Updating w/new polling info only

[markjrieke]

Rerunning the model in its entirety is computationally expensive. A better method would be to save parameter outputs on day $d$, then pass them in as priors on day $d+1$.

This would require some thinking around:

• how the fundamentals-based prior gets updated (maybe it's better to use one large generative model)
• a method for recovering parameters correctly (looking at you, half-normal)
• how to account for daily tracking polls

There's a huge upside in runtime efficiency, but given that we're in the thick of the campaign right now, I shouldn't really dig into this until the next cycle.

[markjrieke]

Ah --- this also only works when the priors don't move due to external information. E.g., forecasted incumbent net approval will become more certain & will shift throughout the cycle. I wonder if there's a way to remove information from a model? (E.g., given a posterior, can we "remove" a prior?)

Edit: is this as simple as subtracting from the log prob ???