I'm reminded of Tetlock and Gardner's excellent book 'Superforecasting', which was essentially a study of people who consistently score at the top of prediction markets. One key thing that these 'superforecasters' had in common was that any new information caused them to update their model of the world, but none caused them to update it very much - typical people making predictions either didn't update their model or…
I guess we just need to tune our relaxation factors [0] or, perhaps better, recalibrate our Kalman filters.
[0] https://en.wikipedia.org/wiki/Successive_over-relaxation