Very excited about this brand new preprint arxiv.org/abs/2007.05035 with @jonassjuul et al. We argue that most current epidemiological predictions do not appropriately capture the uncertainty related to their predictions. Specifically, they tend to underestimate peaks!(thread)
Essentially the problem is that - even though the models produce ensembles of curves - each time step is treated separately in the typical state-of-the-art statistical analysis.
It's easy, for example, to construct examples where 100% of peaks lie outside the confidence intervals calculated in this way (25-75% shown in gray below). 
In an example of a real-world simulations corresponding to three (best, medium, and worst) case social-distancing scenarios, we found that 38%, 54% and 67% (!) of trajectory peaks were located beyond the 75th percentile.
As a possible solution, we suggest that modelers switch to *curve-based descriptive statistics* to summarize curve ensembles. These do a much better job of capturing our intuitive notion of what the confidence intervals should imply.
A Python package (curvestat) to produce the curve-based descriptive statistics used in this manuscript can be found at github.com/jonassjuul/cur…
