I appreciate this blog post, how you explain that R^2 near 1 can be easily achieved without indicating that the model is right. Since we assume there is a growth, wouldn’t it make sense to scale down each error by the corresponding value of the modeled growth function before squaring it in the loss function? Observing that quadratic growth fits better than exponential for a while during an outbreak does not have to be so wrong, if most of the current spreading is local and spatial, like a circle with radius growing linearly with time. Right?