The short answer is that the main reason for using a decay parameter is that future minutes are uncertain, but that's already built in in Review's xMin decay. So you are discounting future EV twice.
Possibly, but I find it extremely hard to believe that the "different thing" decays at a rate of 15% per gameweek (which is what implied by the default 0.85 parameter)
Then I'm also wondering, and I suspect you would be too, how they arrived at recommending .85?
"Past studies suggest..." is how they word it, but is this simply an attempt to reconcile some distribution of data after the fact with what the model predicted?
If that's the case (big assumption on my part), I wonder if variance is being lumped in with uncertainty...which might be a sound way of looking at it? I dunno. Also seems like it could be a suboptimal conflation, perhaps?
Not wanting to account twice for the xMin decay, as you said, made a lot of sense to me. I read up a bit (just using glossary on fplreview) and was going to ask something similar to what Simon said up above (time decay is perhaps applied to other factors/stats in addition to xMin?).
But time decay itself (and how it intersects with uncertainty) is what I'm most curious about. Is it an attempt, essentially, to create a time value of information (specifically the predictive value of the information/solve)? Similar to present/future value of money in economics?
Do you mind if I ask you some questions? Or can you point me towards existing discussions where I could read up on the treatment of uncertainty in fpl solvers? I'm somewhat familiar with poker solvers but have never looked at the ones for this game
Comments
"Past studies suggest..." is how they word it, but is this simply an attempt to reconcile some distribution of data after the fact with what the model predicted?