# Time Series Analysis and Forecasting of COVID-19 Cases Using LSTM and ARIMA Models

**Authors:** Arko Barman  
**arXiv:** [2006.13852](https://arxiv.org/abs/2006.13852)  
**Date:** 2020-06-05  
**Summary page:** https://www.summarizepaper.com/en/paper/2006.13852v1/

## Key points

- The study focuses on Time Series Analysis and Forecasting of COVID-19 Cases using LSTM and ARIMA models.
- Accurate prediction of country-wise COVID-19 cases is crucial for aiding policymakers and healthcare providers in preparing for the future.
- Performance evaluation of LSTM and ARIMA models was conducted to predict confirmed COVID-19 cases.
- Daily cumulative case data was used to generate 1-day, 3-day, and 5-day forecasts with various LSTM models and ARIMA.
- Two innovative k-period performance metrics - kMAPE and kMdSA - were introduced to evaluate accuracy over multiple days.
- Results showed low prediction errors for both LSTM models and ARIMA, with slight underestimation by LSTMs and slight overestimation by ARIMA in their forecasts.
- While ARIMA required longer sequences for accurate predictions, LSTMs could perform well even with smaller sequence sizes as small as 3 but needed a larger number of training samples for optimal performance.
- The development of k-period performance metrics proposed in this study is expected to be beneficial for evaluating time series models' performance accurately over multiple periods.
- Comparison between LSTM models and ARIMA revealed their value as tools for time series analysis and forecasting of COVID-19 cases.
- The detailed analysis presented provides valuable insights into the capabilities of these models in predicting case numbers accurately over short-term and long-term periods.

## Summary

The study on Time Series Analysis and Forecasting of COVID-19 Cases using LSTM and ARIMA models explores the critical need for accurate prediction of country-wise COVID-19 cases. This is crucial in aiding policymakers and healthcare providers in preparing for the future. The research evaluates the performance of Long Short-Term Memory (LSTM) models and Auto-Regressive Integrated Moving Average (ARIMA) model to provide insights into their effectiveness in predicting confirmed COVID-19 cases. Daily cumulative case data was used to generate 1-day, 3-day, and 5-day forecasts with various LSTM models and ARIMA. Two innovative k-period performance metrics - k-day Mean Absolute Percentage Error (kMAPE) and k-day Median Symmetric Accuracy (kMdSA) - were introduced to evaluate accuracy over multiple days. Results showed low prediction errors for both LSTM models and ARIMA, with slight underestimation by LSTMs and slight overestimation by ARIMA in their forecasts. It was observed that while ARIMA required longer sequences for accurate predictions, LSTMs could perform well even with smaller sequence sizes as small as 3. However, LSTMs necessitated a larger number of training samples for optimal performance. The development of k-period performance metrics proposed in this study is expected to be beneficial for evaluating time series models' performance accurately over multiple periods. Comparison between LSTM models and ARIMA revealed their value as tools for time series analysis and forecasting of COVID-19 cases. The detailed analysis presented provides valuable insights into the capabilities of these models in predicting case numbers accurately over short-term and long-term periods. Overall, this study significantly contributes to enhancing our understanding of effective forecasting methods for managing public health crises such as the ongoing COVID-19 pandemic.