COVID-19 Pandemic Prediction using Time Series Forecasting Models
CCOVID-19 Pandemic Prediction using Time SeriesForecasting Models * Note: This paper is accepted in the 11th ICCCNT 2020 conference. The final version of this paper will appear in the conference proceedings.
Naresh Kumar
Department of Information TechnologyDelhi Technological University
Delhi, [email protected]
Seba Susan
Department of Information TechnologyDelhi Technological University
Delhi, Indiaseba [email protected]
Abstract —Millions of people have been infected and lakhs ofpeople have lost their lives due to the worldwide ongoing novelCoronavirus (COVID-19) pandemic. It is of utmost importanceto identify the future infected cases and the virus spread rate foradvance preparation in the healthcare services to avoid deaths.Accurately forecasting the spread of COVID-19 is an analyticaland challenging real-world problem to the research community.Therefore, we use day level information of COVID-19 spreadfor cumulative cases from whole world and 10 mostly affectedcountries; US, Spain, Italy, France, Germany, Russia, Iran,United Kingdom, Turkey, and India. We utilize the temporal dataof coronavirus spread from January 22, 2020 to May 20, 2020.We model the evolution of the COVID-19 outbreak, and performprediction using ARIMA and Prophet time series forecastingmodels. Effectiveness of the models are evaluated based onthe mean absolute error, root mean square error, root relativesquared error, and mean absolute percentage error. Our analysiscan help in understanding the trends of the disease outbreak, andprovide epidemiological stage information of adopted countries.Our investigations show that ARIMA model is more effective forforecasting COVID-19 prevalence. The forecasting results havepotential to assist governments to plan policies to contain thespread of the virus.
Index Terms —ARIMA, COVID-19, Pandemic, Prophet, Timeseries forecasting
I. I
NTRODUCTION
The novel Coronavirus (COVID-19) has infected millions ofpeople worldwide since it emerged from China in December2019. COVID-19 has very high mutating capability, and it canspread very easily. Infected people from this virus suffer fromsevere respiratory problems, and may develop serious illnessif suffering from chronic diseases like cardiovascular diseaseor diabetes or having weak immune system or being olderin age [14]. World health organization (WHO) declared on11th March, 2020, the outbreak of COVID-19 as a pandemic.There are challenges to contain the disease because an infectedperson shows symptom after a long time or no sign of thedisease. At present, no vaccination has been discovered forCOVID-19. In this situation, social distancing, identifying thepositive cases using testing at large scale, and containment ofinfected person is the only option to prevent the spreading ofthe virus [6]. The spread of COVID-19 can be classified [8] under threemajor stages- 1. Local outbreak: at this stage, spreading chainof the virus among the people can be tracked, and the sourceof infection can be found out. The cases in this stage mostlyrelate to within family or friends, or the local exposure. 2.Community transmission: at this stage, source of the chain ofinfected people cannot be found out. The infected cases growthrough cluster transmission in the communities. 3. Large scaletransmission: at this stage, the virus spreads rapidly to otherregions of a country due to uncontrolled mobility of people atlarge scale.Due to high scale community impact and easy spreadingworldwide, national governments imposed lockdown to controlthe spread of corona virus. As of 20th May, 2020, 4996472cases have been confirmed, 1897466 cases have recovered,2328115 deaths have been reported, and 2770891 active caseshave been identified worldwide. The statistical data is collectedfrom [4], and the number of COVID-19 cases is calculatedbetween 22 Jan, 2020 to 20 May 2020.As no vaccine has been discovered of the disease, somotivation behind this paper is to model spreading of thecorona virus, and predict the impact to optimize the plan-ning to manage the various services and resources for thepublic by the governments. Some of the studies [10], [12],[13], [17] have been published showing statistical analysis,modeling, and artificial intelligence to contain the spreadof the virus, and highlight impacts in coming days. Theseearly studies are carried out using very limited informationavailable at early stage of the outbreak. Now, the virus hasspread at large scale, and much information is available forthe analysis. Predictive analysis of COVID-19 has become ahot research area to support health services and governmentsto plan and contain the spread of the infectious disease [3].Modeling and forecasting the daily spread behavior of thevirus can assist the health systems to be ready to accommodatethe upcoming number of patients. Accurate forecasting ofthe disease is a matter of concern because it may impactgovernments policy, containment rules, health system, andsocial life. Regarding this context, we explore the predictive a r X i v : . [ phy s i c s . s o c - ph ] J u l apability of the ARIMA [16], and Prophet [15] forecastingmodels. The models are widely used and accepted due totheir more accurate forecasting capability. We use the daylevel cumulative cases of COVID-19 worldwide and 10 mostlyaffected countries; US, Spain, Italy, France, Germany, Russia,Iran, United Kingdom, Turkey, and India for our analysisstudy.The objective of this paper is to provide evaluative study ofprediction models using COVID-19 cases, and forecasting theimpact of the virus in the affected countries, and worldwide.We present trend analysis of COVID-19 cases, and comparedthe performance of the models using the metrics such as themean absolute error (MAE), root mean square error (RMSE),root relative squared error (RRSE), and mean absolute per-centage error (MAPE). We generate forecasting results forCOVID-19 confirmed, active, recovered, and death cases. Theresults show that ARIMA outperformed the Prophet model.The rest of the paper is organized as follows. SectionII presents a literature survey. Section III provides trendanalysis of COVID-19 cases. Section IV describes overviewof time series forecasting models. Modeling framework andused COVID-19 dataset are described in section V. Statisticalanalysis and model evolution are presented in section VI.Section VII concludes the paper.II. L ITERATURE SURVEY
Intensive research work is going on to evaluate and containthe worldwide disaster of COVID-19 on the human race.Research studies include predictions about the future cases[17], and analysis of the variables responsible for spread ofthe coronavirus [11].In the literature, time series forecasting problems havebeen studied widely in which COVID-19 forecasting is anemerging problem. Forecasting models can be used to forecastthe impact of the disease on the community which can helpto control the epidemic. In [13], authors have performedforecasting evaluation study of the models using COVID-19day level cases from 10 mostly affected states from Brazil.According to the authors, the stacking ensemble and SVRperformed better as compared to ARIMA, CUBIST, RIDGE,and RF models for the adopted criteria. In [5], the authorhas developed ARIMA(p,d,q) model and studied the COVID-19 epidemiological trend in the three most affected countries;Spain, Italy, and France of Europe continent using the databetween 21 Feb to 15 April 2020. The author studied thevarious orders (p, d, q) of the model, and selected best per-forming order based on lowest values of MAPE for the threecountries. He has suggested that ARIMA models are suitablefor forecasting the COVID-19 prevalence for the upcomingdays. Chintalapudi et al. [2], adopted seasonal ARIMA modelfor forecasting of COVID-19 cases in Italy using the datatill 31st March 2020. They have analysed the impact of twomonths lockdown in Italy, and observed decrement in theconfirmed cases and growth in the recovered cases due tolockdown. Alabi et al. [1] have adopted the Facebook Prophetmodel to forecast spread of COVID-19. They have performed prediction for confirmed and death cases. Their forecastingaccuracy of Prophet was 79.6% for the data from WHObetween 7th April to 3rd May 2020. Parikshit et al. [9] havepresented medical perspective of COVID-19, and predictionusing Prophet model. They have recommended Prophet forprediction due to open source algorithm, accuracy, and fasterdata fitting. Using the Prophet model, they have predictedinfected cases worldwide as 1.6 million by the end of May2020, and 2.3 million by the end of June 2020.III. COVID-19 T
RENDS
We have collected cumulative day level data of COVID-19 cases from github repository [7] latest by 20 May, 2020.The repository is supported by ESRI living atlas team, AppliedPhysics Lab (APL), and maintained by the Center for SystemsScience and Engineering (CSSE), both at Johns HopkinsUniversity, USA. The repository contains worldwide COVID-19 reported cases starting from 22 January, 2020 on day-to-day basis. We study COVID-19 confirmed cases, recoveredcases, death cases, and active cases for 10 adopted countriesand worldwide. The adopted countries are the badly affectedcountries by the virus in the world latest by 20 May, 2020.The impact of coronavirus in 10 adopted countries from 1stMarch, 2020 to 20th May, 2020 is shown in Fig. 1. Trendsof confirmed, recovered, deaths, and active cases in adoptedcountries show that impact of the virus from highest to lowestis in the sequence of labeling order of the countries. The figureclearly shows that US is the most affected country, it hashighest confirmed and death cases. Except US and Russia,other countries are able to flatten the graphs after some levelof outbreak.Fig. 1: Trend of Confirmed, Recovered, Deaths, and Activecases of COVID-19 in adopted countries.Fig. 2 shows worldwide spread trend of coronavirus from 22January, 2020 to 20 May, 2020. The trend explains that growthof the virus is almost exponential after mid-March 2020.Fig. 3 shows recovery rate of COVID-19 in the adoptedcountries and worldwide from 22 Jan 2020 to 20 May 2020.From the figure, we can say that highest recovery is done byig. 2: Trend of COVID-19 cases worldwide.Iran whereas Turkey shows exponential growth in the recoverycases, and other countries follows the recovery pattern similarto growth pattern of confirmed cases. Various studies [6],[14], [17] show that the disease recovers automatically aftersometime but causes major health problem which can lead todeath, if not taken care.Fig. 3: Recovery rate of COVID-19 cases.Fig. 4 shows fatality rate of COVID-19 patients worldwideand in adopted countries from 22 Jan 2020 to 20 May 2020.The figure shows that Iran faced highest death rate which waslater over taken by US along with France. Spain has alsoshown significant death rate over the period of time. Othercountries were able to control deaths to some extent usinglockdowns or following social distancing etc. The virus hastaken lives of many people worldwide.The historical data depicts that the COVID-19 badly af-fected the countries which do not impose lockdowns or donot followed social distancing. Some variations in virus spreadrate, recovery rate, and death rate can be seen in differentcountries based on population density, available health systemin a country, testing capability, and action taken to contain theoutbreak.IV. T
IME S ERIES FORECASTING MODELS
Time series forecasting models are used to predict thefuturistic outcomes based on historical information. We haveadopted ARIMA and Facebook Prophet (FBProphet) modelin our evaluative and forecasting study. An overview of themodels is given in the following sections. Fig. 4: Fatality rate of COVID-19 cases.
A. Autoregressive Integrated Moving Average (ARIMA)
ARIMA(p,d,q) [16] is composite of Autoregressive (AR)model, Moving Average (MA) model, and 'I' stands for in-tegration; where p is order of autoregression, d is order ofdifferencing, q is order of moving average.The AR(p) model is defined as a linear process given as thefollowing equation. z t = α + φ z t − + φ z t − + ... + φ p z t − p + w t (1)where z t − , z t − , z t − p are the lags (past values); φ , φ , ...φ p are lag coefficients which are estimated bythe model; w t is the white noise, and α is defined as follows. α = (cid:32) − p (cid:88) i =1 φ i (cid:33) µ (2)where µ is mean of the process.The MA(q) model is defined as the following equation. z t = α + w t + θ w t − + θ w t − + ... + θ q w t − q (3)Where w t , w t − , ..w t − q are error terms of the model for therespective lags i.e. z t , z t − , ...z t − q .ARIMA is able to fit if the data is stationary i.e. data meanand standard deviation is constant. The differencing parameterd is the order of transformation to make dataset stationary.Second order differencing is shown in the following equation. z t = ( Z t − Z t − ) − ( Z t − − Z t − ) = Z t − Z t − + Z t − (4)Finally the equation for the ARIMA(p,d,q) is defined asfollows. z t = α + p (cid:88) i =1 φ i z t − i + w t + q (cid:88) j =1 θ j w t − j (5) . Facebook Prophet Taylor et al. [15] proposed the Facebook Prophet(FBProphet) which uses several non-linear and linear methodsas components with time as a regressor. Prophet is developedand released as open source software by data science team ofFacebook. The model ignores the temporal dependence of thedata, and training is framed just as curve-fitting exercise. So,irregular observations are also allowed in a dataset. The modeloffers various advantages like it can accommodate multipleperiod seasonality; it can accommodate custom and knownholidays; it provides flexibility by offering two options fortrend: 1. a piecewise linear model, 2. a saturating growthmodel; and the model fits very fast. The model includes onemore term holidays as the components of time series, so atime series can be defined by the following equation. z t = T t + S t + H t + (cid:15) t (6)where T t is trend, S t is seasonality, H t is holiday, and (cid:15) t is error term.V. D ATASET AND F ORECASTING FRAMEWORK
This section describes about the dataset we have used toforecast COVID-19 cases of adopted countries, and world-wide. It also describes the modeling framework which we havefollowed.
A. Modeling Dataset
It is observed from the trends in section-III that rate of thereported COVID-19 cases in each country increases with timeand flattens after sometime if large scale testing is performed,lockdown is imposed, and containment is followed. For ourstudy, we disaggregated the available day level data of theadopted 10 countries. We discarded the initial 5 days data foreach country in our study. The reason to discarding the initialsamples is that testing of the samples grows slowly in startingphase which does not depict the actual rate of the spread. Theutilized samples detail is given in Table I. The end date of thecollected samples is 20 May 2020. In this study, we consider80 percentage samples for training and 20 percentage samplesfor testing the models for each country and worldwide data.TABLE I: Total COVID-19 samples used for modeling till 20May, 2020.
Region Sample size (Days) Start Date Confirmed Recovered Deaths Active
Worldwide 120 2020-01-22 4996472 1897466 328115 2770891US 115 2020-01-27 1551853 294312 93439 1164102Spain 105 2020-02-06 232555 150376 27888 54291Italy 106 2020-02-05 227364 132282 32330 62752France 113 2020-01-29 181700 63472 28135 90093Germany 110 2020-02-01 178473 156966 8144 13363Russia 106 2020-02-05 308705 85392 2972 220341Iran 87 2020-02-24 126949 98808 7183 20958UK 106 2020-02-05 249619 1116 35786 212717Turkey 66 2020-03-16 152587 113987 4222 34378India 107 2020-02-04 112028 45422 3434 63172
B. Forecasting Framework
Fig. 5 describes about the adopted framework for predictionand analysis of the COVID-19 cases using ARIMA, andFBProphet models. For the analysis, we have split the datasetsof confirmed, active, recovered, and death cases into trainingand testing. We performed prediction after removing trendswherever applicable, and used statistical measures to evaluatethe performance.Fig. 5: Framework to evaluate the forecasting models.
C. Performance Measures
To evaluate the prediction models, we use the followingstatistical measures.Mean Absolute Error (MAE):
M AE = 1 N N (cid:88) k =1 | z k − z ˆ k | (7)Root Mean Square Error (RMSE): RM SE = (cid:118)(cid:117)(cid:117)(cid:116) N N (cid:88) k =1 ( z k − z ˆ k ) (8)Root Relative Squared Error (RRSE): RRSE = (cid:118)(cid:117)(cid:117)(cid:116) (cid:80) Nk =1 ( z ˆ k − z k ) (cid:80) Nk =1 ( z ¯ − z k ) where z ¯= 1 N N (cid:88) k =1 z k (9)Mean Absolute Percentage Error (MAPE): M AP E = 100 N N (cid:88) k =1 (cid:12)(cid:12)(cid:12)(cid:12) z k − z ˆ k z k (cid:12)(cid:12)(cid:12)(cid:12) (10)where z k denotes actual value and z ˆ k denotes predictedvalue for the k th instance. z ¯ stands for the average value ofz, and N is the total number of testing samples.I. R ESULTS AND DISCUSSION
The adopted framework is implemented in Python 3.8,and we have used ARIMA and Prophet models from openlyavailable packages statsmodels and fbProphet respectively. Wehave performed our experiments in Intel Core i5 processorclocked at 2.40 GHz, 8 GB RAM, and 4GB NVIDIA GTX-1650 GPU. In this section, we will discuss about forecastingaccuracy of adopted models for active, recovered, deaths, andconfirmed cases.
A. Forecasting of active cases
Active cases are the number of infected people who areunder medical supervision. Active cases are derived as shownin the following equation.
Active = Conf irmed − Recovered − Deaths (11)We use ARIMA and FBProphet models to predict the futurecases. ARIMA can be used for prediction if data is stationary.It is clear from the trends in section-III that data of active casesis not in stationary form. So, we have applied techniques toconvert the data into stationary form for ARIMA evaluation.We have applied square root scaling and one lag differencingto convert the data into stationary form. We have performeddicky-fuller test to check stationarity of the data. We also haveused PACF and ACF plots to identify appropriate values of qand p order of ARIMA. The FBProphet is applied directly onactual data. Forecasting accuracy results for active cases of10 adopted countries and worldwide are shown in Table II.We have mentioned order of ARIMA along with the accuracyresults in the tables. The mentioned ARIMA order performedbetter to fit the model accurately. Best MAPE scores are 0.586and 1.481 for US and UK data by ARIMA and FBProphetrespectively. From the results, we can clearly say that ARIMAhas far better performance as compared to FBProphet modelwith respect to all types of error measures i.e. MAE, RMSE,RRSE, and MAPE.
B. Forecasting of recovered cases
To predict and analyse the recovery rate of the disease, wehave performed evaluative study of the adopted models usingrecovery data of the 10 adopted countries, and worldwide.We can see from the trends in section III that recovery datais also non-stationary. So, we have performed stationaritytechniques similar to as discussed in section-VI(A) to evaluatethe ARIMA model. We have applied FBProphet directly onactual data to fit the model and generated the forecastingresults. Table III shows accuracy results of the models for therecovered cases. Best MAE results are 78.19 and 69.11 for UKdata by ARIMA and FBProphet respectively. Results show thatARIMA prediction almost matches the actual values whereasFBProphet did not perform as well. We can see that maximumMAPE value of ARIMA is 15.6 and minimum value is 2.5which are very much acceptable to generate forecasting results,whereas maximum and minimum MAPE for FBProphet are31.822 and 3.759 respectively. TABLE II: Performance results of the models for COVID-19active cases in adopted countries.
Region Model MAE RMSE RRSE MAPE
Worldwide ARIMA(9,1,2) 19141.89 21377.14 0.086 0.816FBProphet 168452.05 182230.63 0.706 6.943US ARIMA(10,1,3) 5732.16 8050.31 0.079 0.586FBProphet 95766.22 108424.76 1.07 9.12Spain ARIMA(8,1,4) 2191.68 2603.02 0.346 3.293FBProphet 67132.86 69748.42 9.274 109.40Italy ARIMA(9,1,3) 3197.25 4266.60 0.320 3.411FBProphet 26934.34 30963.76 2.325 35.55France ARIMA(5,1,4) 10974.15 11489.85 6.166 11.75FBProphet 44596.16 48195.48 25.864 48.340Germany ARIMA(11,1,4) 2114.09 2597.193 0.407 9.052FBProphet 50902.42 52259.90 8.197 277.26Russia ARIMA(10,1,2) 6456.26 6786.96 0.158 4.238FBProphet 36430.36 40232.57 0.936 20.748Iran ARIMA(4,1,2) 328.28 379.79 0.147 2.202FBProphet 12856.19 12902.11 5.009 82.503UK ARIMA(4,1,2) 8090.84 8637.25 0.375 4.66FBProphet 2954.65 4649.43 0.202 1.481Turkey ARIMA(8,1,2) 3631.37 3655.74 0.884 9.485FBProphet 59801.55 60725.11 14.678 158.59India ARIMA(11,1,5) 7007.09 7330.06 0.61 16.74FBProphet 10245.17 12085.37 1.005 21.429
TABLE III: Performance results of the models for COVID-19recovered cases in Adopted countries.
Region Model MAE RMSE RRSE MAPE
Worldwide ARIMA(9,1,2) 34932.99 36992.53 0.128 2.523FBProphet 185584.71 214741.61 0.712 12.49US ARIMA(5,1,2) 31899.89 33109.68 0.667 15.635FBProphet 53970.19 57816.45 1.165 24.174Spain ARIMA(8,1,4) 9683.45 9774.06 0.786 7.361FBProphet 3021.53 3766.69 0.303 2.22Italy ARIMA(9,1,3) 12910.06 13078.23 0.693 12.78FBProphet 8721.87 10057.88 0.533 7.881France ARIMA(3,1,1) 5780.87 5853.29 1.21 10.574FBProphet 7323.90 8362.88 1.729 12.613Germany ARIMA(5,1,3) 13702.61 13901.04 1.287 9.808FBProphet 25017.26 28763.20 2.664 16.969Russia ARIMA(4,1,0) 2376.69 3212.50 0.141 5.103FBProphet 26988.80 33858.56 1.484 60.964Iran ARIMA(1,1,1) 4213.14 4496.75 0.736 4.933FBProphet 5638.72 6037.87 0.988 6.267UK ARIMA(4,1,2) 78.19 91.12 1.177 8.311FBProphet 69.11 79.44 1.026 7.326Turkey ARIMA(8,1,2) 4242.09 4333.57 0.44 4.321FBProphet 45986.27 46211.42 4.688 45.536India ARIMA(2,1,0) 721.17 1066.65 0.096 2.911FBProphet 11395.90 14381.55 1.295 42.882
C. Forecasting of death cases
Coronavirus has taken many lives. So, it is necessary toanalyse the fatality rate of the virus, and forecasting tohighlight future cases which can guide governments to actin advance. In this section, we have evaluated the forecast-ing models for death cases of the adopted countries, andworldwide. We have converted the non-stationary fatality datainto stationary form to fit the ARIMA model similar to asdiscussed in section-VI(A). FBProphet model is applied on theactual data to forecast the prediction results. Table IV showsprediction accuracy of the models for the fatality cases. Wecan see that prediction errors of ARIMA are very less whereasABLE IV: Performance results of the models for COVID-19fatality cases in adopted countries
Region Model MAE RMSE RRSE MAPE
Worldwide ARIMA(9,1,2) 661.98 821.20 0.026 0.257FBProphet 21666.12 24874.45 0.735 7.465US ARIMA(2,1,0) 1924.08 1988.71 0.19 2.571FBProphet 4799.16 5856.54 0.56 5.751Spain ARIMA(2,1,0) 940.85 953.08 0.907 3.577FBProphet 2573.67 2961.76 2.818 9.525Italy ARIMA(2,1,0) 1240.10 1254.32 0.892 4.128FBProphet 3008.94 3433.46 2.443 9.703France ARIMA(3,1,1) 1335.79 1355.02 0.983 5.139FBProphet 6545.93 7270.98 5.274 24.382Germany ARIMA(1,1,0) 318.04 341.57 0.668 4.382FBProphet 1446.64 1668.89 3.262 18.761Russia ARIMA(2,1,0) 43.31 48.98 0.082 2.252FBProphet 628.39 709.50 1.184 30.597Iran ARIMA(1,1,1) 836.66 836.86 2.929 12.487FBProphet 257.44 291.70 1.021 3.759UK ARIMA(2,1,0) 959.53 984.02 0.343 3.119FBProphet 4171.84 4867.84 1.699 12.639Turkey ARIMA(8,1,2) 113.54 117.61 0.619 2.909FBProphet 280.83 312.96 1.647 6.945India ARIMA(2,1,0) 48.94 60.75 0.085 2.704FBProphet 771.35 897.58 1.26 31.822
FBProphet prediction have high error factor in the results. Theresults suggest that ARIMA can be used for actual forecastingof the cases to plan the services accordingly. (a) ARIMA Forecasting for US confirmed cases(b) FBProphet Forecasting for US confirmed cases
Fig. 6: Actual and predicted values plots of ARIMA andFBProphet for covid-19 confirmed cases of US.
D. Forecasting confirmed cases
In this section, we have highlighted the fitted accuracy ofthe models using confirmed cases. For this analysis, we havechosen only two countries; US and India. The results of US and India by both the models ARIMA and FBProphet areshown in Fig. 6 and Fig. 7 respectively. We have showntraining and testing data split using vertical line in the figures.Forecasted and actual data are plotted together to visualizethe fitting accuracy of the models. We can see that FBProphetmodel is able to fit well in case of US data as shown in Fig. 6whereas ARIMA is able to perform well in case of India dataas shown in Fig. 7. FBProphet adopts successively progression,and avoid outliers during modeling and forecasting. The resultsalso depicts that FBProphet can fit well in case of less datawhereas ARIMA requires sufficient data to model and predictthe results. (a) ARIMA Forecasting for India confirmed cases(b) FBProphet Forecasting for India confirmed cases
Fig. 7: Actual and predicted values plots of ARIMA andFBProphet for covid-19 confirmed cases of India.VII. C
ONCLUSION AND F UTURE WORK
WHO has declared COVID-19 as pandemic because it hasinfected most of the countries, and it is a major threat to humanrace. In this paper, we have done analysis and predictionstudy of the disease using widely accepted forecasting models;ARIMA and FBProphet. We have collected COVID-19 dataof 10 highly affected countries US, Spain, Italy, France,Germany, Russia, Iran, UK, Turkey, India, and worldwidelatest by May 20, 2020. For the most of the countries data,ARIMA has better performed compared to Prophet on scaleof MAE, RMSE, RRSE, and MAPE error matrices. Thetrend analysis shows rapid growth in the infected cases, andprediction study shows great rise in the expected active,recovered, and death cases worldwide. However, lockdownsand containment policies may affect the prediction results. Theadopted models have performed well but it limits our study tothe effectiveness of the models, which can be further improvedsing ensemble of multiple prediction models. The obtainedforecasting results further can be improved by taking variousvariables into account like population density, weather, healthsystem, patient history etc. using deep learning techniques, andartificial intelligence. R
EFERENCES[1] Rasheed Omobolaji Alabi, Akpojoto Siemuri, and Mohammed Elmus-rati. Covid-19: Easing the coronavirus lockdowns with caution. medRxiv ,2020.[2] Nalini Chintalapudi, Gopi Battineni, and Francesco Amenta. Covid-19 disease outbreak forecasting of registered and recovered cases aftersixty day lockdown in italy: A data driven model approach.
Journal ofMicrobiology, Immunology and Infection , 2020.[3] IHME COVID, Christopher JL Murray, et al. Forecasting covid-19impact on hospital bed-days, icu-days, ventilator-days and deaths byus state in the next 4 months.
MedRxiv , 2020.[4] Ensheng Dong, Hongru Du, and Lauren Gardner. An interactive web-based dashboard to track covid-19 in real time.
The Lancet infectiousdiseases , 20(5):533–534, 2020.[5] Duccio Fanelli and Francesco Piazza. Analysis and forecast of covid-19 spreading in china, italy and france.
Chaos, Solitons & Fractals ,134:109761, 2020.[6] Milad Haghani, Michiel CJ Bliemer, Floris Goerlandt, and Jie Li.The scientific literature on coronaviruses, covid-19 and its associatedsafety-related research dimensions: A scientometric analysis and scopingreview.
Safety Science , page 104806, 2020.[7] Github Inc. Covid-19 cases. https://github.com/cssegisanddata/covid-19(accessed in 21 may, 2020).[8] Lin Jia, Kewen Li, Yu Jiang, Xin Guo, et al. Prediction and analysis ofcoronavirus disease 2019. arXiv preprint arXiv:2003.05447 , 2020.[9] Parikshit N Mahalle, Nilesh P Sable, Namita P Mahalle, and Gitanjali RShinde. Data analytics: Covid-19 prediction using multimodal data. , 2020.[10] Manotosh Mandal, Soovoojeet Jana, Swapan Kumar Nandi, AnupamKhatua, Sayani Adak, and TK Kar. A model based study on thedynamics of covid-19: Prediction and control.
Chaos, Solitons &Fractals , page 109889, 2020.[11] Barbara Oliveiros, Liliana Caramelo, Nuno C Ferreira, and FranciscoCaramelo. Role of temperature and humidity in the modulation of thedoubling time of covid-19 cases. medRxiv , 2020.[12] Ratnabali Pal, Arif Ahmed Sekh, Samarjit Kar, and Dilip K Prasad.Neural network based country wise risk prediction of covid-19. arXivpreprint arXiv:2004.00959 , 2020.[13] Matheus Henrique Dal Molin Ribeiro, Ramon Gomes da Silva, Vi-viana Cocco Mariani, and Leandro dos Santos Coelho. Short-termforecasting covid-19 cumulative confirmed cases: Perspectives for brazil.
Chaos, Solitons & Fractals , page 109853, 2020.[14] Catrin Sohrabi, Zaid Alsafi, Niamh ONeill, Mehdi Khan, Ahmed Ker-wan, Ahmed Al-Jabir, Christos Iosifidis, and Riaz Agha. World healthorganization declares global emergency: A review of the 2019 novelcoronavirus (covid-19).
International Journal of Surgery , 2020.[15] Sean J Taylor and Benjamin Letham. Forecasting at scale.
The AmericanStatistician , 72(1):37–45, 2018.[16] Shaun S Wulff. Time series analysis: Forecasting and control.
Journalof Quality Technology , 49(4):418, 2017.[17] Xiaolei Zhang, Renjun Ma, and Lin Wang. Predicting turning point, du-ration and attack rate of covid-19 outbreaks in major western countries.