UMass professor develops ensemble forecast for coronavirus pandemic; projects 103,000 deaths nationwide by end of May

A man crossed the Isle of View over the Campus Pond at the UMass Amherst.Blake Nissen for the Boston Globe/File

How many people are going to die in the coronavirus pandemic?

That heartbreaking statistic is something everyone -- from average citizens worried about their loved ones to the White House -- wants to know. And the models have been spitting out varying predictions.

Now, University of Massachusetts professor Nicholas Reich is collecting the various coronavirus pandemic models and has developed a combined, or ensemble, forecast that is intended to reflect their collective wisdom.

The current ensemble forecast predicts that 103,000 Americans will have died from the coronavirus by the end of this month, including 7,433 in Massachusetts, where 4,420 deaths had been reported by Wednesday.

Over here at the COVID Forecast Hub, our team has curated forecasts of #COVID19 deaths in the US from 15 research groups. Our new national-level ensemble combines 7 models and predicts that we will see ~103K deaths by May 30 (80% PI: 93K-119K).https://t.co/jWFRZLgUBu
1/8 pic.twitter.com/KZFnZzzVgc

— Nicholas G. Reich (@reichlab) May 5, 2020

Nicholas Reich, a professor at UMass, headed a lab that is collecting coronavirus pandemic models and developing an ensemble forecast from them.UMass Amherst/UMass

The ensemble forecast, which only looks four weeks ahead, is tragic enough. But some individual models look farther into the future and see tens of thousands of more deaths through the summer.

Reich, an associate professor of biostatistics in the university’s School of Public Health, heads a lab that focuses on using statistics and data to solve public health problems.

The lab gets data from the groups of modelers who are trying to sketch out the trajectory of the pandemic and then posts them on its website, the Reich Lab COVID-19 Forecast Hub.

Reich said one goal of the project was “to create a central resource where anyone can come and look at these publicly available data and make apples-to-apples comparisons of all the models that are being made forecasting the trajectory of COVID-19.”

Another goal was to develop the ensemble forecast. Reich said his lab has been working with the US Centers for Disease Control and Prevention for three to four years on building real-time influenza forecasting models, including building ensemble models, and has brought that experience to forecasting the progress of the coronavirus pandemic for the CDC.

To better coordinate #COVID19 forecasting efforts we have:

(i) built a centralized data repo that standardizes forecast data from many models into a common format,https://t.co/SEN07vHb28
(ii) created an ensemble model,
(iii) built an interactive vizhttps://t.co/jWFRZLgUBu

3/ pic.twitter.com/IZ9mW3odQJ

— Nicholas G. Reich (@reichlab) April 17, 2020

“We’re not starting from scratch,” he said. “We’ve done this before -- under less pressure -- but we’re in a good position to do this now for COVID-19.”

The ensemble brings together a “broad array" of models, said Reich, whose lab is funded by the CDC.

“You want models that are bringing different perspectives to the table,” he said. “You don’t want five models that were all created by one team."

After the data is assembled, it’s run through an algorithm that is “fairly simple," a matter of essentially “taking the average across all these different numbers," he said.

Future refinements, he said, may include weighting the models by how they’ve performed in the past.

He said the lab collects 20 models from 15 different research groups, though not all the models are included in the ensemble.

Reich said that the ensemble forecast only goes four weeks out because he and his colleagues feel like there’s a “limit of predictability” of about four weeks.

Reich’s lab collaborates with scientists at the CDC. The lab also feeds the data it’s collected and its ensemble forecast to the agency, which posts the data on its own website. The data is also used to feed the fivethirtyeight.com website.

Models have sparked controversy as the pandemic crisis has unfolded, with outcries raised when their estimates have been too high or too low. Model predictions have also played a big role in the unfolding debate over whether it’s safe to relax social distancing measures and gradually reopen the economy.

Reich declined to single out a favorite individual model, though he gave a shout-out to the work of the Los Alamos National Lab, which, he said, has previously consistently done a good job modeling the spread of seasonal influenza.

In general, he said, “I am glad that models are part of this national conversation about how we can understand when the country is ready to reopen. I think they can provide really useful data-driven inputs into that decision.”

While the ensemble forecast predicts 103,000 deaths by the end of the month, Reich said that the “best-guess” ranges for the array of models his lab uses range from 93,000 to 111,000 nationwide by the end of the month.

He said that the range “just speaks to how complicated a process this is to predict,” with human decisions having an impact on the spread of the deadly disease.

“We are not predicting a system that is a sort of immutable physical system like the weather. We are predicting a system intimately tied with human behavior and political decision-making, and there’s just so much uncertainty about how all that will play out," he said.