# Combination forecasting of COVID-19 in Bangladesh: guiding public health policy through integrated time series, ML, and DL models

## TL;DR

This paper proposes a weighted ensemble model combining time series, ML, and DL to improve forecasting of daily new COVID-19 cases in Bangladesh.

## Contribution

A novel weighted ensemble model integrating SARIMA, XGBoost, and RNN for context-specific forecasting in Bangladesh.

## Key findings

- The weighted ensemble model outperformed individual models in forecasting accuracy.
- The model integrates linear seasonality and nonlinear dynamics for better prediction.
- Results suggest the model can guide efficient public health interventions in Bangladesh.

## Abstract

Accurate prediction of trends in infectious diseases is crucial for early public health actions, which are difficult to achieve in low-resource countries such as Bangladesh, where the healthcare system is significantly challenged. Although many researchers worldwide have utilized machine learning (ML) and deep learning (DL) models for making COVID-19 predictions, limited research has focused on context-specific ensemble forecasting approaches for Bangladesh, creating a gap in local epidemic modeling. This study aims to design a Weighted Ensemble forecasting model that combines the best time series models (SARIMA, ETS), ML (XGBoost), and DL (LSTM, RNN, GRU) models to improve the accuracy of daily new COVID-19 cases in Bangladesh.

We developed a context-specific Weighted Ensemble model integrating Seasonal Autoregressive Integrated Moving Average, XGBoost, and Recurrent Neural Network, with weights (0.1531, 0.4319, 0.4150). Log transformation handled zeros, with scaling for neural networks. Forecasts extended to May 2027.

Out-of-sample results (SARIMA: RMSE = 14.56, MAE = 9.02, MAPE = 216.87%; XGBoost: RMSE = 5.16, MAE = 3.45, MAPE = 134.52%; Recurrent Neural Network: RMSE = 5.37, MAE = 2.71, MAPE = 41.11%; Weighted Ensemble: RMSE = 5.87, MAE = 3.49, MAPE = 111.02%) show the Weighted Ensemble outperforms individual models, including the best time series, machine learning, and deep learning models, despite misspecification in traditional models.

The Combination model based on Seasonal Autoregressive Integrated Moving Average, XGBoost, and Recurrent Neural Network can improve prediction performance for Bangladesh’s complex COVID-19 data better than the XGBoost model does. The power of ensemble modeling in capturing linear seasonality and the nonlinear dynamics it is evident from these results. This novel integration can inform public health policy, allowing health authorities to fine-tune interventions and make more efficient use of resources while avoiding unnecessary lockdowns as COVID-19 becomes an endemic disease.