Open brookslogan opened 9 months ago
wild unsubstantiated hunch -- this may be related to:
It might be due to bad Weekday model fitting, but Clarabel replaced ECOS this May, so I don't think it's responsible for these.
Clarabel is the fix for ECOS instability, and these spikes happened well before May, so it makes sense to me that ECOS is the reason for them. The PR i linked was intended to force doctor_vists
off of ECOS in order to prevent similar "spikes" that caused validation failures (drvisits was still running a pre-1.5 version of cvxpy
in late May when #1966 was created). I could be missing something though. ¯\_(ツ)/¯
Ah I misunderstood, thought that issue was about accidentally using Clarabel.
Actual Behavior:
chng
smoothed_adj_outpatient_covid
signal has several spikes/outliers:Zooming in on one of these, it appears to impact only a single day:
Expected behavior
smoothed_adj_outpatient_covid
should be smooth.Context
I was browsing around signals in EpiVis to try to sanity-check COVID-19 hospitalization forecasts.