STATS CONFIDENCE FACTOR/NUMBER

[0909dylan]

This is a plea from a NON-MATHS PERSON for a SIMPLE METHOD to give some sort of CONFIDENCE FACTOR/NUMBER to any set of stats.The idea is to give some sort of way of COMPARISON between various sets of stats.There are a number of such formula around that use WINS+ STRIKE RATE + TOTAL STATS /etc.............to give a NUMBER and a +/- CONFIDENCE NUMBER.......but which is RELIABLE + SIMPLE to use ???????????..........e,g you might have a stat of..............3/7.......+ 42%..................or.............23/345...........+ 16%...............or.............45/234...............+ 9%.........or............230/1356...............+ 23%................How can I RANK these 4 sets of stats in terms of RELIABILITY + ACCURACY ???...........and can I give each set a NUMBER to RANK its reliability in comparison with any other set of stats.

[pippex]

well explained Dylan

I have some stats books here but I am slow to open them at my age

[Pittsburgh Phil]

You could try the archie test as demonstrated by Steve Tilley in an old Smartsig edition. This is basically a chi test which gives a figure of how likely past results are due to luck, or if the results are sound.

I think vcamrn posted up a formula for testing the worthiness of a system. This was a formula based on share performance and took into consideration inflation / interest rates? I'll see if I can find it.

[Pittsburgh Phil]

Here is vcamrn's posting:

viewtopic.php?f=12&t=7196

[Pittsburgh Phil]

I found a copy of the Archie test from the 0dd5.com site. The site has now expired but a copy exists in the internet archive:

http://web.archive.org/web/200410092028 ... chie.shtml

[0909dylan]

Thanks for that........I remember the VCRAM POST ( I did not really understand it then )................Is there a " Stats Made Simple " version anywhere ?????????

[Pittsburgh Phil]

You can setup the vcamrn formula in a spreadsheet - just enter the wins, runs, roi etc. and see the resulting figure. If I get time I'll knock one up.

Obviously that formula needs one BIG change to it:

Bank of England Rate = 4.75%

[0909dylan]

PHIL..........Any help appreciated....... ............We all know if you have a set of stats.............30/100.........+ 35%...........its loads better than.......3/10............+ 35%.............and much worse than............300/1000.........+ 35%.............BUT IT WOULD BE NICE TO GIVE A RANKING/NUMBER to all 3 sets of stats..............and if you are comparing with other stats........e.g.....................45/356..........+ 27%.....................or...........3/7.............+ 67%...............give these NUMBERS also so you can compare all 5 sets of stats.

[pippex]

perfect Dylan..thats what I want and one can be implemented easily on the grid data

[Pittsburgh Phil]

I don't know what you mean by adding a ranking to the stats Do you mean sort them by total runs, so that those with the biggest sample are at the top?

---

Low Sample Sizes
Over at paceadvantage.com poster GameTheory suggests using a version of Laplace Smoothing. This is where you add 1 to the wins, and say 9 to the runs in order to make a a stat not so erratic.

Something that happens a lot with a new sire / trainer / jockey is that they may have a flying start and produce unusually high strike rates - strike rates which can not be maintained. If you apply a smoothing at the start then this should dampen spikes for new stats.

e.g. New Sire starts having runners on the sand.

In his first 10 runs he has 6 winners = 60%. There is no way that this will be maintained and the 60% strike rate is sure to drop as more of his offspring run.

Applying Laplace Smoothing at this 10 runner stage:

Wins +1, Runs +9

New Sire smoothed strike rate = (6 + 1) / (10 + 9) * 100% = 37%

This is more closer to reality. If you were to plot a chart using this method you won't get the sharp spikes at the beginning.

I don't know if this smoothing is of use to you. It's something to consider if you are dealing with low samples a lot.

http://www.paceadvantage.com/forum/show ... ostcount=5

[pippex]

very good link PP

Will get my teeth into this while I wait for my other computer to be fixed(I hope)

i am member of that forum so reading thread

[0909dylan]

What I am trying to say is...............lets assume you have a set of results that say..........30,000/1,000,000...........+ 35%.......you would probably be sure 100% that these results would be STATISTICALLY SOUND for the next selection............BUT if your stats say 3/10.......+ 35%........you would not. ............What would be the holy grail if you could assign a % CONFIDENCE to any possible 3/10...........+ 35% combination of figures in between ( for example ).

[pippex]

I think what it means is say the stat is 2 wins from 3

2+1/3+10 = 3/13..this is the adjusted strike rate =23%

[Pittsburgh Phil]

I think what it means is say the stat is 2 wins from 3

2+1/3+10 = 3/13..this is the adjusted strike rate =23%

Yes - If the average strike rate of all sires is around 10% then it is smoothed / normalised / dragged closer to that 10% level.

---

Dylan - for your answer the archie / chi test should be used. This will give a percentage figure indicating if the results were due to chance.

The holy grail is to find figures of 95% or even 99.5%. I don't fall for that though. In some cases it can be right to go with much lower confidence levels.

e.g.

If you have four systems which were all profitable (assume 30% strike rate, 10% ROI). Archie / chi tests indicate the following:

System A 50% chance the results were due to chance
System B 50% chance the results were due to chance
System C 50% chance the results were due to chance
System D 50% chance the results were due to chance

The guy in the lab with the geeky glasses and a pocket full of pens would not follow these systems because they don't meet the 95% confidence rule. He would only back a system if there was a better than 1 in 20 chance that the system results were due to luck.

Common sense says you should follow all four and strive to get better results!

[jth]

The guy in the lab with the geeky glasses and a pocket full of pens would not follow these systems because they don't meet the 95% confidence rule.

That was vcarme Gawd bless him