Ecuadorian flower exports and their seasonality, an econometric

analysis using GARCH modeling

INTRODUCTION

The seasonality condition allows identifying certain patterns that are repetitive
for some variables such as exports (X) of natural flowers, sale of toys, seasonal
illnesses due to climate changes, etc. The flowers of Ecuador are one of the
best in the world, its characteristics are mainly: "thick stems and large
extension, large buds and bright colors. In addition, the great distinguishing
feature of the Ecuadorian rose is its long vase life after cutting. (Corporación
Financiera Nacional [CFN], 2017) and among the varieties are: Gysophilia,
carnation, roses, lilies and others. According to the ISIC (International Standard
Industrial Classification), flowers fall under code A0119.03 Growing of flowers,
including the production of cut flowers and flower buds.(National Institute of
Statistics and Census [INEC], 2012).This classification is used by the Central
Bank of Ecuador (Banco Central del Ecuador [BCE], n. d.) to prepare the
statistics used in this research.
According to the National Finance Corporation (CFN), in 2017 there were 204
companies whose economic activity was the cultivation of flowers located
mostly in Pichincha with 77%, in 2018 of the 208 companies, 74% are also
located in this same province. In 2019 of the 181 companies registered with the
Superintendence of Companies, 74% remain in Pichincha and according to
2020 of the 237 companies, 73% continue their economic activity in Pichincha,
followed by Cotopaxi as the most representative. The industrialization process
of this sector demands important sources of economic resources and, as we
have seen, there is a high concentration of companies in the highlands due to
the high (Henao et al., 2022) of companies in the highlands due to climatic
conditions and irrigation water. (Mena et al., 2017) Table 1 below shows the
classification of the companies dedicated to this activity:
Table 1.
Classification of companies by size
Source: (National Finance Corporation, n.d.)
Prepared by: Authors

As can be seen in Table 2, in the last four years the highest number of
employees is recorded in the year 2018 with 33,494 people whose labor is used
in the flower sector, then in the year 2017 with 29,867, to the year 2020 with
28,775 and finally to the year 2019 were 25,177 employees.
Table 2.
Classification of companies by number of employees
Source: (National Finance Corporation, n.d.)
 Prepared by: Authors

With respect to production and harvested area, its evolution over the last five
years can be seen in Illustration 1 below, taking into account that the rose is the
species most sold in the local and international market.
Illustration 1.
Source: (National Finance Corporation, n.d.)
Flowers are a non-traditional export product for Ecuador, whose figures for the
last 13 years, including the current year 2022, are shown in Illustration 2, which
highlights that in 2021 they grew by 12% compared to 2020, when flower
exports fell to $827,142.1 as a result of the decrease in demand for many
exportable products due to the effects of the COVID-19 pandemic.
Illustration 2.
Source: Central Bank of Ecuador (BCE)
Prepared by: Authors
The main destinations of Ecuadorian exports in 2021 were: United States with
37%, Russia with 14%, Holland with 10%, Ukraine and Italy with 3%, Canada
with 4% and the rest of the world with 29%, (CFN, 2021) Therefore, the trade
balance for this product is favorable, including the year 2020, which was $827
thousand dollars FOB due to the limited import of flowers from the world. With
respect to imports (M) to the year 2019 there was a highly significant value in its
historical evolution, reaching $255,036.85 FOB imports, being the countries of
Colombia with 99% and China with 1% respectively the countries of origin to the
years 2020 and 2021. The financing received by this sector comes in 93% from
Private Banks and 7% from the Public; the latter, through the CFN mostly in the
province of Pichincha with 56%, followed by Cotopaxi, Imbabura, Carchi and
Cañar with 16%, 14%, 8% and 6% respectively between 2018 and 2021.
(C(CFN, 2021)

METHODOLOGY
The research was descriptive and used a mixed approach, where the
quantitative analysis was based on the Box - Jenkis (1970) methodology better
known as ARIMA (Perez, 2017) by studying a seasonal factor that allows
understanding this effect in the longitudinal monthly historical series,
corresponding to 259 observations from January 2001 to July 2022, obtained
from the (Central Bank of Ecuador, n.d.) in statistical-documentary form through
the econometric program Eviews 9. The application of this methodology begins
with the specification, verification, prediction and use of the model, (Altamirano,
2021) It began with the revision of the original model through the application of
stationarity tests to ensure that the series is stable and allows capturing the
seasonal component for prediction; when this condition was not met, the first
difference was applied to the series of natural flower exports to subsequently
obtain a robust SARIMA model of monthly frequency (f=12) that was the basis
of the GARCH model (1,1) (Farrukh & Pär Österholm, 2022) model, which was
subjected to validation tests for use in forecasting five data points between
August and December 2022.

RESULTS
By generalizing the ARMA (p,q) model incorporating a difference to stationarize
the process, it becomes ARIMA (p,d,q) to eliminate the unit root, considering
that the three components of this model are: autoregressive, integration by
difference and moving averages. (Fernandez, 2008; Gujarati & Porter, 2010;
Leon, 2017). The ARIMA equation is as follows:
Yʼt =1 Yʼt-1 + Yʼt-2 + ... +p Yʼt-p -  t-1 -  t-2 - ... -q t-q +U t (1)

By having seasonality conditions to the ARIMA model, then an important

component related to the SAR and SMA models happens to them, precisely
related to seasonal autoregressive and moving average processes that together
make up the SARIMA model (P,D,Q) where D is the integration process or
seasonal difference. (Rosales et al., 2010) This model allows the variable Yt to
correlate with its seasonal lags. (Franses et al., 2014).
The seasonal autoregressive variant of the SAR (p,d,q) (P,D,Q)f equation:
12Ytʼt =1 Yʼt-1 + ... +p Yʼt-p +2 Yt-12 + ... +p Yt-f +Ut (2)

The seasonal moving average variant of the SMA (p,d,q) (P,D,Q)f equation:
12Yt = -  t-1 - ... -q t-q - - t-12 ...- q t-f + Ut (3)

The SAR and SAM combination in SARMA would be:

12Yt =1 Yt-1 +... +p Yt-p +2 Yt-12 +... +p Yt-f -  t-1 - ... -q t-q - - t-12 ...- q t-f + U t(4)
The SARIMA combination (p,d,q) (P,D,Q)f:
12Yt =1 Yʼt-1 +... +p Yʼt-p +2 Yt-12 +... +p Yʼt-f -  t-1 - ... -q t-q - - t-12 ...- q t-f + Ut (5)

The Garch (p,q) model is mathematically defined as follows (Economipedia,

2022) as follows:
p q
σ =ω+ α ∑ ε + β ∑ σ t−q
2 2 2
t t−p
p=1 q =1

(6)

The GARCH model (generalized autoregressive conditional heteroscedastic

conditional model) uses the variance and its lags (σ 2t−q) where β is the volatility
factor and α the shocks that affect the model and it is required that:
∑ (α + β)<1 para ser estacionario and in the model C>0
When analyzing the original series in Illustration 3 (a), it is possible to observe
values in the form of peaks that determine seasonal points that are repeated
every period throughout the historical series under study.
Illustration 3.
Prepared by: Authors

By obtaining a graph in Eviews that breaks down the seasonality of the export
series, it is possible to evidence monthly periods that experience high FOB
values, such as February, March and May in the entire statistical series; and, in
the case of Ecuador it naturally corresponds to events such as the day of love
and friendship, as well as Mother's Day which is when sales soar at national
and international level, and that according to Reinoso (2016) cited by (Cedillo et
al., 2021). there are periods of high demand at high prices as well as the
opposite to the point of limiting exports, as shown in Illustration 3 (b).
The ARIMA methodology begins with the specification, i.e. the evaluation of the
natural flower export series and the contrasts that were applied: Dickey Fuller
Augmented (DFA), Phillips Perron (PP), Kiatkowski, Phillips, Schmidt and Shin
(KPSS) and Elliott Rotheberg and Stock (ERS). (Altamirano, 2021). The results
are presented in Table 3 below:
Table 3.
Contrasts to determine the stationarity of the flower export series.
Source: Format obtained from (Altamirano, 2021)
Prepared by: Authors

As can be seen, the series is not stationary; therefore, it requires a

transformation process through a first difference, in such a way that allows the
mean to be zero E(wt )=0 and the variance more or less constant Var(wt )=σ2 ,
a process known as white noise. (Hamilton, 1994; Morales, 2022). We therefore
proceed to evaluate the criteria again in Table 4:
Table 4.
Contrasts to determine the stationarity of the Differentiated Export Series
Source: Format obtained from (Altamirano, 2021)
Prepared by: Authors

With these data, we proceed to determine the appropriate model to be selected

and proceed to its verification according to the Box-Jenkis methodology, as
shown in Table 5 below.
Table 5.
Significance of the coefficients of the proposed models
Note: Coefficients in color are significant
Prepared by: Authors

As can be seen, 11 models were obtained, of which we proceeded to verify the

one whose coefficients are significant, as in the case of models 1, 5, 6, 7 and 10
whose probability (p < 0.05) and through the application of the validation criteria
Akaike (AIC), Schwarz (BIC) and Hannan Quinn (HQ), the one with the smallest
coefficient was chosen as technically adequate, as can be seen in the records
of Table 6. (Sánchez, 2021) was chosen as the most technically adequate, as
shown in Table 6:
Table 6.
Validation Criteria for the models
Prepared by: Authors

According to the evaluation shown in Table 6, the model that technically adjusts
to be the adequate one is model 10, although under the parsimony criterion
they could be the first ones, but they would not capture the essence of the
seasonality that is intended to be demonstrated in this research.
Model 10: d_ord_export c ar(2) ma(1) sar(12) sma(12)

The model was subjected to unit root analysis, in order to rule out the possibility
that the model is unpredictable and can be reliable, as shown in Figure 4 (a)
below:

Illustration 4.
Prepared by: Authors

In this sense, the model does not contain unit roots and can pass to the next
level, the analysis of the residuals by means of the Durbin Watson Criterion,
allows analyzing their independence as shown in Figure 4 (b). Altamirano
(2021) The Durbin Watson criterion indicates the criterion to rule out the
existence of independence of the residuals. (Todoeconometría, n.d.)
n

∑ ( e t −e t−1 )2
t−2
DW = n (1)
∑e 2
t
t =1

The DW coefficient indicates 1.925606 and allows us to accept the null

hypothesis of independence of the residuals. (Altamirano et al., 2022). With
respect to the analysis of the squared residuals, it is observed that the model
suffers from heteroscedasticity, since the probability values are less than 0.05,
the null hypothesis on homoscedasticity is rejected, see Illustration 5 (a):
Illustration 5.
Prepared by: Authors
 As can be noted the probability value is less than 0.05 (p<0.05) so the model is
volatile and is complemented through the heteroscedasticity test in Figure 5 (b)
whose p=0.0000 value indicates the presence of problems in variance.
To correct the variance, it has been proposed that the model be structured
under the ARCH family of heterocedastic models developed by Engle (1982)
and its variant or extension GARCH (generalized autoregressive conditional
heterocedastic conditional model) proposed by Bollerslev (1986) which
abandons the traditional assumption of normality for the presence of time-
varying variance (Farrukh & Pär Österholm, 2022) to correct these problems
that affect the predictive process and which are based on maximum likelihood
criteria. The results of the GARCH model are shown in Figure 6:
Illustration 6.
Prepared by: Authors

By analyzing the unit root processes, residuals and squared residuals to detect
autocorrelation and heteroscedasticity, the problem has been corrected and the
next stage of the Box-Jenkins methodology can be passed, considering a
determination coefficient of 76%, which is quite acceptable.
Illustration 7.
Prepared by: Authors

It should be noted that the probability p-values in Figure 7 (a), exceed the
condition of 0.05, thus there is independence of the residuals. In Figure 7 (b),
similarly the p-values exceed 0.05 which allows certifying the stability of the
model and even more if it is complemented with the Heterocedasticity Test
which gave the following values shown in Figure 8 which exceed the 0.05
probability values below:
Illustration 8.
Prepared by: Authors

Now, it is time for the prediction stage according to the proposed ARIMA
methodology; therefore, we proceed to determine the predictability criteria in
Illustration 9 (a) of this model 10, the results of which are presented below:

Illustration 9.
Prepared by: Authors

As can be seen in this phase, the criteria that determine that the model is
reliable are analyzed; its mean absolute percentage error MAPE is 10%, which
means that it can be used in a moderate range for forecasting, considering the
high volatility that can occur in the international market due to different
circumstances that affect the exports of different Ecuadorian products, as has
happened with the pandemic and the current Russian invasion of Ukraine.
Illustration 9 (b) shows the results obtained from the technical proposal of using
the GARCH model using the SARIMA model as a fundamental basis with the
original data, concluding that it adequately captures and models the real
historical values.
The forecasts for the next five months, from August to December 2022 with this
model are shown in Figure 10 (a) in graphical form:
Illustration 10.
Prepared by: Authors

The projection equation that allows the modeling of this macroeconomic series
of exports is detailed below:
Corresponds to a: SARIMA (2,1,1)(1,0,1)12 where the coefficient ar(1) remains
muted in the model.
The projected values for the five out-of-sample data are broken down
graphically and numerically in Illustration 10 (b) and allow us to have good
growth expectations that will help sustain the national economy by having
resources that will allow us to make decisions related to the use of foreign
currency in the General State Budget (PGE).

DISCUSSION
The application of the SARIMA model is multiple in different fields such as
economics through the prediction of exports of natural flowers, in business
administration through the prediction of the sale of toys for children and even in
fields such as health through the forecast of patients with diseases that depend
on the climatic seasons, among other fields such as tourism, environment and
agriculture and hydraulic infrastructure to name a few where the ARIMA
methodology is adequately applied with its seasonal factor.
In the case of Ecuadorian flower exports, a SARIMA model (2,1,1)(1,0,1) 12 was
obtained, which, supported by a MAPE of 10%, allows predicting the results that
would be obtained in the last months of the year 2022. In this sense, an
important macroeconomic variable that is used due to seasonal characteristics
is inflation, and in this regard (Santana, 2006) considers the SARIMA(1,1,1)
(0,1,1) model12 together with the analysis of neural networks and concludes that
any of them ends up being reliable although more accurate for predicting is the
second one. Likewise, a factor that affects inflation is the prices in the domestic
market, for example the sale of motorcycles that depend on seasonal market
conditions, (Moro et al., 2013) In their study, they determined a SARIMA (1,0,1)
(2,1,0) model12 with a MAPE of 9.43% that allows to adequately model the
demand for motorcycles in the Brazilian market being important for the planning
of the supply according to local needs.
 An important aspect that they point out Tudela et al., (2022) in their study on
the application of the Box-Jenkis methodology in the demand of tourists to Peru
is the acceptance of a seasonal ARIMA model, with characteristics (1,1,1)
(0,1,1)12 that captures the cyclical essence of tourism, however in their work
they point out that the affectation of the COVID-19 affects the forecasts and
establishes as a fundamental factor the public policies aimed at reestablishing
collective health. Thushara et al., (2019). They also agree in the case of tourism
in Sri Lanka has grown generating foreign exchange and employment and
therefore the importance of the development of a model that allowed making
decisions on investments in infrastructure, hotels and others for the tourist, for
this it was decided to use a SARIMA model (3,1,3) (2,1,2) 12 with MAPE
characteristics of less than 10% becoming adequate for the respective
modeling.
For De Vasconcelos et al., (2022) the application of three SARIMA models was
appropriate to establish the number of admissions that the clinic would have
resulting from health conditions deriving in pneumonia (3,1,3) (1,0,0,0),
bronchitis (4,1,4) (0,0,2) and asthma (1,1,2) (0,0,2) which is closely related to
the seasonal climatic changes of autumn and winter 23 years ago. In this
regard, they agree Raj et al., (2019) and associates that public policies should
also focus on mitigating the impact of climate change on crop yields, using
integrated seasonal SARIMAX (0,1,0)(2,0,0,0) models 4 , ANN neural networks,
and VAR autoregressive modeling, and propose the use of multivariate models
for better prediction of the results.
According to Zou et al., (2018) about the short-term seasonal relationship is
suitable for Sarima SARIMA (1,1,1)(1,1,1)(1,1,1) models 12 , up to four months
ahead, their study focuses on the trend analysis of the deformation of the
hydroelectric dam in Vietnam called Hoa Binh, for which it is compared with two
other prediction models, the backpropagation neural network (BPNN) and the
multiple regression model, concluding that the three models allow projecting the
deformation trend for the respective safety actions that are required to be taken
due to the fact that it is an important axis of economic development. With this
they agree (G. B. Lima et al., 2015). in their study on the flow of the Doce River
and its seasonal behavior, obtaining a SARIMA model (1,1,1)(1,1,2) 12 with
important predictive characteristics, especially in critical months when the
tributary tends to overflow and the management of the Municipality of Colatina
is required to prevent possible natural disasters.
Calis et al., (2017) point out that the decomposition and Box - Jenkis methods
are suitable for forecasting the number of customers attending the Izmir
financial center in Turkey being more consistent the SARIMA (1,1,1) (0,1,1)
(0,1,1) model5 with a MAPE close to 11% considered then of good accuracy, to
consider the projection of air conditioning consumption in these financial centers
due to the agglomeration of people, which is associated with higher costs that
require to be foreseen in advance.
Submedio São Francisco River Valley is an important place in Brazil from where
mango is exported that generates important sources of income, the model
proposed by the authors is related to the study of the price of this fruit through a
SARIMA (4,1,1)x(0,1,3)2 whose seasonality is semiannual and focuses on the
determination of public policies related to subsidies towards producers. (De
Lima et al., 2013) . In this field, (Da Silva et al., 2018) in their study on the
forecasts of recovery of areas and vegetation of Caatinga in the Serra das
Capivaras National Park in the state of Piauí - Brazil with the purpose of
conservation of the biomes, they determined a SARIMA model (1,2,2) (0,1,2) 12
in comparison with the Difference Vegetation Index (NDVI), observing seasonal
characteristics that are collected to make decisions aimed at the protection of
natural areas.

CONCLUSIONS

Ecuadorian natural flowers, as a non-traditional product, present seasonal

characteristics that can be modeled using the Box-Jenkis methodology,
obtaining a SARIMA (2,1,1)(1,0,1) model12 selected from a set of models that,
after evaluation, was adequate to build a GARCH model using maximum
likelihood techniques that allowed capturing the cyclical characteristics of the
monthly time series.
The forecasts obtained for the months of August to December 2022 are quite
reliable when analyzing the characteristics of the predictability criteria, which
ensures that the model is suitable for short-term forecasting.
Finally, the field of applicability of SARIMA models is wide as long as the series
are seasonal and the methodology used is wide to obtain robust models,
however, this research opens the door to expand new prediction techniques
such as neural networks, multiple regression models, VAR models, among
others that allow enriching the practical theoretical scaffolding of applied
econometrics.