583 If the actual values and forecast values are closer to
Sum squared resid 1.29E+09 Schwarz criterion 16.7991
Log likelihood -9548.14 Hannan-Quinn criter. 16.79084
each other, a small forecast error will be obtained. Thus,
F-statistic 26.93512 Durbin-Watson stat 1.994544
Prob(F-statistic) 0 Table 5: Parameter estimation of Vector Autoregressive
Inverted AR Roots -0.22 (VAR) Model
Inverted MA Roots 0.26 -0.26
Variable Coefficient Std. Error t-Statistic Prob.
C 136.0968 64.91865 2.09642 0.0363
Table 4: Parameter estimation of GARCH (1, 1) DPRICERSPERTONE(-1) -0.20834 0.029069 -7.16727 0
Variable Coefficient Std. Error z-Statistic Prob. QTYARRIVTONE(-1) -0.13796 0.055088 -2.50443 0.0124
C -50.5779 29.88697 -1.69231 0.0906 R-squared 0.046421 Mean dependent var -4.8594
AR(1) -0.07402 0.060618 -1.2211 0.222 Adjusted R-squared 0.044741 S.D. dependent var 1091.123
Variance Equation S.E. of regression 1066.435 Akaike info criterion 16.78466
C 169872.4 15899.79 10.68394 0 Sum squared resid 1.29E+09 Schwarz criterion 16.79794
RESID(-1)^2 0.32953 0.04783 6.889595 0 Log likelihood -9547.47 Hannan-Quinn criter. 16.78968
GARCH(-1) 0.563632 0.039367 14.31728 0 F-statistic 27.62633 Durbin-Watson stat 2.023037
R-squared 0.022527 Mean dependent var -4.8594 Prob(F-statistic) 0
Adjusted R-squared 0.021666 S.D. dependent var 1091.123
S.E. of regression 1079.238 Akaike info criterion 16.14087 Table 6: Information criterion for different models
Sum squared resid 1.32E+09 Schwarz criterion 16.163
Model AIC SIC
Log likelihood -9179.15 Hannan-Quinn criter. 16.14923
F-statistic 6.544996 Durbin-Watson stat 2.262935 ARIMA 16.78583 16.7991
Prob(F-statistic) 0.000033 GARCH 16.14087 16.163
Inverted AR Roots -0.07 VAR 16.78466 16.79794

Table 7: Forecast performance of different forecast methods.

Forecast Forecast Days
Test statitics Models 5 10 15 20 30 45 60
MAE ARIMA 41.28 39.28 43.12 41.66 60.32 46.21 43.63
GARCH 64.68 53.52 55.47 54.59 51.55 57.03 53.10
VAR 80.80 73.70 78.06 78.42 87.37 85.31 78.86
ARIMA 0.01 0.01 0.01 0.01 0.01 0.01 0.01
GARCH 0.01 0.00 0.00 0.00 0.00 0.01 0.00
RMAPE VAR 0.01 0.01 0.01 0.01 0.01 0.01 0.01
MSE ARIMA 2300.92 2274.23 2653.59 2487.82 8958.32 6138.25 5169.01
GARCH 4306.60 3203.16 4007.69 3730.37 3606.98 4213.95 3665.52
VAR 8010.80 7481.28 7992.94 7792.23 12274.14 10901.32 9650.13
RMSE ARIMA 47.97 47.69 51.51 49.88 94.65 78.35 71.90
GARCH 65.62 56.60 63.31 61.08 60.06 64.91 60.54
VAR 89.50 86.49 89.40 88.27 110.79 104.41 98.24
 68 Bhardwaj et al. : Study on onion price forecast using time series methods

Table 8: Difference in actual and forecast price of onion

Forecast Actual Forecast Price in different models Percent deviations from actual price
Day Price ARIMA GARCH VAR ARIMA GARCH VAR
1 7000 6974.15 6945.68 7098.97 0.37 0.78 1.41
2 7080 7051.53 7019.76 6995.67 0.74 0.28 0.06
3 7100 7082.18 7044.20 7082.63 1.17 0.63 1.18
4 7100 7090.59 7045.68 7211.28 1.29 0.65 3.02
5 7040 7045.96 6990.12 7020.20 0.66 0.14 0.29
6 7130 7107.63 7069.02 7068.93 1.54 0.99 0.98
7 7130 7118.11 7075.68 7165.80 1.69 1.08 2.37
8 6900 6943.15 6862.70 6990.24 0.81 1.96 0.14
9 6980 6971.50 6919.76 6899.96 0.41 1.15 1.43
10 6980 6971.39 6925.68 7060.66 0.41 1.06 0.87
11 7170 7121.45 7101.61 7040.87 1.73 1.45 0.58
12 7170 7150.13 7115.68 7116.30 2.14 1.65 1.66
13 7170 7160.58 7115.68 7055.24 2.29 1.65 0.79
14 7170 7162.58 7115.68 7097.88 2.32 1.65 1.40
15 7150 7147.72 7097.16 6981.61 2.11 1.39 0.26
16 7150 7144.84 7095.68 7219.64 2.07 1.37 3.14
17 7380 7323.06 7308.65 7272.66 4.62 4.41 3.90
18 7380 7357.59 7325.68 7335.42 5.11 4.65 4.79
19 6920 7011.49 6899.73 7029.33 0.16 1.43 0.42
20 6830 6874.29 6782.34 6841.01 1.80 3.11 2.27
21 6830 6836.60 6775.68 6804.87 2.33 3.20 2.79
22 6600 6647.79 6562.70 6745.50 5.03 6.25 3.64
23 7090 6992.32 6999.41 6937.09 0.11 0.01 0.90
24 6630 6694.66 6609.73 6710.80 4.36 5.58 4.13
25 6820 6797.27 6751.61 6800.21 2.90 3.55 2.85
26 6800 6789.65 6747.16 6764.50 3.01 3.61 3.36
27 6800 6793.77 6745.68 6801.91 2.95 3.63 2.83
28 6800 6793.24 6745.68 6862.91 2.95 3.63 1.96
29 6520 6575.30 6486.40 6493.66 6.07 7.34 7.23
30 6800 6800.00 6800.00 6800.00 2.86 2.86 2.86

smaller RMSE, MAE and RMAPE values are preferred. Lkkjka'k

Most of the forecast errors from GARCH model are
smaller than that from other model. Therefore, we can mRiknu dks tksf[ke ,oa vkfuf”prrkvksa ls pfj=hdj.k fd;k tkrk
conclude that GARCH model performs better than other gS tks eq[;r% fuf”pr mit rFkk ekax ds vuq:i de dher yksp
two models. In other words, GARCH is a better forecast okyh vf/kdka”k oLrq,a gksrh gSA oLrq ewY; xfrf”kyrk dk eq[;
model for daily prices of onion. çHkko lEiw.kZ o O;kid vkfFkZd çn”kZu ij iM+rk gSA vr% oLrq
This study was undertaken to obtain a suitable models
ewY; iqokZuqeku O;kid vkfFkZd fufr ;kstuk rFkk dk;kZUo;u dk
for forecasting Onion prices. In this study, the model ,d egRoiw.kZ fuos”k gSA I;kt esa ewY; vfLFkjrk Hkkjr esa iwjh rjg
that has been selected for forecasting onion prices is Kkr gSA ;g v/;;u I;kt dk cktkj esa igq¡pus ls iwoZ ewY;
ARIMA (1, 1, 2). This model gave reasonable and fuèkkZj.k djuk gS rFkk fo”ks’kr% de vken vof/k tcfd I;kt dk
acceptable forecasts; it did not perform very well when T;knk ewY; dk lk{kh curk gSA “kklu çca/k dks çpqj le; vkiwfrZ
there exists volatility in the data series. In this study, dks lek;ksftr djus ds fy, fey tkrk gS ftlls vf/kd ewY; dh
GARCH (1, 1) has also been used to forecast prices. fLFkfr ls cpk tk ldrk gSA v/;;u le;] Ük`a[kyk vkadM+k ds
lkFk lfp= çfrfnu vktkniqj ekdsZV] ubZ fnYyh esa 1 tuojh
In Vector Autoregressive (VAR) Model, the lag quantity
2009 ls 30 flrEcj 2012 rd fy;k x;kA ,d mi;qDr iqokZuqeku
arrival and lag prices of onion in the mandi influence
vkn”kZ v/;;u I;kt dher Kkr djus ds fy, fd;k x;kA , vkj
the forecasts of onion prices to some extent. The
GARCH(1,1) was concluded to be a better model than vkbZ ,e , ¼1]1]2½ vkn”kZ us mfpr rFkk Lohdk;Z iwokZuqeku vkadM+k
other models in forecasting spot price of Onion because Øe dh vfLFkjrk dh otg ls ;g mfpr fu’iknu ugha fn;kA bl
the percent deviation in forecast values from actual v/;;u esa th , vkj lh ,p ¼1]1½ dk Hkh mi;ksx ewY; iwokZuqeku
values were smaller in GARCH model. Also both the ds fy, fd;k x;kA ;g vkn”kZ , vkj vkbZ ,e , ¼0]1]1½ dh
AIC and SIC values from GARCH model were smaller rqyuk esa vPNk Fkk D;ksafd l”krZ fopj.k dh vfLFkjrk dks idM+us
than that obtained from other model. Therefore, it dh {kerk vPNh Fkh tks vfLFkj :i ls lEiw.kZ le; rd FkkA
showed that GARCH is a better model for estimating osDVc vkVks fjxzsflo ¼oh , vkj½ eYVhosfj;sV ekMy dk Hkh mi;ksx
daily prices. iwokZuqeku ds fy, fd;k x;k ysfdu vkn”kZ dk fuLiknu th , vkj
Vegetable Science, Vol. 43, January - June 2016 69

lh ,p ds Åij lq/kjk gqvk ugha ik;k x;kA fu’d’kZ ds rkSj ij and Forecasting of Sarawak Black Pepper Price. MPRA
th , vkj lh ,p ¼1]1½ dh vU; dh rqyuk esa I;kt ds ewY; ds Paper No. 791, posted 07 November, pp16.
iwokZuqeku ds fy, mŸke ik;k x;k D;ksafd ijh{k.k lkaf[;dh; Nochai R and Nochai T (2006) Arima model for forecasting oil
lax.kd gsrq ;g vknZ”k fu;e FkkA vU; ekMy rFkk , vkbZ lh o palm price. Paper presented in IMT- GT regional conference
on Mathematics. Statistics and Applications. Universiti
,l vkbZ lh tks th , vkj lh ,p ekMy ls çkIr gq, os NksVs Fks Sains Malasyia, Penang June 13-15.
rFkk iwokZuqeku ewY; ls okLrfod ewY; ds rqyukRed :i ls th
Bowman C and Aasim M. Husain (2004) Forecasting commodity
, vkj lh ,p ekMy esa de fofo/krk çfr”kr FkhA bl çdkj ;g prices: futures versus judgment. IMF Working Paper. WP/
Li’V gksrk gS fd çfrfnu ewY; fu/kkZj.k esa ,-vkj-vkbZ-,e-,- dh 04/41.March.
rqyuk esa th-,-vkj-lh-,p ,d mŸke ekMy gSA Bollerslev T (1986) Generalised autoregressive conditional
heteroskedasticity. J Econometrics 31: 307-327.
Reference Engle R (1982) Autoregressive conditional heteroscedasticity with
Assis K, Amran A, Remali Y and Affendy H (2010) A comparison estimates of the variance of united kingdom inflation.
of univariate time series methods for forecasting cocoa bean Econometrica 50: 987–1007.
prices. Trends Agric Eco 3: 207-215. Dickey DA and Fuller WA (1979) Distribution of the estimators
Moghaddasiand R and Badr BR (2008) An econometric model for for autoregressive time series with a unit root. J Am Statis
wheat price forecasting in Iran. Paper presented in Assoc 74: 427-431.
International Conference on Applied Economics–ICOAE. Box G and Jenkins G (1970) Time series analysis: forecasting and
Sen LK, Shitan M and Hussain H (2007) Time Series Modelling control. San Francisco: Holden-Day.