Infectious diseases are the leading cause of human mortality around the world. The occurrences of infectious diseases can be traced back to the Middle Ages. COVID-19, an unusual coronavirus, was discovered in Wuhan, China, in December 2019. The newly identified virus has been linked to 470 million confirmed infections [1]. Vaccinations for SARS-CoV-2 and antiviral measures such as wearing masks and avoiding large crowds are critical in preventing viral spread. Many vaccines have shown the efficacy of more than 95% in preventing SARS-CoV-2 symptoms. Approximately 80 M COVID-19 infections and 1 M coronavirus-related fatalities have been reported in the USA alone [1]. Diverse mathematical epidemic surveillance models [2–4] have been given in the literature [2]. Biological and disease mechanisms that use epidemiological methodologies are important in most modeling approaches to predict the progression of outbreaks and pandemics.

An effective method of accomplishing this is accurately predicting the number of active cases at any time. Time-series data include confirmed daily COVID-19 cases, recoveries, and deaths, to name a few instances. Time-series data are a series of numeric values measured at intervals equal in length (e.g., per minute, hour, or day) [3]. To anticipate the future dynamics of influenza, malaria tuberculosis, and other infectious diseases [4,5], different time series models were utilized to forecast COVID-19 in the United States, Italy, India, and other countries [6].

Despite this, no general concept for selecting models for projecting the spread of COVID-19 has been established. It is also recommended that multiple models can be used to forecast the spread of the epidemic in different states and under different conditions of the pandemic propagation. Deep learning models, for example, were demonstrated in [7] to have the lowest rates of forecast errors while tracking the dynamics of infection cases in the four nations under consideration [7,8]. The following are the main significant advantages of ARIMA: (1)

Dealing with tiny data.

In addition, the following are some of the characteristics of deep learning models: (1)

There are no pre-requisites (stationarity, no level shifts).

In this paper, the proposed methodology uses various data published across the globe for the active cases of patients identified with coronavirus infection (in this example), the total number of deaths recorded due to the virus impact, and the total number of cases that have been recovered. To ensure the precision of the Artificial Intelligence (AI) models, epidemiological time series are employed to monitor and manage the spread of infection and its containment. In addition, this work intends to investigate various ways to forecast an outbreak and identify abnormalities in historical data. This may be used without compromising the data accuracy, or periodicity, assisting management of the regions to combat the spread of infection.

As a result, data analytics and ML techniques may be used to develop a more comprehensive understanding of viruses. A comparative review of COVID-19 forecasting models, including ML techniques, is presented in this paper, along with examining the distribution and transmission of COVID-19 in various nations. Thus, this research has examined the empirical performance of several classical univariate time series models, as well as ML-based regression algorithms such as Logistic Regression, Lasso, Ridge, ElasticNet, Huber Regressor, Lasso Lars, K-Neighbors Regressor, Passive Aggressive Regressor, Random Forest Regressor, Extra Trees Regressor, SVR, AdaBoost Regressor, Bagging Regressor, Decision Tree Regressor, Gradient Boost, Croston, and NaiveBayes.

The following sections are organized as follows. First, a literature survey for some of the recently published related studies is introduced. The next section includes data description materials and methods. Subsequently, the discussion of modeling and forecasting outcomes are presented. The final section provides the recommendations, conclusions, and future works.

The purpose of time series forecasting as an asset is to ease the forecasting process and to create a convenient and easy-to-use application/tool for any kind of situation dealing with the need to have a futuristic visualization of the trend of the data involved in the planning of any domain.

COVID-19 scatter forecasting, or simply future impact projection, provides essential inputs for government, public health agencies, corporations, and citizens to schedule, organize, and manage an outbreak. As a result, most outbreak models used in the tracking and forecasting of COVID-19 are based on epidemiological patterns, such as susceptible, infected, and eliminated individuals [8]. The Susceptible-Exposed-Infected-Removed (SEIR), Susceptible-Exposed–Infected–Recovered–Dead (SEIRD), and extended Susceptible–Infected–Removed (SIR) [9] models incorporate variable-time quarantine procedures such as macro-isolation rules at the federal level and standard isolation policies. To forecast infection, several AI-based models have been published in the literature. These models include an interior searching algorithm and a multi-layer Artificial Neural Network (ANN) feedforward. Several trends in this field include the following models [10]: (1) modified stacks for transmission dynamics, (2) nonlinear hybrids for predicting affected, (3) recovered and lethal molecules, (4) agent-based AI simulation platform (EnerPol) for predicting growth and containment strategy, (5) multi-input deep convolutional neural network (CNN) for predicting the cumulative number of confirmed cases, (6) topological autoencoder for generating a similarity map of transmission dynamics, and (7) SEIR.

The work in [11] explores the contemporary pattern or trend of COVID-19 transmission in any highly impacted country like India and the regression analysis of other data (like different information for Indian data). In their work, five models were used in this study: Linear Regression (LR), Exponential Smoothing (ES), K-Nearest Neighbors (KNN), Random Forest (RF), and Support Vector Machines (SVM). Each model contains two types of predictions: (1) a newly positive number of eases verified and (2) several people killed.

In research conducted by author in [12], the authors have developed a new nonlinear deterministic model based on ordinary differential equations with six compartments, quarantine, and isolation. The model has a positive invariant area, while the balance points are investigated concerning the simple reproductive number in terms of its local stability. Furthermore, a new mathematical model for predicting the spreading of epidemics in Egypt has been suggested in [13]. In the proposed model, the number of cases was raised significantly after two months, concluded from exploring the data. Consequently, a series of proliferation prevention initiatives had to be adopted, such as implementing systematic prohibition, isolation, and social divergence policies over a specific timeframe.

Many researchers have attempted to forecast the outbreak of COVID-19 by using diverse mathematics methods like the classic SIR model and its derivatives [14–20]. This analysis simulated the outbreak from February 14 to April 11 in Isfahan Province of Iran and predicted the remaining direction with three different socially distinguishing scenarios. To anticipate the pandemic, further advanced models and a detailed understanding of the epidemic biomedical and epidemiological dimensions were needed [14].

In summary, the pandemic characteristics of COVID-19 are inconsistent with the SIR modeling system. For most of which quantitative evidence is not yet available, the mechanisms of this epidemic were subject to different parameters. The new approach suggested produced improved outcomes and showed the value of social distance [15]. Their research proposed ML and deep learning (DL) algorithms such as Long Short-Term Memory (LSTM), eXtreme Gradient Boost, and polynomial regression in forecasting to estimate COVID-19 count in advance and enhance precision [16].

To find out the forecast for the upcoming months, the whole data should be used to check for the pattern and stationarity of the data. The outliers can be treated using either Mean-Standard Deviation (MSD) or Median Absolute Deviation (MAD) [17]. According to [18], not much research was done on this particular application (to have thorough research on all the time series, ML, and deep learning models for the forecasting problem) using the different ML techniques. Once implemented, this algorithm could be used in any domain such as heart rate measurement, climate changes, forecasting of foods, sales of retail industries, etc. [19]. The researchers presented some of the well-known models used widely for forecasting demand, and an extensive comparison of their performances is made, and inferences are drawn [20–25]. Many different models are available in the research fields that vary from the statistical time series algorithms and regression-based ML algorithms rather than deep learning-based algorithms like the LSTM [21].

The methods used to determine and compare the results of Canadian Foundries’ orders are NN, RNN, and SVM [22]. In [23], many statistical approaches like the ARIMA are used along with the Artificial Neural Networks (ANNs) to use the linearity of the ARIMA and nonlinearity of the ANNs to make a hybrid model. The same was also proven in [24], so the linear and the nonlinearity of the data can be captured. The hybrid model or the ensemble of several models is standard for better accuracy since the well-known M-competition [24].

Various studies have been done to understand combining the models to give a better forecast. In an analysis of the time-series data, which has a powerful trend and seasonality, a comparative study is done with the ANN and other traditional methods, including the Winters Exponential Smoothing, Box-Jenkins ARIMA model, and multivariate regression [25]. In [26], the authors focused on the extensive work on analyzing the time series, ML-based regression models, and deep learning to combine in ensemble methods. A lot of algorithms such as Auto-Regressive time series models, also known as AR models, Simple Moving Average (SMA), Simple Exponential Smoothing (SES), Weighted Moving Average (WMA), and Holts-Winter Exponential Smoothing (HWES) models [27] are used. Some naïve forecasting techniques also impact forecast generation through the time series models. The ML models can be classified into linear and nonlinear models. Some of the algorithms that are used in the linear ML regression models are Linear (LR), Lasso, Ridge, Elastic Net (EN), Huber, Lasso Lars (LLARS), Passive-Aggressive (PA), and regression models. Also, in the case of the nonlinear ML regression models, K-Neighbors, Decision Tree, Extra Tree, SVR, AdaBoost, Bagging, Random Forest, Extra Trees, and Gradient Boosting regressor models are used. In addition, deep learning models such as LSTMs and Multilayer Perceptron (MLP) [28,29] are used.

The goal of implementing the time series forecasting algorithm is to create genericity of the modeling, which gives better accuracy and presents a highly stable prediction model that generates reliable forecasts for any relevant data type. In this work, we concentrate on assembling with a weighted average as the second level learner.

Numerical experiments on numerous datasets from various nations have been conducted to evaluate the suggested method and demonstrate the generalizability of the designed scheme. As clarified in Fig. 2, the confirmed dataset of the monthly mean resample is depicted in green, and the original data is displayed in blue. As a result, we need to determine whether or not the dataset is stationary at this point, as clarified in Fig. 3.

When the statistical characteristics of a dataset, such as the mean, variance, and autocorrelation, remain constant across time, the dataset is said to be stable. According to Fig. 4, the augmented Dickey-Fuller test is applied to verify whether the dataset is stationary or not. Detrending, differencing, or a combination of the two is used to complete the rationalization of our dataset, as indicated in the accompanying Figs. 5–7.

The same analysis is performed for all of the five countries under examination to get the best stationary data inputs that will help increase the accuracy of the predictions, as shown in Figs. 8 and 9. We only choose to present confirmed case modeling and forecasting results to keep our research as abbreviated as possible. We found that SES, Holt, Holt-Winters, and ARIMA perform best. Amongst many ML-based forecasting models, we found that RegressionTrees, ExtraTrees, and K-Nearest Neighbor give the best forecasts.

Below is a graphical representation of the best-fitted model for each country. Training and validation datasets observed actuals against forecasted are visualized in Figs. 10, 14, 18, 22, 26 for Australia, Canada, India, UK, and US, respectively. From these figures, it is obvious that the COVID-19 epidemic has propagated differently for each geolocation. This could be a result of differentiated weather and population demographics.

Figs. 11, 15, 19, 23, and 27 depict the five days forecasts for confirmed cases using the best performing ML/TS algorithms mentioned above for Australia, Canada, India, UK, and US, respectively. The graphs show a comparison between the actual observations of the confirmed case count in the period from 18th December to 23rd December and the different forecasts resulting from the best performing ML and statistical time series models, along with the forecasts obtained upon the implementation of ensemble and the weighted ensemble of these outperforming models. The corresponding RMSE for the best-performing forecasting models mentioned above is depicted in Figs. 12, 16, 20, 24, and 28 for Australia, Canada, India, UK, and US, respectively. In addition to the RMSE measure, the results from diagnostics utilizing the plot diagnostics technique are utilized to guarantee that none of the model assumptions are broken and that no out-of-the-ordinary behavior occurs.

This technique results in the four visual outputs depicted in Figs. 13, 17, 21, 25, 27, and 29 for Australia, Canada, India, UK, and US, respectively. The autocorrelation graph on the bottom right shows that the time series residuals are weakly correlated with their lag-adjusted counterparts. However, by proving all four of the arguments stated above, one can conclude that the residuals of this model are almost normally distributed. This signifies that we have found a well-suited model for our dataset.

Additionally, Table 1 presents an example of the resulting forecasts of each model along with the RMSE for UK. Also, Table 2 presents the forecasts obtained from the best performing ML and statistical time series forecasting models in addition to the ensemble and weighted ensemble models. We can conclude that weighted ensemble models outperform any other model using all of these illustrations.

In this work, we have highlighted the importance of time series forecasting models for highly accurate and reliable identification of the spread of infectious diseases. It has been proven that the forecasting time series models are very important to be utilized in identifying the spread of infectious diseases. The developed time-series regression modeling introduced in this article succeeded in collecting historical data rigorously and methodically to develop the most excellent model that can illustrate the underlying natural structure of the series in question. Thus, when tracking the evolution of an epidemic, it is vital to identify how many people will be impacted. Consequently, it is critical to tailor a suitable model to the time series. All suggested time-series models developed for forecasting infectious diseases prove their good performance when tested on different datasets. In future work, the presented work can be further well-developed to be adapted to the international society to forecast COVID-19 cases. Furthermore, we intend to propose a model for any similar pandemic outbreak forecasting.