Fee reserve algorithm

[darosior]

What are the different fee reserve algorithm to study, what is the best one we should go with? What breathing room does it give in each case?

(Below is taken from #1)

X * 95Q30

Define the 95%-ile of fee-rates for a rolling window over the previous 30 days, call it 95Q30. Anything above that stands a strong chance of being accepted, unless there is high volatility in the fee-market. In which case, does it make sense to double the MA95 and use that as the fee-reserve estimate?
- A time-frame of 30 days is more characteristic of the recent fee-market than of historic high-fee periods.
- by choosing a lower bound of 95%-ile, the top 5%-ile of "extreme outliers" are excluded
- An average over 30 days would be smooth

X * MAX95Q

Consider the month(s) which historically showed the highest demand for block-space, and determine the 95%-ile fee-rates for that period. Call it MAX95Q.
- This time-frame represents the "worst" that the fee-rate market has ever been. If our algorithm would have performed well then, it will likely perform well now.
- If the recent month becomes the "worst" then the estimator adjusts to a new time-frame. If the fee-market subsequently calms down, the fee-estimator stays the same.
- by choosing a lower bound of 95%-ile, the top 5%-ile of "extreme outliers" are excluded
- The estimator is only volatile if the fee-market is growing constantly, in which case the estimatore should be adjusting.

[darosior]

We could also have the watchtower be set with an arbitrary lower bound (such as a MAX95Q but more reasonable) and use max(95Q30, lower_bound). The lower bound would of course be updated should a new high be encountered.