Quote:
Originally Posted by BaseMetal2
All data is welcome but 5000 games will be about the same level predictor as a batch of 500 STT games, i.e. not too accurate.
The 30% + roi's are also more than decent, 5 to 10k worth of them gets you well onto the first page of tracking sites and most of these top players are having a good run - it's very easy to over estimate the achievable roi's.
I'll fill in a little on this because it now seems to me I was a bit vague with accuracy, I'll give a concrete example.
In the Normal distribution (I also typo'd and used 'binomial' in an earlier post) it is well known that approx 68% results will fall within +/- one sd of the mean.
I'll calculate some values for a mediocre player 68% confidence range that has an roi of 0.0% in both STT's and 180 games.
So for games of BI $0.95 + $0.05 :
the variance for 1 game of 9s, at 0% roi = 2.25
the variance for 1 game of 180s, at 0% roi = 25.6
the variance for 1k games of 9s, at 0% roi = 1000 * 2.25 = 2250
the variance for 1k games of 180s, at 0% roi = 1000 * 25.6 = 25600
the stdev for 1k games of 9s, at 0% roi = Sqrt(2250) = 47.4
the stdev for 1k games of 180s, at 0% roi = Sqrt(25600) = 160
So after playing 1k games the 9s player is 0.0 +/- 47.6 BI's (i.e. +/- 1sd) whereas the 180s player is 0.0 +/- 160 BI's, both sets cost 1000 BI's = $1000.
and in roi% terms
the 9s player is +/- 47.6 * 100 / 1000 = +/- 4.7%
the 180s player is +/- 160 * 100 / 1000 = +/- 16%
So after 1k games we could be 68% certain the 9s players' roi is found in the range 0.0 +/- 4.7%, but we only 'know' the 180s player roi to +/- 16%.
How many games before we know the 180 roi at 4.7% accuracy?
Sqrt(n *25.6 ) * 100/n = 4.7
or n = 11589
With a 0.0 roi playing 180s after 11.6k you will know your roi as 0.0% +/- 4.7% with a 68% confidence, or 11.5 times as many games as for the STT case.
So to know roughly, using the Normal model, how many more games you need to predict roi to about the same percentage in either STTs or 180s it is reasonable to use a factor of 10x when comparing them.
Things to note are the Normal model is a good model of tourny results, but games change and tilt happens etc. and I might have screwed all this up (a good chance of this
) I was surprised how large a factor it seems.