Cross-Validation Comparison of COVID-19 Forecast Models

Area of Science:

• Epidemiology
• Data Science
• Time Series Analysis

Background:

• Forecasting models for COVID-19 have shown variable accuracy.
• Data quality for COVID-19 is a significant concern impacting model reliability.

Purpose of the Study:

• To compare univariate time series models for COVID-19 case forecasting.
• To identify the optimal forecasting model using cross-validation and varying forecast periods.
• To address data quality issues in COVID-19 datasets.

Main Methods:

• Utilized COVID-19 case data from "Our World in Data" (December 31, 2019 – November 21, 2020).
• Compared univariate models: Error Trend Season (ETS), Exponential Smoothing (multiplicative), and ARIMA.
• Employed cross-validation and Mean Absolute Percentage Error (MAPE) for model evaluation.

Main Results:

• The ETS model with additive error-trend and no seasonality demonstrated superior performance.
• A minimum of 100 days of data was required for the ETS model to achieve a 5% MAPE threshold.
• Identified ETS as the best-fit univariate model for the analyzed COVID-19 data.

Conclusions:

• The ETS model offers a reliable approach for COVID-19 case forecasting.
• Sufficient historical data (at least 100 days) is crucial for accurate forecasting.
• This study provides a benchmark for selecting time series models in epidemiological forecasting.