Known Unknowns: Coronavirus Models

Sometimes a picture is worth a thousand question marks.

Take a glance at the graph below. The source is a recent New York Times article, which discusses the models that scientists are using as they attempt to forecast the course of the coronavirus.

Coronavirus Mortality Model - New York Times
Coronavirus Mortality Model - New York Times

If you noticed the dramatic variability in the forecast, you’re not alone.  The point isn’t lost on the article’s authors at The Times, either. To wit:

However good the modelers’ mathematical strategies may be, many of the descriptive facts about the virus are still unclear. Researchers aren’t sure about the rate at which people who become infected die, or about the rate of transmission to other people. They don’t know for sure how many people have already been infected and have some immunity to the disease — or how long that immunity will last. Even the count of coronavirus deaths itself is uncertain.

Or, perhaps even more on the nose:

Several epidemiologists said it was hard to expect the models to offer precise forecasts at this point because they rely on such uncertain inputs. “It’s like trying to repair a car while it’s still running,” said Andrew Noymer, an associate professor of public health at the University of California, Irvine

If you’re wondering what’s going on here — how such wild variability in forecasting models is possible — I direct your attention to this epidemic calculator, hosted on the code repository and development platform GitHub.
Epidemic Calculator Screenshot - Click for interactive calculator
Epidemic Calculator – Static Screenshot

First, the disclosures: I’m not a statistician or an epidemiologist (obviously) — so I can’t vouch for the precision of the differential equations or how well the outputs correlate to real-world disease outcomes.

However, assuming the calculations are a first-order approximation* of the model from The Lancet the author cites, the point is this: The calculator model exhibits extreme sensitivity to even minuscule variations in user input.

Push one input slider a bit to the right, and the curve grows dramatically; ooch another input slider a tiny bit to the left, and the curve shrinks dramatically.

Moreover, combining the categories of Transmission Dynamics and Clinical Dynamics results in eleven independent variables being modeled — with most, in various combinations, exhibiting dramatic curve-shifting properties.

What makes a visual model like this so useful — particularly to those of us who aren’t epidemiologists — is that it clearly and simply conveys an essential truth: Namely, that the scientists who are attempting to forecast the path of this virus face a truly daunting challenge.

Ash Bennington and Tony Greer discuss the epidemic calculator during an episode of Real Vision Daily Briefing. Watch the full episode here.

*NOTE: Real Vision attempted to contact the author of the Epidemic Calculator to verify the source code and underlying formulas but was unable to reach him. Real Vision was, however, able to independently confirm his status as a researcher at OpenAI, a well-known research laboratory in San Francisco, which focuses on artificial intelligence.

RELATED CATEGORIES: Coronavirus, Market Analysis, Stock Market