Source-linked AI summary
Real-time forecasts and risk assessment of novel coronavirus (COVID-19) cases: A data-driven analysis
Tanujit Chakraborty, Indrajit Ghosh
TL;DR
The paper addresses short-term COVID-19 forecasting and case-fatality risk assessment under limited early outbreak data. It combines ARIMA with wavelet-based forecasting for five countries and applies an optimal regression tree to identify CFR-related variables across 50 countries. It reports ten-day forecasts and seven key parameters associated with CFR, under stated simplifying assumptions.
Problem
Limited early COVID-19 data make short-term forecasts uncertain, while reliable country-level case-fatality risk assessment requires identifying relevant demographic and disease characteristics.
Method
The paper combines ARIMA and wavelet-based forecasting for five countries and applies an optimal regression tree to COVID-19 CFR data from 50 countries.
Results
The hybrid model produced ten-day-ahead forecasts, while the regression-tree analysis identified seven variables significantly affecting CFR estimates for 50 highly affected countries.
Takeaways & Limitations
The forecasts can support early warning and health-care planning, while CFR risk assessment can provide information for planning health-care systems in affected countries.
Takeaways & Limitations
The analysis assumes comparable virus mutation rates across countries, permanent immunity after recovery, and no climate or spatial effects in short-term predictions.
Abstract
from arXiv · showhide
The coronavirus disease 2019 (COVID-19) has become a public health emergency of international concern affecting 201 countries and territories around the globe. As of April 4, 2020, it has caused a pandemic outbreak with more than 11,16,643 confirmed infections and more than 59,170 reported deaths worldwide. The main focus of this paper is two-fold: (a) generating short term (real-time) forecasts of the future COVID-19 cases for multiple countries; (b) risk assessment (in terms of case fatality rate) of the novel COVID-19 for some profoundly affected countries by finding various important demographic characteristics of the countries along with some disease characteristics. To solve the first problem, we presented a hybrid approach based on autoregressive integrated moving average model and Wavelet-based forecasting model that can generate short-term (ten days ahead) forecasts of the number of daily confirmed cases for Canada, France, India, South Korea, and the UK. The predictions of the future outbreak for different countries will be useful for the effective allocation of health care resources and will act as an early-warning system for government policymakers. In the second problem, we applied an optimal regression tree algorithm to find essential causal variables that significantly affect the case fatality rates for different countries. This data-driven analysis will necessarily provide deep insights into the study of early risk assessments for 50 immensely affected countries.
1. Introduction
The paper addresses two COVID-19 needs: short-term forecasts of daily cases in five countries and risk assessment of case fatality rates across profoundly affected countries. It links these analyses to uncertainty from limited data and the need to anticipate health-care demands.
- 11,16,643 confirmed cases and 59,170 reported deaths were recorded worldwide by April 4, 2020.
- The study asks whether daily new COVID-19 cases can be forecast for Canada, France, India, South Korea, and the UK.
- The study seeks causal variables that significantly affect case fatality rates in profoundly affected countries.
- Limited outbreak data make COVID-19 predictions widely uncertain, while rising cases burden health-care systems.
- An optimal regression tree combines case estimates, disease characteristics, and health-care facilities to assess outbreak risks for 50 countries.
2. Real-time forecasting of COVID-19 cases
The forecasting analysis uses five country-level COVID-19 time series and combines ARIMA with wavelet-based forecasting to address stationary and nonstationary patterns. The hybrid approach is designed to reduce component-model biases and produce out-of-sample forecasts.
- Data: Five univariate COVID-19 time series cover Canada, France, India, South Korea, and the UK through April 4, 2020.The datasets contain 64 observations for India, 65 for the UK, 70 for Canada, 71 for France, and 76 for South Korea.
- Hybrid forecasting model: The proposed forecasting model combines stationary ARIMA with nonstationary wavelet-based forecasting.The hybridization is intended to reduce the individual biases of the component models.
- ARIMA Model: ARIMA models linear tendencies in stationary time series using autoregressive, differencing, and moving-average orders.Model development iterates through stationarity identification, parameter estimation using ACF and PACF, and diagnostics using AIC and BIC.
- Wavelet-based Forecasting (WBF) Model: Wavelet analysis represents signals across time and frequency scales, making wavelet-based models suitable for nonstationary data.The WBF procedure uses MODWT with a Haar filter, denoises by removing high-frequency components, and applies ARIMA to the reconstructed series for out-of-sample forecasting.
- Hybrid forecasting model: ARIMA residuals for the complex COVID-19 datasets were oscillatory and periodic, motivating wavelet modeling of the remaining series.
Algorithm 1 Proposed Hybrid ARIMA-WBF Model
The proposed hybrid ARIMA-WBF model combines ARIMA forecasts with wavelet-based modeling of oscillatory residuals to forecast COVID-19 cases. It was applied to five countries and evaluated using RMSE and MAE, with results strongest for Canada, France, and the UK.
- Algorithm 1 Proposed Hybrid ARIMA-WBF Model: The model uses training data to select ARIMA(p, d, q), generate residuals, train WBF, and produce combined in-sample and out-of-sample forecasts.The algorithm selects decomposition levels using WL = int[log(n)] with periodic boundaries.
- Algorithm 1 Proposed Hybrid ARIMA-WBF Model: The two-step pipeline first models linear epidemic components with ARIMA, then remodels ARIMA residuals using a WBF model.Final predictions combine the ARIMA forecasts with WBF residual forecasts.
- Algorithm 1 Proposed Hybrid ARIMA-WBF Model: The hybrid approach targets oscillatory and nonstationary residual behavior that ARIMA fails to model adequately.Wavelet functions are used to capture residual oscillations and remaining autocorrelations.
- Experimental evaluation: The study evaluates five COVID-19 time series—Canada, France, India, South Korea, and the UK—using RMSE and MAE on training data.The datasets are described as nonlinear, nonstationary, and non-Gaussian, with limited observations.
- Results: The hybrid model outperforms individual models for Canada, France, and the UK, while producing results competitive with ARIMA for India and South Korea.The study also generates ten-days-ahead point forecasts for the five countries.
- Results: The forecasts can be updated daily or periodically as country-level actual case values become available, supporting near-term health-care resource allocation.The authors characterize the exercise as real-time rather than ex-post forecasting.
3. Risk Assessment of COVID-19 cases
Using country-level COVID-19 data for 50 affected countries, the study applies an optimal regression tree to identify variables associated with case fatality rates (CFR). The fitted tree selects seven influential variables and provides interpretable CFR relationships and rules.
- 3.1. Data: The dataset covers COVID-19 case fatality rates and candidate demographic, disease, healthcare, socioeconomic, and climate variables for 50 countries.Candidate inputs include case counts, population characteristics, age distribution, lockdown duration, outbreak duration, doctors, hospital beds, income, and climate zones.
- 3.2. Method: Regression Tree: Regression tree methodology grows, prunes, and cross-validates trees before selecting the tree with the lowest estimated prediction error.The method uses recursive partitioning with mean squared error, followed by pruning and cross-validation.
- 3.3. Results: An optimal regression tree with seven variables achieved RMSE = 0.013, R2 = 0.896, and AdjR2 = 0.769.The fitted tree used minsplit = 5 with equal costs for each variable.
- 3.3. Results: Seven variables were identified as important for CFR: case count, percentage aged over 65, total population, doctors, lockdown period, outbreak duration, and hospital beds.These were selected from 10 potential input variables and included four control variables among the seven.
- 3.3. Results: The optimal tree begins with total COVID-19 cases as the most crucial parent-node variable and represents relationships between important variables and CFR.The tree displays average CFR estimates in its boxes and provides additional decision rules.
- 3.3. Results: 10% CFR was associated with countries having more than 14,000 cases and populations between 14 and 75 million.The paper identifies this as the second-highest case fatality rate in one tree rule.
4. Limitations of our findings
The analysis relies on simplifying assumptions about comparable virus mutation rates, permanent immunity after recovery, and omitted climate-change and spatial effects in short-term predictions. It also acknowledges that additional controllable and disease-based factors may affect CFR.
- 4. Limitations of our findings: The study assumes comparable virus mutation rates across countries and permanent immunity for recovered persons.These assumptions simplify the COVID-19 dataset analysis.
- 4. Limitations of our findings: Short-term predictions omit the effects of climate change and spatial data structures.The omission applies to the short-term prediction analysis.
- 4. Limitations of our findings: Additional controllable factors and disease characteristics may affect CFR and remain future scope for the study.The paper presents the identified factors as not necessarily exhaustive.
5. Discussions
The discussion presents the hybrid forecasting model and regression-tree risk assessment as tools for ongoing COVID-19 monitoring. It emphasizes regularly updating forecasts, interpreting short-term trends, and using influential CFR-related parameters for planning.
- 5. Discussions: The hybrid ARIMA-WBF model provides ten-day-ahead forecasts for daily COVID-19 cases in Canada, France, India, South Korea, and the UK.The forecasts address limited early-growth data and can serve as an early warning system.
- 5. Discussions: Real-time forecasts can be regularly updated with new confirmed-case data, similarly to weather forecasting.The discussion distinguishes this system from an ex-post analysis.
- 5. Discussions: The forecasts mostly oscillate over the next 10 days and reflect the impact of broad social-distancing measures that likely helped stabilize the epidemic.The paper reports no stiff decay sooner and no unusual uplift in cases for the five countries.
- 5. Discussions: Short-term forecasts can guide adjustment of lockdown periods according to projected outbreak behavior.This recommendation is explicitly linked to the reported forecasts.
- 5. Discussions: Regression-tree results identify seven influential CFR-associated parameters, including four control variables relevant to government action.The discussion connects these parameters with health-care-system planning and preventive actions.
- 5. Discussions: Suggested preventive actions include social distancing, protecting people over 65, extending lockdowns during sharp increases, and increasing hospital beds.These actions correspond to variables identified as influential in the risk assessment.