Long term pollster error across cycles

[markjrieke]

We have information about pollster bias across many cycles. This can be used to inform poll level priors by running a model that fixes the random walk to the observed outcome on day $d=D$. Pollster, mode, etc. bias can then be given as a random walk over the observed cycles.

See, for example, the last paragraph in section 3.2 in Linzer, 2013

[markjrieke]

This methodology can also be used to set better priors whole-cloth. If parameters change from cycle to cycle (e.g., avg polling error, scale of sampling methodology, etc.), then I should think of them as a random walk over each cycle & use the output to set 2028 priors