https://www.medrxiv.org/content/10.1101/2020.03.11.20024901v2.full.pdf
Batista was one of the first on the ball of data modeling through logistic regression. You should be aware that if those charts look "nice" it is because they are constantly updated and the prediction is readjusted. It's a bit different if you stick to the data that was available 15 days ago and check what it predicted for today.
The code is available here btw
https://www.mathworks.com/matlabcentral/fileexchange/74411-fitvirus
So, those type of predictions, as presented, always look very convincing. Read the fine print and caveats in Batista's paper.
Likewise, the
IHME model (the one that predicted "only" 61000 deaths for the US at one point and was very "hot" a few weeks ago, much less so today, looked very convincing visually. The issues with that model are different (at least as you can guess from the outside). Some conversation I had around a month ago, slightly edited for privacy, italics are my words...
29 MAR 2020, 23:59
something that doesn't "feel" right about the IHME model is that the predicted deaths peaks (and the whole death line) don't seem to trail peak infection rates by much, if at all.
Exactly
Look at some of the states
Deaths go to zero in June
I think Birx hates models and picked the low one
Look how the deaths just stop
There is no tail in some of the states
The overall is going since some states start very late
Reverts to baseline. Approved for one month only. XXXXXXXX will license it to you for a small fee
But I don’t get how individual stares have zero deaths past May
Yes homeostatic
I knew the word!
raw data is a bit all over the place right now, was trying (roughly) to normalize confirmed and death, then find best correlated shifts on normalized curve yesterday, but data on smaller euro countries isn't reliable (lulls then sudden bulk reporting), Germany is a clear outlier, etc...
So, bottom line, those observational predictive models with nice visualizations but very little connexion with the actual underlying process will always appear to be convincing. Batista is very well aware of that and notes the limitations in his paper, but yeah, (almost) no one reads the paper these days...