Utility of human judgment ensembles during times of pandemic uncertainty: A case study during the COVID-19 Omicron BA.1 wave in the USA

Srinivasan Venkatramanan, Juan Cambeiro, Tom Liptay, Bryan Lewis, Mark Orr, Gaia Dempsey, Alex Telionis, Justin Crow, Chris Barrett, Madhav MarathemedRxiv 2022 · DOI

Abstract

Responding to a rapidly evolving pandemic like COVID-19 is challenging, and involves anticipating novel variants, vaccine uptake, and behavioral adaptations. Human judgment systems can complement computational models by providing valuable real-time forecasts. We report findings from a study conducted on Metaculus, a community forecasting platform, in partnership with the Virginia Department of Health, involving six rounds of forecasting during the Omicron BA.1 wave in the United States from November 2021 to March 2022. We received 8355 probabilistic predictions from 129 unique users across 60 questions pertaining to cases, hospitalizations, vaccine uptake, and peak/trough activity. We observed that the case forecasts performed on par with national multi-model ensembles and the vaccine uptake forecasts were more robust and accurate compared to baseline models. We also identified qualitative shifts in Omicron BA.1 wave prognosis during the surge phase, demonstrating rapid adaptation of such systems. Finally, we found that community estimates of variant characteristics such as growth rate and timing of dominance were in line with the scientific consensus. The observed accuracy, timeliness, and scope of such systems demonstrates the value of incorporating them into pandemic policymaking workflows.

Introduction

Epidemiological forecasting typically relies on computational models, whether mechanistic or machine learning-based. While powerful, these models often struggle during "regime shifts"—sudden changes in pandemic dynamics, such as the emergence of a new variant like Omicron. Computational models require time to be re-parameterized with new data and often fail to account for rapid shifts in human behavior or policy. This study investigates whether human judgment, aggregated through structured forecasting platforms, can fill this gap by synthesizing diverse information streams (news, preprints, social media) and adjusting mental models faster than code-based simulations.

Methods and Data

The researchers partnered with the Virginia Department of Health to host the "Real-time Pandemic Decision Making" (RPDM) tournament on Metaculus. Over 18 weeks (November 2021 – March 2022), 129 unique users provided 8,355 probabilistic predictions across 60 questions. The study focused on the "Metaculus Prediction," a weighted ensemble based on historical accuracy and recency.

Forecasts were evaluated using metrics such as Median Absolute Percentage Error (medMAPE) for cases and hospitalizations, and Mean Absolute Error (medMAE) for vaccine uptake. These were compared against ground truth data from Virginia health authorities and benchmarks including the CDC ForecastHub (an ensemble of ~40 computational models) and simple linear baselines.

Key Results

The human ensemble demonstrated remarkable accuracy and adaptability. Case forecasts performed on par with the CDC ForecastHub, even slightly outperforming it at the 1-week horizon. Vaccination forecasts were significantly more robust than linear baselines, typically staying within 1 percentage point of the ground truth with three weeks of lead time.

Percentage error heatmaps for case, hospitalization, and vaccination forecasts.

A key advantage of the human ensemble was its rapid integration of "preprint science." Forecasters adjusted estimates of Omicron’s severity and growth rate in late December 2021 immediately following the release of key preprints, pivoting faster than many formal nowcasting systems. Furthermore, the community correctly identified noise in official CDC nowcasts for Omicron prevalence, acting as a vital sanity check.

Beyond quantitative metrics, the tournament captured qualitative shifts in the wave's prognosis. Forecasters accurately identified the timing of peaks and subsequent declines, often with minimal error in days.

Error in days for timing forecasts of peak and trough activity.

Discussion and Future Work

While successful, the study highlights challenges such as the high overhead of designing precise forecasting questions. Future work should explore "Chimeric" models that integrate human-judged parameters directly into computational simulations. Additionally, conditional "if-then" forecasting could provide policymakers with better tools for exploring counterfactual scenarios. Overall, the study demonstrates that human judgment ensembles are a valuable, timely, and accurate complement to traditional epidemiological modeling, particularly during periods of high uncertainty.

Summary updated

Abstract

Abstract Responding to a rapidly evolving pandemic like COVID-19 is challenging, and involves anticipating novel variants, vaccine uptake, and behavioral adaptations. Human judgment systems can complement computational models by providing valuable real-time forecasts. We report findings from a study conducted on Metaculus, a community forecasting platform, in partnership with the Virginia Department of Health, involving six rounds of forecasting during the Omicron BA.1 wave in the United States from November 2021 to March 2022. We received 8355 probabilistic predictions from 129 unique users across 60 questions pertaining to cases, hospitalizations, vaccine uptake, and peak/trough activity. We observed that the case forecasts performed on par with national multi-model ensembles and the vaccine uptake forecasts were more robust and accurate compared to baseline models. We also identified qualitative shifts in Omicron BA.1 wave prognosis during the surge phase, demonstrating rapid adaptation of such systems. Finally, we found that community estimates of variant characteristics such as growth rate and timing of dominance were in line with the scientific consensus. The observed accuracy, timeliness, and scope of such systems demonstrates the value of incorporating them into pandemic policymaking workflows.
BibTeX
@misc{venkatramanan2022utility,
  author = {Srinivasan Venkatramanan and Juan Cambeiro and Tom Liptay and Bryan Lewis and Mark Orr and Gaia Dempsey and Alex Telionis and Justin Crow and Chris Barrett and Madhav Marathe},
  title = {{Utility of human judgment ensembles during times of pandemic uncertainty: A case study during the COVID-19 Omicron BA.1 wave in the USA}},
  year = {2022},
  doi = {10.1101/2022.10.12.22280997},
  howpublished = {medRxiv},
}

Loading related content...

Figure