There's no "getting elections right"; these models just estimate probability. They gave Trump a 29% change in 2016 -- that's almost 1 in 3!

A better way to assess accuracy would be to bin predictions (0-10% chance, 10-20% chance, etc.) and see if the observed frequency aligns with the predictions.

I like market based approaches as a signal and opportunity creation, price discovery

In parallel your signal could be a data source adding to individuals contribution to price discovery