expected wiiners /number of winners

[acorn54]

i think the reason research isn't properly done with horses is that the proper methods in statistics are not used
if it shows a profit the system is deemed viable
in quirin's book "winning at the races", in the appendix, he gives the formula for testing whether the numbers of winners is greater than should be expected .
i think this step is not applied in alot of testing of horse methodologies/systems
guy (acorn)

[Equineer]

Acorn54,

Software handicappers do use a variety of confidence tests when they measure their results. For the Quirin tests that you referenced, here are links to both an example and the formula.

Example:
http://www.paceadvantage.com/forum/showthread.php?s=&postid=132185&highlight=all+categories#post132185

Formula:
http://www.paceadvantage.com/forum/showthread.php?s=&postid=131345&highlight=standard+normal#post131345

[sjk]

I have always considered the profitability test very compelling.

[formula_2002]

Originally posted by sjk
I have always considered the profitability test very compelling.

as do I, and it has saved me a fortune!!!

[sjk]

I was looking at the formula in Equineer's post and it looks to me that if you are winning over the course of several thousand races (and it is not due to a few extremely large scores) you cannot help but have a relatively healthly score according to the formula.

As I understand it:

If you assume 12% winners paying an average mutual of $20 (for $2) over 5000 races,

NW = 600
EW = 500 (less if you figure in the track take)

so (NW - EW) / SqrRoot(EW * (1 - (EW / NH)))

= 4.71

which I assume is good.

Formula_2002 can tell me if this is a correct application of the formula and how to interpret the results.

[formula_2002]

A CONFIDENCE LEVEL BETWEEN
1.6 AND 1.9 = 90% CONFIDENCE
1.9 AND 2.5= 95% CONFIDENC
>=2.5= 99% CONFIDENCE.
<1.6 = GET MORE DATA

4.72 is correct and would mean you could walk on water and not worry about sinking.
I guess its almost as good a bet on the sun rising in the east every day.

but you want to test on the smallest possible incremental odds range.

To group 2-1 , 3-1 , 10-1 and 20-1 and average those odds to come out with an average is not a good idea.

Also, you have to determine what the expected number of winners is by actuall looking at emperical data.

Equiengineer's tables seem to do just that.

Once you have everything just the way you would like to see it, with promising results., do it all over again with new data!!!

;)

[Equineer]

The formula is exactly the one used in books by W.L. Quirin, mathematics professor at Adelphi University (which presumably was not founded by Adelphi Neumann, the twin brother of Alfred E. Neumann). :)

The formula is a sample-size-sensitive test for the significance of a difference between two proportions, in this case between expected winners (EW) and number of winners (NW) within a sample size (NH = number of horses in sample).

In Quirin's books, if the result is outside the range of -2.5 and 2.5, then the difference is considered significant in relation to sample size.

[hurrikane]

NW/EW sounds like IV to me.

didn't think that was allowed up here. :D

[Equineer]

Originally posted by hurrikane
NW/EW sounds like IV to me.

didn't think that was allowed up here. :D I know a guy who owes his life to an IV the paramedics applied after a series of bad beats dropped him to the floor in a comatose state. :)

[sjk]

Joe,

I have never understood why it would be helpful to divide the sample into a bunch of small pieces based on odds. If you have already determined significance for the broader sample, to use smaller samples would just make it harder to demonstrate what you already know.

I could not do the calculation anyway since I don't know what the losers potential payoffs would have been. I can compare all of the winners with the full sample but if I started dividing the winners into groups I would have no corrresponding way to segment the losers.

[acorn54]

you have to use very narrow odds range groups
lets say your system after 100 bets shows a 25% profit. if those winners were bombs,say 40-1 or 50-1 to give an extreme example those odds ranges may have won at more than their expected number of winners for the horses your system picked at the high end,however your system may have been 0-19,or 1-19 at say 4-1 or 7/2.
guy (acorn)

[sjk]

Acorn,

I would agree that you can't take the results of a 100 race sample very seriously, especially if there are longshot winners in the sample.

I think the formula the Equineer and Formula have referenced would show that such results are suspect.

[acorn54]

in winning at the races quirin developed a route system and it was based on a 300 race sample of route races. (the multiple regression section of the book in the last chapter before the appendix) so 100 races might be too small but 300 must be large enough if professor quirin relied upon that size for his research.
guy (acorn)

[sjk]

I am not familiar with your reference. The number of races that could reasonably be assembled as a sample has changed considerably over the years. Unless a method was so specialized as to only generate a play once in a great while I don't see why you would not want to test it over a large number of plays.

I would expect to use tens of thousands of races in testing a general purpose method.

[karlskorner]

As I stated a couple of years back, I spent the winter of 86/87 with Bill Quirin, Jim Quinn, Joe Takach and others. NEVER SAW THE MAN MAKE A WAGER. His students put together the statistics and theories and he was able to publish 4 books from it. Face reality gentlemen, the man was an author and what he wrote looked good in print.

[hurrikane]

Nice to see you back Karl. How have the tracks held up with all the storms?

[acorn54]

Originally posted by karlskorner
As I stated a couple of years back, I spent the winter of 86/87 with Bill Quirin, Jim Quinn, Joe Takach and others. NEVER SAW THE MAN MAKE A WAGER. His students put together the statistics and theories and he was able to publish 4 books from it. Face reality gentlemen, the man was an author and what he wrote looked good in print.

yes he was an author but he is also knowledgeable in the correct application of statistical research and if he used only a 300 route race sample to validate his mulitple regression formula he probably was doing a valid statistical test of the vialbility of his route system
guy (acorn)

[karlskorner]

Hurrikane;
I was in Africa for the first 3 storms, caught the last one last week, actually S. Florida didnt fare to bad. the last one blew some palm trees around, but CRC came through with little or no damage, drove over to Gulfstream last week, it looks like they are on schedule with their rebuilding, but it's going to be a half ass meet for sure.

Acorn54

"Valid statistical tests of the vialbility of a system" has nothing to do with the business of racing horses. I am quite sure somebody can run a test showing that gray horses, blind in one eye, win certain type of races.

[Jeff P]

I've always felt that the best way to test any method, after its factor definitions and ranges have been finalized, is to confront it with FRESH data. In my experience, if the method performs well with races that it has never seen, races not used during the method's creation, then and only then do you have something.

I have two separate sets of data in my database. First is what I call Development Data. This I use for primary research-discovering the effects of various factors and combinations of factors- identifying trends if you will. From this, I develop specific handicapping models that attempt to exploit whatever it was that my research uncovered.

The second set of data is what I call Validation Data. What I use this for is to test how each specific model performs when confronted with fresh data. When working with Validation Data, I won't re-work factor definitions for a model that almost passes muster. A model either passes validation for me or it dosen't. Models that don't validate I simply throw out. The models that pass Validation are the ones I keep and use.

I do feel that calculating a confidence level and looking at expected winners vs actual winners for tests done on Validation Data does have some use. Mainly, it makes me feel better about the model in question.

and if he used only a 300 route race sample to validate his mulitple regression formula he probably was doing a valid statistical test

Sample size is a separate issue. I've seem some models that perform well going forward, yet their discovery was based on Development and Validation tests involving only a handful or racing days. I've seen other models where the merit (or lack of it) only became apparent after looking at thousands of racing days. I think an awful lot depends on the factors and sometimes combinations of factors involved- how they are used and combined- in what context they are used, etc.

For me, all of this makes racing, despite the best efforts of some very brilliant, and even not so brilliant (LOL) people, a very fascinating and imperfect science.

[Equineer]

Originally posted by karlskorner
As I stated a couple of years back, I spent the winter of 86/87 with Bill Quirin, Jim Quinn, Joe Takach and others. NEVER SAW THE MAN MAKE A WAGER. His students put together the statistics and theories and he was able to publish 4 books from it. Face reality gentlemen, the man was an author and what he wrote looked good in print. Quirin's books should make any reader hesitant to wager. The same is true of other authors if you discount their anecdotal examples of success. Quirin simply deserves credit for quantifying the situation more clearly than most.

Quirin presents hundreds of examples of failure and near misses. While he discovers a few glimmers of hope, he also explains why pari-mutuel wagering will defeat any publicized method or angle for systematically picking winners.

In this day and age, every shred of published information soon becomes broadband knowledge. The only horseplayers who can hope to systematically beat the mutuels by picking winners are those with a narrowband edge.

From valid insider information, whether legal or illegal... to acute skills judging raceday conditions and equine physicality, financial success from picking winners depends on exploiting some form of narrowband information. The track savvy that we associate with legendary winners amounts to nothing more than exploitation of some form of narrowband information, even if it stems from very rare perceptive and mental skills.

Is it rational to suppose that handicapping authors don't understand that to publish is to perish as players? Unlike many other authors, Quirin simply laid the cards on the table... by empirically demonstrating that trying to pick winners by using broadband information is a foolish approach to a game that is really a pari-mutuel competition among players.