[Paper Review] Improving COVID-19 Forecasting using eXogenous Variables

Mohammadhossein Toutiaee, Xiaochuan Li+8 more|arXiv (Cornell University)|Jul 20, 2021

Data-Driven Disease Surveillance10 references4 citations

TL;DR

This paper proposes a hybrid forecasting approach for U.S. COVID-19 mortality using SARIMAX and MCP models enhanced with eXogenous variables—such as hospitalizations, ICU occupancy, and mobility data—demonstrating up to 64.58% improvement in national-level accuracy and 58.79% in state-level forecasts over GCN-LSTM, with significant gains over baseline models including AUG-NN.

ABSTRACT

In this work, we study the pandemic course in the United States by considering national and state levels data. We propose and compare multiple time-series prediction techniques which incorporate auxiliary variables. One type of approach is based on spatio-temporal graph neural networks which forecast the pandemic course by utilizing a hybrid deep learning architecture and human mobility data. Nodes in this graph represent the state-level deaths due to COVID-19, edges represent the human mobility trend and temporal edges correspond to node attributes across time. The second approach is based on a statistical technique for COVID-19 mortality prediction in the United States that uses the SARIMA model and eXogenous variables. We evaluate these techniques on both state and national levels COVID-19 data in the United States and claim that the SARIMA and MCP models generated forecast values by the eXogenous variables can enrich the underlying model to capture complexity in respectively national and state levels data. We demonstrate significant enhancement in the forecasting accuracy for a COVID-19 dataset, with a maximum improvement in forecasting accuracy by 64.58% and 59.18% (on average) over the GCN-LSTM model in the national level data, and 58.79% and 52.40% (on average) over the GCN-LSTM model in the state level data. Additionally, our proposed model outperforms a parallel study (AUG-NN) by 27.35% improvement of accuracy on average.

Motivation & Objective

• To enhance national and state-level COVID-19 mortality forecasting accuracy by incorporating eXogenous variables such as mobility, hospitalizations, and ICU usage.
• To evaluate the effectiveness of spatio-temporal graph neural networks (GCN-LSTM) versus statistical models (SARIMAX, MCP) in modeling pandemic dynamics.
• To determine whether eXogenous variables improve forecast performance compared to models relying solely on lagged mortality data.
• To compare the relative strengths of SARIMAX and MCP models in forecasting at national versus state levels.
• To validate that mobility and health system metrics serve as meaningful predictors for future mortality trends during the pandemic.

Proposed method

• The SARIMAX model is applied to national-level data, incorporating eXogenous variables like hospitalizedIncrease, ICU currently, and ventilator use to improve multivariate time-series forecasting.
• The MCP (Multivariate Conditional Probability) model is used at the state level with a higher lag order (k=14) and regularization to mitigate overfitting and improve generalization.
• A spatio-temporal graph neural network (GCN-LSTM) is employed, where nodes represent state-level deaths, edges encode state-to-state mobility flows, and temporal edges capture time-varying attributes.
• Elastic net regularization is used in the MCP model to balance model complexity and predictive accuracy, particularly effective in high-dimensional state-level forecasting.
• The models are evaluated using sMAPE (symmetric Mean Absolute Percentage Error) across 14 forecasting horizons for both national and aggregated state-level data.
• Mobility data from SafeGraph’s U.S. Flows is used as a key eXogenous input, representing daily and weekly population flows between states from January 19, 2020, to January 19, 2021.

Experimental results

Research questions

• RQ1 Can the inclusion of eXogenous variables such as mobility, hospitalizations, and ICU occupancy significantly improve the accuracy of COVID-19 mortality forecasts?
• RQ2 How do SARIMAX and MCP models compare to deep learning models like GCN-LSTM in forecasting national and state-level mortality?
• RQ3 Does the performance of statistical models (SARIMAX, MCP) vary between national and state-level forecasting due to data heterogeneity?
• RQ4 To what extent do mobility patterns and health system metrics serve as predictive indicators for future mortality trends?
• RQ5 Can regularization in MCP models reduce overfitting and improve performance on state-level data compared to SARIMAX?

Key findings

• The SARIMAX model achieved the best national-level forecast accuracy, with a 64.58% improvement in sMAPE over the GCN-LSTM model on average across 14 horizons.
• The MCP model outperformed all baselines at the state level, achieving a 58.79% average improvement over GCN-LSTM and a 52.40% improvement over SARIMAX.
• The inclusion of eXogenous variables such as hospitalizedCurrently and ICU occupancy significantly enhanced forecasting accuracy, particularly in the SARIMAX and MCP models.
• The SARIMAX model showed superior performance on national data, while MCP excelled at the state level due to its ability to handle higher lag orders and regularization.
• The proposed models outperformed the parallel AUG-NN model by 27.35% in average accuracy, demonstrating the value of structured eXogenous variable integration.
• The study found that adding certain eXogenous variables could degrade SARIMAX performance, suggesting that variable selection and model-specific tuning are critical for optimal results.