On Fri, Oct 25, 2013 at 5:51 PM, Jeremy Spilman <jeremy@taplink.co> wrote:
Gavin, can you confirm the best place to  read  up on the discuss fee estimation changes for v0.9?

The blog post is the best place for high-level overview.

The (closed for now, but it will come back) pull request is the best place for low-level details and nit-picking discussion:
  https://github.com/bitcoin/bitcoin/pull/3024

 
I think fee estimation at its core is about providing a data point, or even call it an API, which can be used however you see fit.

What parameters do I want to see in a 'fee estimation' API?

 - 30 minutes vs 24 hours processing time
 - Confidence Levels (50%/90%)

The pull request adds an 'estimatefees' JSON-RPC api call:

====
estimatefees [prioritymedian=0.1] [feemedian=0.5]
Estimates the priority or fee a transaction needs
to be relayed across the network and included in
the block chain.

prioritymedian and feemedian are values from 0.0
to 1.0, where 0.0 will return the smallest
recently-included-in-a-block priority (or fee) seen,
1.0 the largest, and 0.5 the median priority (or fee)
for transactions that were broadcast on the network and
included in a block.

The default value for prioritymedian (0.1) is
chosen to return a priority for free transactions that
will eventually be confirmed, but might take several hours.
The default value for feemedian (0.5) returns how much
fee you should include to have your transactions confirmed
in an average amount of time.

Values returned are:
 freepriority : priority needed to out-compete a prioritymedian
  fraction of free transactions to be relayed and included in blocks.
 feeperbyte : fee, in satoshis/byte, needed to out-compete a
  feemedian fraction of fee-paying transactions.

Values of -1.0 are returned if not enough transactions
have been seen to make a good estimate.
====

That API doesn't give "30 minute versus 24 hour" confirmation time or confidence intervals. I've always regretted not taking a statistics class; if you want to help write code that estimates confidence intervals send me an email. The API certainly isn't set in stone.

  - Is it globally consistent?

Ummm.... roughly, yes, it will be. Nodes that have just joined the network and haven't seen enough transactions enter and leave the memory pool will have a different estimate than long-running nodes, but in my testing the estimate narrows down very quickly (with three or four blocks enough fee-paying transactions have been seen to make a reasonable estimate; it takes longer to see enough free transactions to get a good estimate of the priority needed to get into the free space of a block).

RE: lots of other comments:

I feel like there is a lot of "in the weeds" discussion here about theoretical, what-if-this-and-that-happens-in-the-future scenarios.

I would just like to point out (again) that this is not intended to be The One True Solution For Transaction Fees And Transaction Prioritization. If you've got a better mechanism for estimating fees, fantastic! If it turns out estimates are often-enough wrong to be a problem and you've got a solution for that, fantastic!

RE: are we already seeing pressure on transaction fees:

I believe we are, yes. As part of the prep work for the smart fee work I spent some time plotting priority (for zero-fee transactions) and transaction fee (for zero-priority transactions) versus confirmation time, and it looks to me like people/services are starting to include more than the hard-coded fees in the reference implementation-- I assume because they want their transactions to be confirmed more quickly.

There is definitely already competition among zero-fee transactions for the "free" block space. One of the reasons I'm comfortable with the fee changes I'm proposing is if the estimation code gets it very wrong we'll see that first as free transactions taking "too long" to confirm, but they'll confirm eventually because priority increases over time.

--
--
Gavin Andresen