Twitter | Search | |
Carl T. Bergstrom
So a few weeks ago alerted me to his forecasts. I finally had a few minutes to take a look tonight (apologies for the delay, Youyang!) and while I haven't had a chance to dig really deep, I'm impressed with what I'm seeing.
Reply Retweet Like More
Carl T. Bergstrom Apr 28
Replying to @CT_Bergstrom
A few preliminary impressions. First, I can be a bit of an ML skeptic. But in this case, don't let the "machine learning" text fool you into thinking this is snake oil. There is a solid mechanistic SEI(R)S foundation to everything here. ML is used for parameter inference on top.
Reply Retweet Like
Carl T. Bergstrom Apr 28
Replying to @youyanggu
What the ML is doing is inferring fixed parameters (latency, infectious period, etc) and location-specific parameters (R0, CRF, etc.). I think the ML-skeptics can could think of this as just straightforward statistical inference; though can correct me if I'm wrong.
Reply Retweet Like
Carl T. Bergstrom Apr 28
Replying to @CT_Bergstrom
(Though above my pay grade, I'd be curious how this would compare to Bayesian MCMC approaches.) Rather than employing a formal model selection procedure the system avoids overfitting using cross-validiation methods. This is critical; the approach seems sensible enough to me.
Reply Retweet Like
Carl T. Bergstrom Apr 28
Replying to @CT_Bergstrom
Unlike the IHME model, Youyang Gu's model tries to predict what happens beyond the current period of intense social distancing. This allows him to avoid the IHME's ridiculously low estimates (with tight confidence bounds) of June deaths.
Reply Retweet Like
Carl T. Bergstrom Apr 28
Replying to @CT_Bergstrom
Of course this requires some assumptions about the rate at which states will lift social distancing. These are guesswork, but they're clearly stated and better guesses than the not-at-all-during-the-duration-of-the-model that we see from IHME.
Reply Retweet Like
Carl T. Bergstrom Apr 28
Replying to @youyanggu
And 's model has the admirable quality that the cone of uncertainty expands, rather in the declining, into the future. (left:covid-projections; right: IHME)
Reply Retweet Like
Carl T. Bergstrom Apr 28
Replying to @CT_Bergstrom
Of course this is just one person's subjective opinion, but the projections pass a gut-check for me. State by state, country by country, the forecasts I've seen seem entirely plausible. That shouldn't impress anyone else, but it was crucial for me to see.
Reply Retweet Like
Carl T. Bergstrom Apr 28
Replying to @CT_Bergstrom
Another thing to recognize is that this model was presumably created with a different purpose than the IHME model. The IHME model was first and foremost intended to forecast peak hospital need. This model seems more focused on forecasting past the peak of the curve.
Reply Retweet Like
Carl T. Bergstrom Apr 28
Replying to @CT_Bergstrom
(That's not to let IHME off the hook; once they started promoting the ability for their model to do anything else, they became responsible for its predictions at other stages of the pandemic.)
Reply Retweet Like
Carl T. Bergstrom Apr 28
Replying to @CT_Bergstrom
In summary, based on a first look I like . It's a model I will start following. In the longer term, it will be necessary to adjust (learn?) the parameters around lifting social distancing. But for now, I think it's making as good of guesses as possible.
Reply Retweet Like
Carl T. Bergstrom Apr 28
Replying to @youyanggu
And in doing so, it's making predictions that seem as good as any I've seen. Of course, I'll be happy to hear others' impressions, criticisms, etc. ID modeling twitter, please let me know what you think! It's a nice David vs. Goliath story, in a way. Well done, .
Reply Retweet Like