Applied Soft Computing Journal 96 (2020) 106610

Contents lists available at ScienceDirect

Applied Soft Computing Journal

journal homepage: www.elsevier.com/locate/asoc

Forecasting of COVID19 per regions using ARIMA models and

polynomial functions
∗
Andres Hernandez-Matamoros b , Hamido Fujita a,b,c , , Toshitaka Hayashi b ,
Hector Perez-Meana d
a
Faculty of Information Technology, Ho Chi Minh City University of Technology (HUTECH), Ho Chi Minh City, Viet Nam
b
Iwate Prefectural University (IPU), Faculty of Software and Information Science, Iwate, 020-0693, Japan
c
Andalusian Research Institute in Data Science and Computational Intelligence (DaSCI), University of Granada, Granada, Spain
d
Instituto Politecnico Nacional, Av. Santa Ana 1000 Mexico D. F., 04430, Mexico

article info a b s t r a c t

Article history: COVID-2019 is a global threat, for this reason around the world, researches have been focused on topics
Received 10 June 2020 such as to detect it, prevent it, cure it, and predict it. Different analyses propose models to predict
Received in revised form 30 July 2020 the evolution of this epidemic. These analyses propose models for specific geographical areas, specific
Accepted 2 August 2020
countries, or create a global model. The models give us the possibility to predict the virus behavior, it
Available online 6 August 2020
could be used to make future response plans. This work presents an analysis of COVID-19 spread that
Keywords: shows a different angle for the whole world, through 6 geographic regions (continents). We propose
Covid-19 epidemic to create a relationship between the countries, which are in the same geographical area to predict the
Forecast advance of the virus. The countries in the same geographic region have variables with similar values
ARIMA model (quantifiable and non-quantifiable), which affect the spread of the virus. We propose an algorithm to
Geographic region performed and evaluated the ARIMA model for 145 countries, which are distributed into 6 regions.
Then, we construct a model for these regions using the ARIMA parameters, the population per 1M
people, the number of cases, and polynomial functions. The proposal is able to predict the COVID-19
cases with a RMSE average of 144.81. The main outcome of this paper is showing a relation between
COVID-19 behavior and population in a region, these results show us the opportunity to create more
models to predict the COVID-19 behavior using variables as humidity, climate, culture, among others.
© 2020 Elsevier B.V. All rights reserved.

1. Introduction forecasting, [15] studied the spread of COVID-19, they realized

forward prediction and backward inference of the epidemic. [16]
In December 2019 in Wuhan, China started the pandemic of applied mathematical models and time-series to describe the out-
COVID-19, commonly known as Coronavirus, which has caused break among passengers and crew members on Princess Cruises
havoc around the world. World Health organization reported on Ship.
June 7 [1], the virus is in 216 Countries, there are 6 750 521 A model using early forecasting from Small Dataset is pro-
active cases, and it has produced 395 779 deaths. For this reason, posed by [6]. In [5] the authors proposed used ARIMA models
scientists around the world have been focused on topics such to predict the spread around the world, in specific they use
detect it [2], prevent it [3], cure it [4], and predict it [5–13]. To two models ARIMA (1,2,0) and ARIMA (1,0,4). An ARIMA (1,1,2)
predict the coronavirus different schemes has been applied, for model is selected to fit the predictions in Italy by [8,10] proposed
example in [11] proposes an approach, which is based Composite ARIMA models to predict cases and deaths per 3 countries Italy,
Monte Carlo enhanced by deep learning and fuzzy rule induction Turkey, and Spain. A ARIMA (2,2,2) model is used by [9] to predict
to predict the COVID-19, [14] detailed models for forecasting the the spread in India. Models for different regions of Italy are
course of the pandemic, these models demonstrate the utility of proposed by [7]. [13] uses (ARIMA) model to analyze two data
parsimonious models for early-time data. Using the official data sets and predict the daily new confirmed cases for the 7-day
period. In [12], the authors have studied 15 countries (USA, Spain,
Italy, France, Germany, United Kingdom (UK), Turkey, Iran, China,
∗ Corresponding author at: Iwate Prefectural University (IPU), Faculty of
Russia, Brazil, Canada, Belgium, Netherlands and Switzerland) to
Software and Information Science, Iwate, 020-0693, Japan.
E-mail addresses: [email protected] (A. Hernandez-Matamoros),
predict the spread of the Coronavirus in these countries.
[email protected], [email protected] (H. Fujita), As we can see, advanced prediction models used ARIMA [17]
[email protected] (T. Hayashi), [email protected] (H. Perez-Meana). to predict the spread of Coronavirus, this is because the ARIMA

https://doi.org/10.1016/j.asoc.2020.106610
1568-4946/© 2020 Elsevier B.V. All rights reserved.
2 A. Hernandez-Matamoros, H. Fujita, T. Hayashi et al. / Applied Soft Computing Journal 96 (2020) 106610

Fig. 1. Example of the available information on ‘‘Our World in Data".

Fig. 2. Block diagram of Building the model.

models give results in terms of its predictive performance. The are the geographic region and the total population in the coun-
models give us the possibility to predict the virus behavior, it try. The geographic regions are North America, South America,
could be used to make future response plans. There are coun- Africa, Asia and Europe. The countries in the same geographic
tries, which have faced the COVID-19 in the same way, (Brazil region (continent) different variables with similar values such as
and Sweden) but with totally different consequences. The dif- quantifiable data (climate, humidity, natural regions, etc.) and
other non-quantifiable (cultural similarities, similar gastronomy,
ferences between these countries are geographical, demographic,
among others).
economic, public health, cultural, poverty, among others. These
We propose an algorithm to performed and evaluated the
differences have caused that Brazil has 694,116 cases while Swe- Auto-Regressive Integrated Moving Average (ARIMA) model for
den has 45,133 cases on June 8. As we can see, Brazil has 15.3 145 countries, which are distributed in 6 geographic regions.
times more cases than Sweden. For this fact, in this work, we The ARIMA models using the available information until April 25
propose to create a relationship between the countries and two 2020. Next, the information is divided into 2 sets, the first set
variables more to predict the COVID-19 behavior. These variables is used to create the ARIMA models and the second set is used
 A. Hernandez-Matamoros, H. Fujita, T. Hayashi et al. / Applied Soft Computing Journal 96 (2020) 106610 3

Fig. 3. COVID-19 time series of Canada.

to calculate the RMSE between the real data and predict data. of time series are different per each country. For example, the
The First set uses 90% of the data and the second set uses 10% of time series of Canada starts on January 26 (123 days until May
the data. Then, the calculated parameters of ARIMA models, the 28), Egypt time series starts on February 15 (103 days until May
population per 1M people per country, the number of cases per 28), time series of China starts on December 31 (149 days until
country are used to create polynomial functions, which are able May 28), Italy time series starts on January 31 (118 days until
to predict the ARIMA parameters. These polynomial functions May 28), time series of Australia starts on January 25 (124 days
generate models for the next geographic regions: North America, until May 28), Brazil time series starts on February 26 (92 days
South America, Africa, Oceania, Asia, and Europe. until May 28), etc.
The results are evaluated using RMSE. The main contributions This paper proposes a model per geographic region. The coun-
can be summarized as follows: tries are separated in 6 regions which are North America (13
countries), South America (12 countries), Africa (43 countries),
• We propose an algorithm to calculate the best ARIMA pa- Asia (40 countries) and Europe (33 countries). To create the mod-
rameters per country with low RMSE. els per region, we use the ‘‘total population in the age groups’’
• The algorithm to calculate the best parameter of ARIMA is available in the website of United Nations [19]. The population in
tested with 10% of the original data. the age groups is added to generate a total population. The values
• Our approach is analyzing 145 countries, almost 10 times for each country are shown in the Tables in Appendix A.
more than another proposed scheme.
• The approach starts analyzing particular cases (countries) to 3. Proposed approach
create a general case (geographic region).
• Our approach is able to show a relation between the pre- The proposed approach consists of two stages ‘‘Building the
diction error and other variables. In this work, a relation model’’ and ‘‘Evaluating the model’’. These stages are applied
between the prediction error and the population per 1M 6 times, one time per region. We use the time series ‘‘Total
people is shown. confirmed COVID-19 cases’’. The first stage ‘‘Building the model’’
requires the time series per country, which starts on the day
The organization of the paper is as follows. Section 2 briefs the when each country presented the first case of COVID-19 and it
databases used. Section 3 presents the proposed approach. The finishes on April 25. The second stage ‘‘Evaluating the model’’
paper ends with the Results, a Discussion section and Conclusion. requires the information of COVID-19 on May 28. Then the fore-
casting between May 12 and May 28 is calculated and compared
2. Databases with the real values. In the following subsections, the proposed
approach is explained in a general way and using an example. The
The time series created in this work using the data of ‘‘Our example calculates the ‘‘p, D, q’’ values of ARIMA to Canada and
World in Data’’ [18], which is completely open access. They collect builds the North America model.
the data from the European Centre for Disease Prevention and
Control (ECDC), the WHO, Johns Hopkins, United Nations, World 3.1. Building the model
Bank, Global Burden of Disease, Blavatnik School of Government,
etc. They standardized names of countries using ‘‘Our World Fig. 2 shows a block diagram of this stage, which consists of
in Data’’ [18] standard entity names, they discarded detected the ARIMA and polynomial functions. The inputs of this stage
inconsistencies in the original data, detailed documentation for are time series of the countries per region and Rc, which are
each country is available [18]. Multiple time series for a country explained in Section 3.1.1 (Arima Stage) and 3.1.2 (Polynomial
are collected, the complete COVID-19 dataset only includes the Functions), respectively.
most complete number of people tested, confirmed cases and
deaths. The data on the coronavirus pandemic is updated daily. 3.1.1. ARIMA stage
‘‘Our World in Data’’ has 77 charts on COVID-19. Fig. 1 shows We use the time series ‘‘Total confirmed COVID-19 cases’’ per
an example of one chart. The data of charts contain information country. Then, we have a time series presented in the following
from 207 countries. Then, we can explore the statistics on COVID- equations:
19 for the countries in the world. This work uses the available
information until May 28. The consulted chart is ‘‘Total and daily y = {yt , t ϵ T } (1)
confirmed COVID-19 cases’’, which is used to create the time
T = {T1 , T2 , T3 , . . . , T1+n } (2)
series per country called ‘‘Total confirmed COVID-19 cases’’. The
time series start on the day when each country presented the first In Eq. (1), y means the total confirmed cases per day presented
case of COVID-19 and finish on May 28. This fact means the length in a country. In Eq. (2), T1 means the day when each country
4 A. Hernandez-Matamoros, H. Fujita, T. Hayashi et al. / Applied Soft Computing Journal 96 (2020) 106610

Fig. 4. Canada time series separated into training and testing time series.

detected the first COVID-19 patient. Then D1+n represents the

number of days elapsed until April 25. For example, Canada de-
tected the first patient on January 26. Then, we can rewrite Eq. (2)
as shown Eq. (3).

T = {Jan26, Jan27, Jan28, . . . , Apr25} (3)

The Fig. 3 shows the time series of Canada. To compute the

best parameters of ARIMA, the time series is separated into train-
ing time series and testing time series. To train time series is
created using 90% of the data from the original time series. At
the same time, test time series is created using 10% of the data
from the original time series. The Fig. 4 shows the example using
time series of Canada.
Then, we have two time series (Train and Test). The next
equations presented in a formal way these time series.

k = round(n ∗ .9) (4)

Train = {yt , t ϵ U } (5)
U = {T1 , T2 , T3 , . . . , Tk } (6)
Test = {yt , t ϵ V } (7)
V = {T1+k , T2+k , T3+k , . . . , T1+n } (8)

In Eq. (4), the k value is calculated, this value is the threshold

to separate the data between Train and Test. Eqs. (5) and (7)
present the Train and the Test time series, respectively. Then,
Algorithm 1 is applied to calculate the best parameters of an
autoregressive integrated moving average (ARIMA).
ARIMA is a statistical analysis, it uses time series data. The
ARIMA predicts future values by examining the differences be-
tween values in the time series. An ARIMA model consists of
3 components Auto regression (AR), Integrated (I), and Moving
average (MA). Each component is a parameter. To represent these In the previous Algorithm, Root Mean Square Error (RMSE)
[22] measures the stability between the original data and forecast
parameters, ARIMA models use a standard notation p, D, and
data, RMSE is calculate using Eq. (9).
q. This standard notation indicates the type of ARIMA model
used. Where p means the number of lag observations, D means

 n
 ∑
the degree of difference, and q means the order of the moving RMSE = √n (Xi − Si )2 (9)
average, for further details refer to [14,20,21]. i=1
 A. Hernandez-Matamoros, H. Fujita, T. Hayashi et al. / Applied Soft Computing Journal 96 (2020) 106610 5

Fig. 5. Comparison between Real data and Forecast Data of Canada.

Table 1
Data of North America.
Country Population per million Total confirmed COVID-19 cases per 1M people ARIMA Parameters RMSE average
people (ppMp) April 25 May-11 p D q 640.61

Belize 0.390 45.269 45.269 0 2 0 2.830E−07

Canada 37.411 1162.546 1824.167 3 2 3 2.611E+02
Cuba 11.333 113.450 155.916 5 1 3 2.128E+01
Dominican Republic 10.739 529.964 953.825 5 0 3 6.851E+01
Grenada 0.112 133.311 186.635 5 1 3 2.452E−01
Guatemala 17.581 24.001 58.720 2 2 5 1.671E+01
Haiti 11.263 6.314 15.961 3 3 1 2.392E+00
Honduras 9.746 59.669 199.099 4 3 1 1.719E+01
Jamaica 2.948 97.259 169.528 5 3 3 7.446E+00
Mexico 127.576 99.835 271.630 2 3 0 1.614E+02
Nicaragua 6.546 1.660 2.415 5 5 3 3.859E−01
Panama 4.246 1237.146 1957.927 3 1 5 2.141E+01
United States 329.065 2690.383 4017.488 4 1 3 7.750E+03

In our example, the algorithm 1 is applied into Train Canada

Time series. Then, we have p, D, q values, p is 3, D is 2 and q is ⎡
DC 1
⎤
4. Fig. 5 shows a comparison between the Real data (Test Time ⎢DC 2 ⎥
series) and Forecast Data of Canada. ⎣ .. ⎦
VD = ⎢ (11)
⎥
This process is applied to each country. Appendix A presents .
the p, D, q values of the ARIMA model and the RMSE to each DCk
analyzed country.
qC 1
⎡ ⎤
⎢qC 2 ⎥
3.1.2. Polynomial functions ⎣ .. ⎦
Vq = ⎢ (12)
⎥
The ARIMA stage calculated the p, D, q values to each coun- .
try. In this stage, we use these values, the information of cases qCk
confirmed COVID-19 per million people on April 25, and the
RcC 1
⎡ ⎤
Population per million people (ppMP). Table 1 shows an example
of the obtained values in the previous stage, which are used in ⎢RcC 2 ⎥
⎣ .. ⎦
Rc = ⎢ (13)
⎥
this stage.
.
This stage calculates the polynomial function for ARIMA pa- RcCk
rameters (p, D, q). This fact means, we must calculate 3 polyno-
mial functions, one for each ARIMA Parameter. Then, we create
the Matrixes Vp , V D , Vq , which have the p, D, q values for the Ck = 1, 2, 3, . . . , Number of countries in the region (14)
countries in the region. Rc is a matrix, each one of the elements
In our example k=1, 2, 3, . . . , 13. Then we need to apply the
of Rc belongs to one country. These elements are calculated per Algorithm 2 on Vp , V D , Vq .
country as follows: the numbers of confirmed COVID-19 cases on The Algorithm 2 creates the vectors t and d. The vector t starts
April 25 multiply per the population (ppMp). in 1 and finishes in the maximum value of Rc. The vector d has the
⎡
pC 1
⎤ information of the ARIMA parameters. Then, we need to calculate
⎢pC 2 ⎥ a polynomial p(t) of degree n, that is the best fit for the data in
⎣ .. ⎦
Vp = ⎢ (10) the vector d, as shown Eq. (15).
⎥
.
pCk p (t ) = p1 t n + p2 t n−1 + p3 t n−2 + · · · + pn t + pn+1 (15)
6 A. Hernandez-Matamoros, H. Fujita, T. Hayashi et al. / Applied Soft Computing Journal 96 (2020) 106610

t1n t1n ... 1 p1 d1

⎛ ⎞⎛ ⎞ ⎛ ⎞
The next Eqs. (18)–(20) show the polynomial functions for
⎜ t1n t1n ... 1 ⎟ ⎜ p2 ⎟ ⎜ d2 ⎟ North America, the polynomials of all regions are available in
⎜ . .. .. ⎟
⎝ . .. ⎠ ⎝ .. ⎠ = ⎝ .. ⎠ Appendix B.
⎜ ⎟ ⎜ ⎟
. . . . . . (16)
North America
t1n t1n ... 1 pn+1 dm
n is degree of polynomial pNAp (t ) = p1 t n + p2 t n−1 + p3 t n−2 + · · · + pn t + pn+1 ; n = 14
To solve Eq. (15), we present the problem as shown the (18)
Eq. (16). Then, t is used to form Vandermonde matrix [15] V pNAd (t ) = p1 t + p2 t
n n−1
+ p3 t n−2
+ · · · + pn t + pn+1 ; n = 15
with n+1 columns and m rows. Where m is the length of d. After (19)
to solve the Eq. (16), we find the values of P1 , P 2 ,. . . , P n+1 . To
calculate the best values of Pn , we propose the Algorithm 3. pNAq (t ) = p1 t n + p2 t n−1 + p3 t n−2 + · · · + pn t + pn+1 ; n = 47
(20)

3.2. Evaluating the model

The Fig. 7 shows the block diagram of the Evaluating model

stage. This stage has 3 inputs Ec, time series of the country until
May 11, and days to predict. Ec has the value of Total confirmed
COVID-19 cases multiply per the population for the country to
be evaluated in the region. In this evaluation, we use the data
on May 11 and the polynomial functions creates in the previous
stage. The Algorithm 4 is used to calculate the ARIMA parameters.
We apply the Algorithm 4 using the values of Canada on May
11 and the functions PNap (t), PNad (t), and PNaq (t) Eqs. (18)-(20).
Canada belongs to North America, so we use the functions of
North America to calculate the ARIMA parameters. To calculate
At the beginning of the Algorithms 2–3, we can take one of another country, we must use the functions which belong to the
the ARIMA parameters, this fact means that the Algorithms 2 and region of the country.
3 must be applicate 3 times, one time per ARIMA parameter.
After the Algorithms 2 and 3 are applied the values to North
America model are: 14 is degree of polynomial to parameter p,
15 is degree of polynomial to parameter D, and 47 is degree of
polynomial to parameter q. The Eq. (18) shows an example of
the polynomial created for ARIMA parameter D. Fig. 6 shows the
polynomial functions for ARIMA parameters.
pNAd (t ) = p1 t 15 + p2 t 14 + p3 t 13 + p4 t 12 + p5 t 11
+ p6 t 10 + p7 t 9 + p8 t 8 + p9 t 7 + p10 t 6
+ p11 t 5 + p12 t 4 + p13 t 3 + p14 t 2 + p15 t
After we applied the Algorithm 4, the Canada ARIMA param-
+ p16 (17) eters are p=3, D=2, q=3. The next step is the prediction using
 A. Hernandez-Matamoros, H. Fujita, T. Hayashi et al. / Applied Soft Computing Journal 96 (2020) 106610 7

Fig. 6. Polynomial Functions for North America Model.

the Algorithm 5. The Canada results using the model of North 4. Results
America are shown in Figs. 8–9.
The Eq. (9) is applied to calculate the RMSE between the fore- This section presents the results for each region analyzed.
cast values and real values. Fig. 8 shows a comparison between Table 2 shows the average RMSE per region. In the table, the
the real and forecast signals and Fig. 9 shows the forecast of RMSE is calculated between the forecast and the real values.
Canada with confidence interval of 95%. Fig. 10 presents a comparison using RMSE and the forecast for one
8 A. Hernandez-Matamoros, H. Fujita, T. Hayashi et al. / Applied Soft Computing Journal 96 (2020) 106610

Fig. 7. Block Diagram of Evaluating the model.

Fig. 8. Comparison between Real Data and Forecast Data.

country per each region in the following way: (a) North America, 5. Discussion
(b) South America, (c) Africa, (d) Oceania, (e) Asia, and (f) Europe.
The Appendix A presents the results per country before to
create the geographic models. These results belong to each coun-
try in the different regions. As we mention in Section 3.1.1, the
time series are separated into modeling (90% of the signal) and
testing (10% of the signal). Below, we will discuss each region in
particular.
North America region has 13 countries; this region presents
a RMSE average of 640.61. The RMSE average of this region is
the most bigger between the regions. This fact appears, because
the United States presents the most bigger RMSE between the
145 countries (7749.99), this country has the largest number of
population in the region (329.06 ppMp). On the other hand, Belize
presents the lowest RMSE. The United States has almost 96 times
the population of Belize (0.39 ppMp).
Europe region consists of 33 countries. In this experiment, the
countries, which present a ppMp major than 45 ppMp presents
biggest RMSE. Spain presents an RMSE of 1892.33 with a 46.73
ppMp, Italy has 60.55 ppMp and presents an RMSE of 566.88,
 A. Hernandez-Matamoros, H. Fujita, T. Hayashi et al. / Applied Soft Computing Journal 96 (2020) 106610 9

Table 2
Average of RMSE results.
Region Average of RMSE between Average of RMSE between
original and forecast signal in original and forecast signal
training stage from May 12 to May 27
North America 640.61 3.6051e+04
South America 104.78 2.0828e+04
Africa 13.80 1.4913e+03
Oceania 6.79 161.2570
Asia 89.46 3.52964e+03
Europe 218.59 2.88212e+04
a
Average 144.81 1.2723e+04
a
The average is calculated using the 145 countries.

Table 3
Comparison between [10] and this work.
Country RMSE
[10] This work
Italy 1150.31 566.88
Turkey 138.35 1892.33
Spain 379.89 696.35
a b
Average 556.183 144.81
a
Using 3 countries.
b
Using 145 countries.

it has countries with 200 ppMp. This fact could be means a rela-
tion between the virus spread and the climate for example. For
the region Asia the average RMSE is minor to 90, this area was the
first area infected by COVID-19 so there are more available data
in this area. Thus, we have more data to calculate the Forecast.
Table 3 shows a comparison between [10] and this work
before to create the geographic models. As shown Table 3 this
approach has better RMSE to forecast the virus in Italy, on the
contrary [10] has better RMSE to predict the virus in Turkey and
Fig. 9. Forecasting of Canada.
Spain. At first, it seems that their proposal is better than ours,
but when the RMSE averages are compared, we can see that our
proposal has a lower RMSE than them, besides we are analyzing
United Kingdom presents an RMSE of 728.13 with a 467.53 ppMp, 145 countries while they only analyze 3.
Germany has 83.51 ppMp and presents an RMSE of 1075.02, and Fig. 8 presents the results per one country in each region.
Russia presents an RMSE of 958.44 with a 145.87 ppMp. The In Fig. 8(a–f), we can see an upward trend in the number of
RMSE average of this is 218.59. cases, with the exception of the American States, which marks
Brazil presents a RMSE of 591, this country has the largest a decrease in the number of cases. Let us remember that from
number of population in the region (211.04 ppMp). In contrast, the beginning, the American States had the highest RMSE among
Paraguay presents the lowest RMSE. Brazil has almost 30 times all countries. When the geographic models are created, these
the population of Paraguay (7.05 ppMp). These countries belong models are used to predict new cases in a country. The results are
to South America region, the RMSE average of this region is shown in Table 2. The forecast is made 17 days after the models
104.78. are calculated, we take this decision to have a real difference
Asia region consists of 40 countries. In this region, Turkey between the cases on April 25 and May 11 as shown the tables
presents the most bigger RMSE (696.35) with a population (83.43 in Appendix A. As expected, the RMSE error grew because, the
ppMp). China and India have a ppMp major to one thousand, but prediction is making 17 days after the models were created and
the RMSE are 117.88 and 250.83, respectively. On the other hand, we calculate 15 days of prediction cases. In these time interval,
Yemen with a population less than 30 ppMp has a RMSE close to the actions as quarantine control, stay at home campaign, social
zero. distance taken by governments significantly affect the prediction.
Egypt presents a RMSE of 84.08, this country has 110.38 ppMp. If the lector wants current predictions, the information needs to
In contrast, Namibia presents a RMSE close to zero. Egypt has be updated and repeat the building the model stage.
almost 41 times the population of Namibia (2.49 ppMp). These
countries belong to Africa region, the RMSE average of this region 6. Conclusion
is 13.8.
Oceania region consists of 4 countries. In this region, Australia We can conclude that the algorithm to model and evaluate the
presents the most bigger RMSE (24.76) with a population (25.20 ARIMA models is able to develop models, which have low RMSE.
ppMp). The lower RMSE is presented by Fiji with a population On the other hand, this work shows a way to model the COVID
less than 1 ppMp. spread started in particular cases to generate a general case. We
For the regions North America, South America, Oceania, and can conclude, this work contributes to researchers working in
Europe, there is a relation between the major ppMp and the error COVID-19 prediction. It shows there is a relation between the
on the prediction. The RMSE of Africa is minor to 15, even though virus spread and the different variables present in the countries,
10 A. Hernandez-Matamoros, H. Fujita, T. Hayashi et al. / Applied Soft Computing Journal 96 (2020) 106610

Fig. 10. An example of results per region. RMSE and Forecast (a) North America, (B) South America, (c) Africa, (d) Oceania, (e) Asia and (f) Europe.

which belong to the same geographic region. Interestingly, we work different variables could be analyzed, for example, the date
can find a show relation between the population in a country and when the first coronavirus case is detected in the country, humid-
RMSE error in a prediction. In future challenges of the proposed ity, temperature, among other variables. Other kinds of clusters
 A. Hernandez-Matamoros, H. Fujita, T. Hayashi et al. / Applied Soft Computing Journal 96 (2020) 106610 11

Fig. 10. (continued).

Table A.1
Data of the countries separated by geographical regions.
North America
Country Population per million Total confirmed COVID-19 cases per million people ARIMA parameters RMSE average
people (ppMp) April 25 May-11 p D q 640.61

Belize 0.390351 45.269 45.269 0 2 0 0.000000283

Canada 37.411038 1162.546 1824.167 3 2 3 261.0992307
Cuba 11.333484 113.45 155.916 5 1 3 21.28255446
Dominican Republic 10.738957 529.964 953.825 5 0 3 68.51095035
Grenada 0.112002 133.311 186.635 5 1 3 0.245202091
Guatemala 17.581476 24.001 58.72 2 2 5 16.71287027
Haiti 11.263079 6.314 15.961 3 3 1 2.392474128
Honduras 9.746115 59.669 199.099 4 3 1 17.1948204
Jamaica 2.948277 97.259 169.528 5 3 3 7.44624079
Mexico 127.575529 99.835 271.63 2 3 0 161.3544175
Nicaragua 6.545503 1.66 2.415 5 5 3 0.385861038
Panama 4.24644 1237.146 1957.927 3 1 5 21.41200047
United States 329.064917 2690.383 4017.488 4 1 3 7749.995645
South America
Country Population per million Total confirmed COVID-19 cases per million people ARIMA parameters RMSE average
people (ppMp) April 25 May-11 p D q 104.78

Suriname 0.581363 17.046 17.046 2 2 0 2.42E−09

Uruguay 3.461731 162.074 203.528 1 2 4 0.536253127
Guyana 0.782775 92.809 132.221 5 0 1 0.945527112
Paraguay 7.044639 31.265 99.965 4 3 1 3.977159335
Venezuela 28.515829 11.183 14.559 4 2 4 7.768192986
Bolivia 11.513102 69.134 218.966 5 5 4 13.19241251
Chile 18.952035 643.747 1510.027 5 3 4 31.85255857
Argentina 44.780675 75.737 127.8 2 1 2 34.307185
Colombia 50.339443 95.926 217.421 1 2 0 42.19486859
Ecuador 17.373657 633.847 1675.39 3 4 4 206.1259375
Peru 32.510462 656.56 2041.348 4 2 1 325.2189929
Brazil 211.049519 249.319 765.428 4 2 1 591.2483357
(continued on next page)

could be applied like cultural behavior, religious behavior, hy- to make current predictions, just the information needs to be
giene habits, feeding habits, among others. The approach is able updated.
12 A. Hernandez-Matamoros, H. Fujita, T. Hayashi et al. / Applied Soft Computing Journal 96 (2020) 106610

Table A.1 (continued).

Africa
Country Population per million Total confirmed COVID-19 cases per million people ARIMA parameters RMSE average
people (ppMp) April 25 May-11 p D q 13.80

Namibia 2.494524 6.297 6.297 4 2 0 3.38E−09

Mauritania 4.525698 1.505 1.721 1 2 0 1.88E−08
Seychelles 0.097741 111.857 111.857 0 3 0 0.00000302
South Sudan 11.062114 0.447 13.936 3 4 0 0.231695815
Angola 31.825299 0.761 1.369 2 0 1 0.272840126
Madagascar 26.969306 4.406 6.175 0 0 0 0.5
Botswana 2.303703 9.355 9.78 3 0 3 0.561333394
Mauritius 1.26967 260.268 261.054 5 2 0 0.577046983
Togo 8.082359 10.871 21.018 2 1 1 0.597788103
Central African Republic 4.745179 3.934 37.062 2 3 4 0.690491146
Zimbabwe 14.645473 1.951 2.422 4 1 4 0.920238223
Eritrea 3.497117 10.997 10.997 4 0 3 1.098426352
Benin 11.801151 4.784 26.313 1 1 5 1.123599466
Chad 15.946882 2.435 19.603 2 2 3 1.486804327
Libya 6.777453 8.878 9.314 5 1 0 1.802690408
Zambia 17.861034 4.569 14.524 4 1 5 2.264558145
Ethiopia 112.078727 1.018 2.079 3 0 3 2.448214102
Burundi 11.530577 1.009 1.598 1 4 4 2.535566664
Cape Verde 0.549936 158.277 442.456 4 2 2 2.574723428
Burkina Faso 20.321383 29.947 35.927 3 0 5 2.824155601
Uganda 44.269587 1.64 2.645 5 2 5 2.97488775
Sierra Leone 7.813207 10.28 38.486 3 3 1 3.214739643
Liberia 4.937374 23.133 39.346 4 1 5 3.84699105
Niger 23.310719 28.133 33.916 4 2 2 3.917381572
Sudan 42.813237 3.968 31.084 2 2 4 4.558386235
Mali 19.658023 16.049 34.764 3 2 2 5.436959696
Kenya 52.573967 6.249 12.497 5 2 2 5.646461767
Mozambique 30.366043 2.08 2.911 5 3 5 6.014669312
Malawi 18.628749 1.725 2.927 1 4 3 6.1920973
Tunisia 11.694721 78.013 87.32 1 0 0 8.775869024
Gabon 2.172578 77.278 296.981 4 3 2 11.10807476
Nigeria 200.963603 5.312 21.34 5 4 0 12.60385557
Somalia 15.442906 20.638 66.318 2 1 4 14.76084313
Senegal 16.296362 32.549 102.067 5 3 3 26.08610014
Djibouti 0.973557 1011.132 1224.694 1 1 3 26.43208425
Cameroon 25.876387 52.852 97.153 1 2 2 30.35431279
Algeria 43.053054 71.31 130.51 2 0 3 30.43462472
Equatorial Guinea 1.355982 151.106 312.904 1 2 1 41.34353663
Ghana 30.417858 41.161 137.193 5 2 2 47.53364114
Guinea 12.771246 72.643 163.408 3 5 0 59.82419028
South Africa 58.558267 71.153 168.862 3 3 5 65.82375285
Morocco 36.471766 101.814 164.262 1 1 0 70.07161038
Egypt 100.388076 39.987 91.856 5 1 3 84.08557986
Oceania
Country Population per million Total confirmed COVID-19 cases per million people ARIMA parameters RMSE average
people (ppMp) April 25 May-11 p D q 6.79

Fiji 0.889955 20.079 20.079 0 2 0 0.000000105

Papua New Guinea 8.776119 0.894 0.894 3 2 2 0.321552716
New Zealand 4.783062 231.635 237.857 3 3 0 2.092622312
Australia 25.2032 262.237 272.197 1 3 1 24.7684965
Asia
Country Population per million Total confirmed COVID-19 cases per million people ARIMA parameters RMSE average
people (ppMp) April 25 May-11 p D q 89.46

Yemen 29.161922 0.034 1.71 1 2 1 8.46E−16

Laos 7.169456 2.611 2.611 0 2 0 0.000000474
Brunei 0.433296 315.441 322.298 5 1 1 0.201374663
Mongolia 3.225166 10.981 12.812 5 3 2 0.42737412
Bhutan 0.763094 9.072 11.664 0 5 3 0.854706352
Vietnam 96.462108 2.774 2.959 0 0 0 0.894427191
Syria 17.070132 2.4 2.686 1 2 3 1.2290986
Kyrgyzstan 6.415851 101.928 155.728 4 3 1 1.764150638
Myanmar 54.045422 2.647 3.308 4 1 3 1.810985483
Jordan 10.101697 43.222 52.925 5 2 0 1.976423483
Lebanon 6.855709 101.971 123.802 3 0 2 3.248815136
Nepal 28.608715 1.682 4.119 3 2 4 4.05357834
Georgia 3.996762 111.301 159.181 0 1 4 4.895198267
Maldives 0.530957 236.799 1544.746 5 4 1 5.924589657
Cyprus 1.198574 917.914 1025.232 4 0 4 6.066194022
Azerbaijan 10.047719 157.015 248.442 4 0 5 11.83730777
(continued on next page)
 A. Hernandez-Matamoros, H. Fujita, T. Hayashi et al. / Applied Soft Computing Journal 96 (2020) 106610 13

Table A.1 (continued).

Thailand 69.625581 40.888 43.195 3 0 0 12.58960875
Iraq 39.309789 42.464 68.792 1 3 3 15.52536699
Uzbekistan 32.981715 54.856 72.246 5 1 3 18.9777337
Armenia 2.957728 565.936 1118.035 0 3 4 19.92595139
SriLanka 21.323734 19.614 40.302 3 2 4 20.83410943
Malaysia 31.949789 175.833 205.648 2 1 5 27.80701783
Kuwait 4.207077 612.097 2034.392 0 2 2 44.83279931
Oman 4.974992 350.525 665.606 2 0 2 48.21943289
Iran 82.913893 1050.017 1281.096 5 3 0 50.19983585
Afghanistan 38.041757 34.705 113.08 5 2 0 75.99268382
Bangladesh 163.046173 28.472 88.998 5 3 0 79.01328851
Qatar 2.832071 2958.98 7816.568 1 2 0 97.6076415
Bahrain 1.641164 1479.799 2903.767 1 2 2 99.73990182
Philippines 108.116622 65.632 98.502 4 0 0 103.1445092
Indonesia 270.625567 30.019 51.301 5 1 5 110.4601355
Israel 8.519373 1739.695 1903.636 0 2 4 113.630396
China 1433.783692 58.291 58.368 5 0 0 117.8792989
United Arab Emirates 9.770526 938.385 1839.966 5 5 0 188.3682723
Saudi Arabia 34.268529 433.793 1121.622 3 3 2 226.4705085
India 1366.417756 17.758 48.661 5 2 3 250.833
Pakistan 216.565317 54.053 140.073 4 3 4 266.1868764
Japan 126.860299 101.932 124.909 1 1 2 399.4484529
Singapore 5.804343 1910.657 3988.826 1 2 2 449.3506577
Turkey 83.429607 1243.931 1644.042 5 3 3 696.3506539
Europe
Country Population per million Total confirmed COVID-19 cases per million people ARIMA parameters RMSE average
people (ppMp) April 25 May-11 p D q 218.59

Montenegro 0.627988 507.912 515.873 2 2 0 0.988109524

Malta 0.440377 1012.368 1123.344 5 3 0 3.539156254
Latvia 1.90674 415.65 497.826 5 1 4 3.594220963
Iceland 0.339037 5242.491 5277.656 3 2 5 5.129808852
Croatia 4.130299 489.371 532.73 1 0 5 7.425098776
Estonia 1.325649 1209.915 1310.93 4 0 5 10.28679943
Hungary 9.68468 252.889 339.946 3 1 5 13.10475646
Norway 5.378859 1366.477 1493.938 5 2 1 14.35761421
Slovakia 5.457012 249.101 266.867 4 2 1 15.43096651
Albania 2.880913 235.597 301.619 5 2 4 15.44541938
Lithuania 2.759631 517.946 543.292 1 0 5 15.54234467
Luxembourg 0.61573 5902.782 6207.906 1 2 4 18.86761887
Bulgaria 7.000117 170.974 282.797 5 3 3 26.9764822
Slovenia 2.078654 660.435 700.841 4 2 5 29.45549614
Czech Republic 10.689213 679.15 758.522 5 1 4 32.36197356
Austria 8.955108 1673.033 1752.865 4 1 4 41.48747638
Serbia 8.772228 1099.698 1486.348 3 2 2 43.17539749
Finland 5.532159 793.218 1076.034 0 3 2 46.3725334
Denmark 5.771877 1417.423 1800.524 2 1 1 50.60316184
Sweden 10.036391 1739.433 2606.327 3 2 5 87.03004062
Romania 19.364558 541.489 798.537 4 2 5 102.4828358
Poland 37.887771 287.793 422.653 5 1 4 105.4066726
Portugal 10.226178 2282.207 2704.893 4 1 5 108.4530628
Ukraine 43.993643 185.783 348.289 2 5 2 159.7185476
Ireland 4.882498 3682.615 4657.139 1 4 2 172.2604087
Belgium 11.539326 3821.783 4580.048 5 2 0 207.3927744
Belarus 9.452409 928.426 2431.18 5 2 5 268.7103234
Netherlands 17.097123 2132.201 2487.734 5 0 0 387.229697
Italy 60.550092 3191.997 3623.278 4 3 3 566.8847446
United Kingdom 67.530161 2113.307 3228.692 2 1 0 728.1254615
Russia 145.87226 470.225 1436.864 5 5 1 958.4432358
Germany 83.517046 1819.418 2023.956 4 1 0 1075.024974
Spain 46.736782 4454.496 4871.587 0 1 0 1892.331932

Declaration of competing interest Appendix B

The authors declare that they have no known competing finan- North America
cial interests or personal relationships that could have appeared pNAp (t ) = p1 t n + p2 t n−1 + p3 t n−2 + · · · + pn t + pn+1 ; n = 14
to influence the work reported in this paper. pNAd (t ) = p1 t n + p2 t n−1 + p3 t n−2 + · · · + pn t + pn+1 ; n = 15
pNAq (t ) = p1 t n + p2 t n−1 + p3 t n−2 + · · · + pn t + pn+1 ; n = 47
Acknowledgments
South America
This study is supported by JSPS, Japan KAKENHI (Grants-in-Aid pSAp (t ) = p1 t n + p2 t n−1 + p3 t n−2 + · · · + pn t + pn+1 ; n = 23
for Scientific Research) #JP20K11955. pSAd (t ) = p1 t n + p2 t n−1 + p3 t n−2 + · · · + pn t + pn+1 ; n = 23
pSAq (t ) = p1 t n + p2 t n−1 + p3 t n−2 + · · · + pn t + pn+1 ; n = 47
Appendix A Africa
pAfp (t ) = p1 t n + p2 t n−1 + p3 t n−2 + · · · + pn t + pn+1 ; n = 47
See Table A.1. pAfd (t ) = p1 t n + p2 t n−1 + p3 t n−2 + · · · + pn t + pn+1 ; n = 47
14 A. Hernandez-Matamoros, H. Fujita, T. Hayashi et al. / Applied Soft Computing Journal 96 (2020) 106610

pAfq (t ) = p1 t n + p2 t n−1 + p3 t n−2 + · · · + pn t + pn+1 ; n = 47 [9] Hiteshi Tandon, Prabhat Ranjan, Tanmoy Chakraborty, Vandana Suhag,
Oceania Coronavirus (COVID-19): ARIMA based time-series analysis to forecast near
future, 2020, arXiv:2004.07859.
pOp (t ) = p1 t n + p2 t n−1 + p3 t n−2 + · · · + pn t + pn+1 ; n = 47
[10] Lutfi Bayyurt, Burcu Bayyurt, Forecasting of COVID-19 Cases and Deaths
pOd (t ) = p1 t n + p2 t n−1 + p3 t n−2 + · · · + pn t + pn+1 ; n = 47 Using ARIMA Models, medRxiv 2020.04.17.20069237.http://dx.doi.org/10.
pOq (t ) = p1 t n + p2 t n−1 + p3 t n−2 + · · · + pn t + pn+1 ; n = 47 1101/2020.04.17.20069237.
Asia [11] Simon James Fong, Gloria Li, Nilanjan Dey, Rubén González Crespo, Enrique
pAsp (t ) = p1 t n + p2 t n−1 + p3 t n−2 + · · · + pn t + pn+1 ; n = 14 Herrera-Viedma, Composite Monte Carlo decision making under high
uncertainty of novel coronavirus epidemic using hybridized deep learning
pAsd (t ) = p1 t n + p2 t n−1 + p3 t n−2 + · · · + pn t + pn+1 ; n = 40 and fuzzy rule induction, Appl. Soft Comput. (ISSN: 1568-4946) 93 (2020)
pAsq (t ) = p1 t n + p2 t n−1 + p3 t n−2 + · · · + pn t + pn+1 ; n = 50 106282, http://dx.doi.org/10.1016/j.asoc.2020.106282.
Europe [12] R.K. Singh, M. Rani, A.S. Bhagavathula, R. Sah, A.J. Rodriguez-Morales, H.
pEp (t ) = p1 t n + p2 t n−1 + p3 t n−2 + · · · + pn t + pn+1 ; n = 12 Kalita, C. Nanda, S. Sharma, Y.D. Sharma, A.A. Rabaan, J. Rahmani, P. Kumar,
pEd (t ) = p1 t n + p2 t n−1 + p3 t n−2 + · · · + pn t + pn+1 ; n = 50 Prediction of the COVID-19 pandemic for the top 15 affected countries:
Advanced autoregressive integrated moving average (ARIMA) model, JMIR
pEq (t ) = p1 t n + p2 t n−1 + p3 t n−2 + · · · + pn t + pn+1 ; n = 50 Public Health Surv. 6 (2) (2020) e19115, http://dx.doi.org/10.2196/19115.
[13] Xingde Duan, Xiaolei Zhang, ARIMA modelling and forecasting of irreg-
References ularly patterned COVID-19 outbreaks using Japanese and South Korean
data, 2020, 105779, http://dx.doi.org/10.1016/j.dib.2020.105779, (ISSN
[1] World Health Organization, Coronavirus disease (COVID-19) outbreak 2352-3409).
situation retrieved from: OnlineResource. [14] Andrea L. Bertozzi, Elisa Franco, George Mohler, Martin B. Short, Daniel
[2] Ali Narin, Ceren Kaya, Ziynet Pamuk, Automatic detection of coronavirus Sledge, The challenges of modeling and forecasting the spread of COVID-
disease (COVID-19) using X-ray images and deep convolutional neural 19, Proc. Natl. Acad. Sci. 117 (29) (2020) 16732–16738, http://dx.doi.org/
networks, 2020, arXiv, 003.10849. 10.1073/pnas.2006520117.
[3] Chen Lin, Yuxiao Ding, Bin Xie, Zhujian Sun, Xiaogang Li, Zixian Chen, Meng [15] Lixiang Li, Zihang Yang, Zhongkai Dang, Cui Meng, Jingze Huang, Haotian
Niu, Asymptomatic novel coronavirus pneumonia patient outside Wuhan: Meng, Deyu Wang, Guanhua Chen, Jiaxuan Zhang, Haipeng Peng, Yiming
The value of CT images in the course of the disease, Clin. Imaging (ISSN: Shao, Propagation analysis and prediction of the COVID-19, Infect. Dis.
0899-7071) 63 (2020) 7–9, http://dx.doi.org/10.1016/j.clinimag.2020.02. Model. (ISSN: 2468-0427) 5 (2020) 282–292, http://dx.doi.org/10.1016/j.
008. idm.2020.03.002.
[4] BioSpace, Quotient Sciences and CytoAgents Accelerate Potential Treatment [16] Kenji Mizumoto, Gerardo Chowell, Transmission potential of the novel
for COVID-19 Cytokine Storm, retrieved from: OnlineResource. coronavirus (COVID-19) onboard the diamond Princess Cruises Ship, Infect.
[5] Domenico Benvenuto, Marta Giovanetti, Lazzaro Vassallo, Silvia Angeletti, Dis. Model. (ISSN: 2468-0427) 5 (2020) 264–270, http://dx.doi.org/10.1016/
Massimo Ciccozzi, Application of the ARIMA model on the COVID-2019 j.idm.2020.02.003.
epidemic dataset, in: Data in Brief, Vol. 29, 2020, 105340, http://dx.doi. [17] G.E.P. Box, G.M. Jenkins, Time series analysis, in: Forecasting and Control,
org/10.1016/j.dib.2020.105340, (ISSN 2352-3409). Holden-Day, San Francisco, 1976.
[6] Fong Simon, Gloria Li, Nilanjan Dey, Ruben Gonzalez Crespo, Enrique [18] A. Max Roser, Hannah Ritchie, Esteban Ortiz-Ospina, Joe Hasell, Coro-
Herrera-Viedma, Finding an accurate early forecasting model from small navirus pandemic (COVID-19), 2020, Published online at OurWorldIn-
dataset: A case of 2019-nCoV novel coronavirus outbreak, Int. J. Interact. Data.org. Retrieved from: ’https://ourworldindata.org/coronavirus’ [Online
Multimed. Artif. Intell. 6 (2020) 132–140, http://dx.doi.org/10.9781/ijimai. Resource].
2020.02.002. [19] United Nations, Department of Economic and Social Affairs, Population
[7] Gaetano Perone, An ARIMA model to forecast the spread and the final size Dynamics, Retreived from: OnlineResource.
of COVID-2019 epidemic in Italy, 2020, arXiv:2004.00382. [20] J. Fattah, L. Ezzine, Z. Aman, H. El Moussami, A. Lachhab, Forecasting of
[8] Guorong Ding, Xinru Li, Yang Shen, Brief Analysis of the ARIMA model on demand using ARIMA model, Int. J. Eng. Bus. Manag. 10 (2018).
the COVID-19 in Italy, medRxiv 2020.04.08.20058636. http://dx.doi.org/10. [21] S.L. Ho, M. Xie, The use of ARIMA models for reliability forecasting and
1101/2020.04.08.20058636. analysis, Comput. Ind. Eng. (ISSN: 0360-8352) 35 (1–2) (1998) 213–216,
http://dx.doi.org/10.1016/S0360-8352(98)00066-7.
[22] J. Serrà, J.L. Arcos, An empirical evaluation of similarity measures for time
series classification, Knowl. Based Syst. 67 (2014) 305–314.