Ground Rules
To answer this question we first need to lay down some ground rules of what we’re trying to solve. We’ll focus on trying to solve the problem for consumer wallets only. We’ll be ignoring microchannels, which dramatically reduce the number of transactions used but still have to put some on the blockchain. We’ll also be assuming that full replace by fee is in effect (see
https://medium.com/@bramcohen/the-inevitable-demise-of-unconfirmed-bitcoin-transactions-8b5f66a44a35 ) because the best solution uses that fairly aggressively.
What should transaction fees be?
Before figuring out how wallets should calculate transaction fees, we first need to know what transaction fees should be. The obvious solution to that question is straightforward: It should be determined by supply and demand. The price is set at the point where the supply and demand curves meet. But supply and demand curves, while mostly accurate, are a little too simple of a model to use, because they don’t take into account time. In the real world, the supply of space for transactions is extremely noisy, because more becomes available (and has to be immediately consumed or it’s lost forever) every time a block is minted, and block minting is an intentionally random process, that randomness being essential for consensus. Demand is random and cyclical. Random because each transaction is generated individually so the total amount is noisy (although that averages out to be somewhat smooth at scale) and has both daily and weekly cycles, with more transactions done during the day than at night.
What all these result in is that there should be a reward for patience. If you want or need to get your transaction in quicker you should have to pay on average a higher fee, and if you’re willing to wait longer it should on average cost less. Inevitably this will result in transactions taking on average longer than one block to go through, but it doesn’t require it of everyone. Those who wish to offer high fees to be sure of getting into the very next block are free to do so, but if everyone were to do that the system would fall apart.
What should the wallet user interface be?
Ideally transaction fees would be handled in a way which didn’t require changes to a wallet’s user interface at all. Unfortunately that isn’t possible. At a minimum it’s necessary to have a maximum fee which the user is willing to spend in order to make a transaction go through, which of course means that some transactions will fail because they aren’t willing to pay enough, which is the whole point of having transaction fees in the first place.
Because transaction fees should be lower for people willing to wait longer, there should be some kind of patience parameter as well. The simplest form of this is an amount of time which the wallet will spend trying to make the transaction go through before giving up (Technically it may make sense to specify block height instead of wall clock time, but that’s close enough to not change anything meaningful). This results in fairly understandable concepts of a transaction being ‘pending’ and ‘failed’ which happen at predictable times.
Transactions eventually getting into a ‘failed’ state instead of going into permanent limbo is an important part of the wallet fee user experience. Unfortunately right now the only way to make sure that a transaction is permanently failed is to spend its input on something else, but that requires spending a transaction fee on the canceling transaction, which of course would be just as big as the fee you weren’t willing to spend to make the real transaction go through in the first place.
What’s needed is a protocol extension so a transaction can make it impossible for it to be committed once a certain block height has been reached. The current lack of such an extension is somewhat intentional because there are significant potential problems with transactions going bad because a block reorganization happened and some previously accepted transactions can’t ever be recommitted because their max block height got surpassed. To combat this, when a transaction with a max block height gets committed near its cutoff it’s necessary to wait a longer than usual number of blocks to be sure that it’s safe (I’m intentionally not giving specific numbers here, some developers have suggested extremely conservative values). This waiting is annoying but should only apply in the edge case of failed transactions and is straightforward to implement. The really big problem is that given the way Bitcoin works today it’s very hard to add this sort of extension. If any backwards-incompatible change to Bitcoin is done, it would be a very good idea to use that opportunity to improve Bitcoin’s extension mechanisms in general and this one in particular.
What information to use
The most obvious piece of information to use for setting transaction fees is past transaction fees from the last few blocks. This has a number of problems. If the fee rate goes high, it can get stuck there and take a while to come down, if ever, even though the equilibrium price should be lower. A telltale sign of this is high fee blocks which aren’t full, but it’s trivial for miners to get around that by padding their blocks with self-paying transactions. To some extent this sort of monopoly pricing is inherent, but normally it would require a cabal of most miners to pull it off, because any one miner can make more money in the short term by accepting every transaction they can instead of restricting the supply of available transaction space. If transaction fees are sticky, a large but still minority miner can make money for themselves even in the short term by artificially pumping fees in one of their blocks because fees will probably still be high by the time of their next block.
Past fees also create problems for SPV clients, who have to trust the full nodes they connect to to report past fees accurately. That could be mitigated by making an extension to the block format to, for example, report what the minimum fee per bytes paid in this block is in the headers. It isn’t clear exactly what that extension should do though. Maybe you want to know the minimum, or the median, or the 25th percentile, or all of the above. It’s also possible for miners to game the system by making a bunch of full nodes which only report blocks which are a few back when fees have recently dropped. There are already some incentives to do that sort of bad behavior, and it can be mitigated by having SPV clients connect to more full nodes than they currently do and always go with the max work, but SPV clients don’t currently do that properly, and it’s unfortunate to create more incentives for bad behavior.
Another potential source of information for transaction fees is currently pending transactions in the network. This has a whole lot of problems. It’s extremely noisy, much more so than regular transaction fees, because (a) sometimes a backlog of transactions builds up if no blocks happen to have happened in a while (b) sometimes there aren’t many transactions if a bunch of blocks went through quickly, and (c) in the future full nodes can and should have a policy of only forwarding transactions which are likely to get accepted sometime soon given the other transactions in their pools. Mempool is also trivially gameable, in exactly the same way as the last few blocks are gameable, but worse: A miner who wishes to increase fees can run a whole lot of full nodes and report much higher fees than are really happening. Unlike with fee reporting in blocks, there’s no way for SPV clients to audit this properly, even with a protocol extension, and it’s possible for full nodes to lie in a much more precise and targetted manner. Creating such a strong incentive for such a trivial and potentially lucrative attack seems like a very bad idea.
A wallet’s best information to use when setting price are the things which can be absolutely verified locally: The amount it’s hand to pay in the past, the current time, how much it’s willing to pay by when. All of these have unambiguous meanings, precise mathematical values, and no way for anybody else to game them. A wallet can start at a minimum value, and every time a new block is minted which doesn’t accept its transaction increase its fee a little, until finally reaching its maximum value at the very end. Full nodes can then follow the behavior of storing and forwarding along several blocks’s worth of transactions, ten times sounds reasonable, ignoring transactions which pay less per byte than the ones they have stored, and further requiring that a new block be minted between times when a single transaction gets replaced by fee. That policy both has the property of being extremely denial-of-service resistant and minimizing the damage to zeroconf. (Zeroconf is a bad idea, but if something is a good idea to do for other reasons reducing the pain to those stuck with zeroconf is a nice bonus.)
An actual formula
At long last, here is the formula I advocate using:
Pick a starting point which is de minimis for your first transaction or 1/2 (or less, configurable) your last fee paid if you’ve sent coin before
Let B = max number of blocks from start before giving up, S = starting fee, M = max fee
For each new block at height H from the start, post a new transaction with fee e^(lg(S) + (lg(M) — lg(S)) * H/B)
To avoid artifacts when multiple wallets use the same magic numbers, do this before the first block: pick V uniformly in [0, 1], let S = e^(lg(S) + (lg(M) — lg(S)) * (V/(V+B)))
The very first time you send coin it makes sense to give it a longer time to do the transaction because it’s starting from a very low value and you don’t want to way overshoot the amount necessary. But if you start from the standard absolute minimum fee in Bitcoin and put the maximum time at several hours it will increase by less than 10% per block, so exponential growth is on your side.
It might be reasonable to, for example, start at a value which is a discount to the minimum paid in the last block if that value is less than what you would start with otherwise and if there’s a protocol extension to put that information in the block headers. Such possibilities should be studied and discussed more, but the formula I gave above should be the default starting point if you simply want something which works and is conservative and reliable.
Sidebar: Handling utxo combining
Whenever a wallet makes a payment, it needs to decide how to structure the inputs and outputs of the new transaction. Generally the output consists of two utxos, one of them going to the recipient and one of them going back into the original wallet. Which input or inputs to use is less clear. Usually an attempt is made to optimize for anonymity, or at least leaking as little information as possible, and there’s usually a comment in the code saying what amounts to ‘I can’t clearly justify any particular strategy here but this is what I’m doing’.
When there are real transaction fees, one might consider trying to optimize utxo combining for fees. The strategy used turns out to matter surprisingly little for fees in the long run. For every separate utxo in your wallet, you’ll eventually have to pay the fee to combine it with something else, and the amount of increase in fee will be the same regardless of whether you do it in the current transaction or a later transaction. It does make sense to include more inputs in earlier versions of a payment though, because the fees at that time are lower, and drop them in later versions once the fees have gone up, in the hopes that the utxo consolidation can be done for cheaper in some later transaction. It may also make sense to do completely separate purely consolidation transactions with no external output during off-peak times. That puts more bytes on the blockchain because of the unnecessary intermediary value it generates though, so there needs to be a significant difference in fees between peak and off-peak times for it to make sense. Both of those techniques have significant and unclear privacy implications and should be studied more.
There are also signing tricks which could potentially save significant amounts of bytes on the blockchain, thus lowering fees. The most elegant would be to create a new extension so that when there are multiple inputs to a transaction which all use Schnorr the signature can be a single combination signature instead of separate signatures for each of them. This has very little downside and I’m strongly in favor of it being done.
A simpler, less elegant trick which saves more bytes would be to allow multiple inputs to the same transaction which use the same key to only result in a single signature. This lowers privacy, because it gives away the association between utxos before they’re consolidated, but if used properly would only push back that reveal a little bit. The danger is that wallets would instead use it improperly and use the same key all the time, which would always save as many bytes as possible but be a privacy disaster.
A trick which is a just plain bad idea, although it would save even more bytes, would be not count the bytes of the reveal of a p2sh script to count if that exact same script has ever been used before. This is clearly a bad idea, because it directly encourages extremely privacy-averse behavior, and because it necessitates a data structure of all p2sh scripts which have ever been done before for validation, which is quite large and costly to maintain.