I was showing that the correct recommended fee is not as correct as implied.
Although 90% accuracy is claimed by 21 on that correct fee/timescale, I was able to call it almost certainly wrong.
And I was correct.
Your right, fee estimate is not perfect at 21, I guess they just use past data and not take into consideration number of unconfirmed transactions in mempool at particular fee, otherwise the recommended fee would go up if almost 4000 unconfirmed transactions used such recommended fee already.
But the fee estimate algorithm become better over time, I dont doubt about that. If I really need transaction to be confirmed in the next block, I pay much higher than recommended fee. If not, I pay lower obviously. What I dont see obvious is what is so special about 1 MB, its not even defined as 2^20 bytes - weird.
21 say, "The predictions are based on blockchain data of the last 3 hours, as well as the current pool of unconfirmed transactions (mempool).
First, a likely future mempool and miner behavior is predicted using Monte Carlo simulation. From the simulations, it can be seen how fast transactions with different fees are likely to be included in the upcoming blocks."
So they are trying.
But the only way to be more "accurate", as you say is to pay more.
The only way 21 can ultimately be more accurate is to recommend higher fees.
Interesting though, according to http://www.bitcoinqueue.com/details/ the last block (418410) contained far more 50 sat/byte fees than 60 sat/byte fees.
(i have issues with the accuracy of bitcoinqueue, but it did seem to tally roughly there with 21. there were 100's of 60 sat/byte fees waiting btw)
So paying the correct recommended fee maybe too low or too high!
(i am avoiding talking of blocksize here, although it leads to this fees problem. You may like my thread, https://bitcointalksearch.org/topic/core-have-been-derelict-in-their-duties-1410211)