Base payouts on exactly the last N shares submitted, where N < the expected number of shares that will yield one block.
Yes, that means on a long round that takes for example 3N shares to reach, there's no credit at all for the first 2N shares submitted. Tough cookies. They had as much a chance as any other to be rewarded, there was no way to know in advance they wouldn't, and there's no incentive to leave after contributing them because the future is unknown and has as much positive expectation as ever. Ignore Sunk Costs.
What if a round takes less than N shares, say N/2 shares? Reach back to the previous N/2 shares those that already paid out once and pay them again. This situation too was just as unpredictable as the chance the next N shares will yield no blocks, so people can't gain any advantage by timing their entry or departure.
Already thought of this. Someone pointed out it's even less fair than the score-based approaches.
How is it unfair? Can you explain or provide a link or query that would find prior discussion?
I think it's like a score system but with a hard cliff for shares over a certain age (in share ticks), rather than slow decay.
Another benefit would be that less state is required to calculate payouts, because you only need a fixed-size rolling window of the last N shares to calculate payout on the next hit. Fewer disk space problems!
For people who believe that a long round means a pool is 'due' (a fallacy but popular), it might even encourage participation in long rounds.
The only way I can see it being considered unfair is if there's a belief that every share needs to actually earn something, no matter how far removed it is from a success, rather than just earn a equal chance at something. But if that is a strong belief, just go with Deepbit-style pay-per-share. Other hard-to-understand improvised blended models create gaming opportunities for miners or the pool operator.