Phase-space model to forecast COVID-19 infection peak for U.S. states | Hacker News Reader

Phase-space model to forecast COVID-19 infection peak for U.S. states

covid.arena-ai.com

37 pointsHorizonXP23 comments

"We are physicists, computer scientists and mathematicians, not epidemiologists or virologists. Therefore this should NOT be viewed as predictions on the crisis, but rather as a technical tool to test various theories and assumptions in order to make more informed decisions."

This really should be at the top in bold. This is not a prediction, do not treat it like one, this is just a thought experiment.

> The best thing you can do to fight COVID-19 is nothing. Stop writing that paper. Don't put it on the arxiv.

> In recent days we've seen an influx in papers on the arxiv modeling the spread of COVID-19. Many of these are relatively simple papers clearly written by physicists using simple SIR models, some basic curve fitting, and even Ising models to model the spread of COVID-19.

> I'm writing to ask you, from the bottom of my heart, to cut that shit out.

> This is not an unexplained X-ray line from the galactic center. This is not the 750 GeV diphoton excess. This is not something where the first paper to correctly guess the peak number of COVID-19 cases on the arxiv gets a Nobel prize. People's lives are at stake and you're not helping. At best, you make physicists look bad. Epidemiology, as a field, already exists. Any prediction from a physicist tinkering with equations pulled from Wikipedia is not going to be a better prediction than that of professional public health experts whose models are far more sophisticated and already validated.

> At worst, people die.

https://www.reddit.com/r/datascience/comments/fsfdn2/the_bes...

I would love to be able to trust what the CDC and WHO are saying on these issues, they are in charge of the official narratives on coronavirus. Unfortunately, they have performed poorly since January, forcing us all to rely on outside sources to help shape our opinions on this issue.

It is primarily their fault that models created by outsiders are getting so much attention at this time.

That's a non sequitur though. The fact that the CDC and WHO have made mistakes does not imply that the models of J. Random Asshole are worth anything.

I'd tend to agree with you that big orgs make mistakes, and of course they do, they're humans, not divine. Calling everyone else J. Random Asshole and thus inevitably crap is a step too far the other way I feel.

Hear hear. There are FAR too many "look how I modeled the data" things out there that are both super-optimistic and super-pessimistic. It's great that people are interested in modeling data, but it's far too easy for folks to think they are sound models and read more into them than they should.

Heck, just yesterday someone started replying to a FB thread of mine with a bunch of Excel screenshots saying how things will be, and when questioned about how he came to that his sole response was "it's just math".

Thanks for the feedback. 100% agree, we will make that bolder and bigger so it's super clear

It doesn't render properly on mobile btw.

Safari on iOS

Likewise Firefox on Android.

(Page jumps around when I try to scroll, graphs are off the edge of the screen with no horizontal scrollbar, and sliders don't do anything)

Thanks for flagging this. Working on a fix now!

Thank you for your responsiveness on this.

> We looked at each country’s trajectory in “phase space” – where we looked at the evolution of the daily case growth rate as a function of the number of cases (rather than as a function of time). Each country has followed a very similar trajectory in this space, where the disease begins on a theoretical SIR curve but then veers off course and decelerates after government intervention.

This has also been done here: https://aatishb.com/covidtrends/

Accompanying minutephysics video: https://www.youtube.com/watch?v=54XLXg4fYsc

Previous discussion: https://news.ycombinator.com/item?id=22715920

This page provides results by country rather than by US state.

Great references -- thanks!

No problem. Thank you for working on this.

Interestingly the I_c(t) cumulative cases is all you need to solve the SIR model. S is an exponential decay of that, R is linearly proportional, and I ends up being the difference of the population N and (S+R) based on the S+I+R = N relation. A straightforward integrating factor.

Yesterday, Governor Coumo said it almost looks like New York was close to a peak but that it was still too early to tell because there's not enough data, yet.

What concerns me is the repeated mentioning of using training data from China, when we suspect the data they've released is inaccurate.

Hey Causality, it's a valid concern. Most groups have been using China's reported deaths and cases as a baseline for the "shape" of the virus. We have modeled things with and without China / international data and are hoping to expose that soon.

Could you give a good source that discusses how the data is thought to be inaccurate and how it may be adjusted for?

Aside from any and all statistical and analytical arguments made by people much smarter than me, the two things which cast the most doubt on their numbers are a massive drop in the number of registered cell and landline phones, and the fact they actively suppressed their SARS numbers in 2003.

https://www.ft.com/content/efdec278-6d01-11ea-9bca-bf503995c...

https://www.nytimes.com/2003/04/21/world/the-sars-epidemic-e...

https://www.theepochtimes.com/the-closing-of-21-million-cell...

We've been seeing these models pop up for more than a month now, do we have any hindsight on the accuracy of earlier simulations?

They are as stretchable as elastic so accuracy is not even a goal. It's more like a live spreadsheet with nice sliders, I don't see any attempt to fidelity worth taking seriously.

Hey simias, we (Arena) would love to post retroactive accuracy for the different models so far, can work on that or post it to github

Reply on news.ycombinator.com