42% winners (!?) across all races, why it must show a profit.

[Carl]

For Hurrikane.

The easiest way for me to show it is
1) I had 5,200 races in my Aug db.
2) I did a make table on the winners.
3) I took the average odds of the lowest paying 2184 of them (sorted the table by nODDS).

This number came out to 1.44 to a dollar, or the average price of the 2184 lowest paying horses was $4.88.

Over 1000 races, the worst case scenerio would be

$ bet------------------$2000
$ collected-----------$2049 (420 winners @ average pay of the lowest paying horses $4.88)

Now in "real life", you would do much better than this, since you are bound to miss a few $4.00 horses "by accident" (they are "steamers", not showing form, etc.), which means your 42% winners would have to include some better paying doggies to make up the 42%. I would expect the "real life" profit on a 42% play across all races to be much closer to 30%, if not higher.

later,
Carl

[Carl]

Glen posted at PA's:
==========================
"A single factor that picks 40% winners over a 5000 race sample is impossible. I'm not saying that Dave S didn't get those results and I'm not calling him a liar. Just saying it is impossible to create a program that using a single factor (power rating, consensus fig or whatever) to select winners in every race without skipping one".
===========================
And the fellow who had a program that picked 42% winners across all races all tracks agreed sorta:
===========================
"It was by no means a "single factor." In fact, it was a 8 different systems, each system using 16 factors (weighted), and some of those factors actually contained as many as 34 other factors... In fact, some of THOSE factors contained several factors.

And thank you for giving me the benfit of the doubt.

The funny thing was that when I first brought it up, I mentioned it in passing. It held no real significance for me because I did not believe that the system had done much more than start with the "best" horse (which is usualy the favorite) and "pick a few fleas" (as Druther would call them.

It was never meant as "See how great this software is?" But someone challenged me about the probability... as they should have. They were right. The numbers did not add up.

Anyway, thanks again for keeping an open mind.

Regards,
Dave Schwartz "
===============================

Interesting stuff for me. And it does "hold real significance for me". ANYTHING that can show a profit betting every race "holds real significance" for me.

Question I have here is:
When does a factor become "a single factor"?

Believe it or not, I have spent some time thinking about this.

My thoughts about this are going along these lines. I'll do it in point form, IF anyone wants to pick up on any of the thread makes it easier to discuss.
1) "FX", although it is a factor made up of other factors, could be called a single factor. Why? Because it appears in every race.
2) Ev1 + AP1 could not be single factor because it does not appear in combination in every race. Every race "does not have one".
3) It would seem to me that the number the gentleman is using to pick 42% winners across all races is a single factor by this definition (admittedly mine), because even this number is made up of "eight systems with 16 weighted factors", it could be reduced to a single (long) equation, and would spit out a winner for every race tested, a second choice, etc. The only races I believe he excepted were hurdles and less than 5F, which he excluded from the study.

I find this fascinating. A verible "money machine" if you will. And betting every race too.

Q for other posters. Would you say Yes/no/maybe to this 42% winners feature being able to be considered a factor on it's own? Kinda like a "Super POW rating"? Would it be programable as such? If "no", why not? If "yes", can I buy it?

later,
Carl

[George]

Carl......nice post on an interesting subject.

Disagree that the system Dave described was a "single" factor. He really had eight different filters with different "weights" for each filter. Think such an approach could easily hit 42% playing all races but would PROBABLY show a small profit. Could also visualize average payoffs being slightly lower and showing a small loss.

I have built systems in the past from BRIS files that built a single composite from the top 20 or so factors in BRIS. Most produced 40 to 45 percent wins leaving out maiden and two year old races. All produced very marginal profits and/or losses.

The HTR number is a good example of a composite factor. While a pretty good number, it usually points you at the two overbet favorites. Mike's comp number is more useful because the pscan component sometimes is based on an overlooked horse.

It would be quite easy to take top 20 htr numbers and make a "supercomp" number. This could be done in Access as a calculated field. Guarantee you it is not worth the effort. Betcha Ken Massa has already tried and tested it :)

[Carl]

Thanks George.

Agree that the "HTR" number is a "super number made up of 20 numbers".

I think that the number Dave talks about is a "super number" also, made up of 8 sorts and then 16 variable factors.

And a Q for you. What is the highest percentage factor you have ever seen in any program that played (practically) every race? I don't think I have ever seen one past 31% in HTR, I don't know for other programs.

IMO, the "general public", all of us together with our "dollar votes" is the all time champ at 34% winners, but I am not sure.

And I agree, it is an interesting topic.

later,
Carl

[George]

If percent is across "all" races and for several thousand races would agree the general public is as good as it gets. Don't remember any program I have used or seen doing any better than the public's 34%. And I have used many over the years.

By applying a little bit of selectivity most good program's can get in the area of 40%. Usually pretty chalky at that percent rate.

I do find Dave's claim believable as he was in effect running 8 "spot" plays homing in on the key attributes of 8 different race types. Think I could get pretty close to 42% using that approach, but doubt that it is worth the effort. And yes, the net effect of that approach is an "super" composite.

Have followed the "ants" topic fairly closely. Made my living for a while programming neural nets and genetic algorithm's (which have neural nets "underneath"). While I wish the HSH folks great success, my experience with "ants" has been that they tend to weight toward "average" ground and produce results very similar to composites. Fairly high hit rate with low average mutuels. Not to say a breakthrough in training methodology could not be done!

I would be happy with a tool that gave me 30% hits on horses the public underbets. Or even 18% hits on horses the public seriously underbets. Ooops! think we already have that.

[Glen]

I respectfully disagree that is can not be done today over 5000 races looking FORWARD. Carl had the Peep Challenge (pick 100 horses and show a positive ROI)....Now I hereby invite anybody to the Pick Forty Two.

Pick Forty Two Rules as Follows:
Pick any track AP to YAV
Start with race 1 on the card picking 1 horse to win per race on EVERY race for the next 100. ie No cherry picking of races...Play all of the next 100 races without skipping.

Objective:
Pick 42 Winners.

[Carl]

Hi Glen,

I agree.

And a Q for you, since you use the modeller and I don't.

Backfitting 100 races, what is the best percentage you can see at any track for any HTR factor? I would think at some of the "small fields" tracks (IMO the key element in winning this wager) would approach 40% winners on some factors (MLO, HTR, P-SCAN). Yes/no?

Later,
Carl

[Glen]

Carl,

Looking back on the modeler over 100 races....Mid to high 30's. Although I think Tom founds somethings at FL that might be higher than that. Dunno cause I don't play FL.

[George]

Glen........not sure who you disagree with? I did not say that hitting 42% over 5000 races was easy or even sure it was doable. Think I can write a program that would, but the effort is more than such a program would be worth. Following are using PL5.

Five tracks in my modeler where a single htr factor hit near that rate at close to 100 races.
PIM 102 Races-HTR---->41%
FL 98 Races----K-------->44%
LAD 136 Races-PRB---->40%
PRM 82 Races-K-------->42%
RET 109 Races-PSCN-->56%
Am sure that none of the above would hold for 5000 straight races.
Do think somebody with a lot of time on their hands could "filter" using the modeler and get lucky over next 100 races.

[Carl]

George,

IF you could make a program that picked 42% winners across almost all races, I would suspect the ROI would be past 20%, as I think the average pay would be close to $6.00. I would think it would be hard to get an average pay less than that playing 100 races and winning 42 of them. You would have to stumble upon the odd $10.00 horse, and, as I pointed out in the original post, the worst you could average for payout would be $4.90 even if you had the very lowest paying horses in the sequence of 42.

later,
Carl

[George]

Carl,
Guess part of my problem is mentally getting "ALL" as in every race at every track! You are certainly right that such a program SHOULD make money. Of course if more than one person had a copy of that program the average mutual would be around $3.00.

BTW, the modeler numbers I showed in above thread surprised me! Should give us hope. Problem is to guess which htr factor at what five tracks will hit 42%. Think I will write a program to analyze that!! :)

[tbrown]

HTR is really hot at Finger Lakes right now.
The only way i can really capitalize on it is win bets and two-horse x-boxes.
Most of the plays are underlays, but overlays aren't any good if they don't win.
Use to shun those $13 exactas ,now I love them 'causse they are hitting so regular.
Got to play the short run to last for the long run. Afte rit cools off and we get a couple of rain showers, things won't be so chalky, but for now....
That is what I really like best about HTR - the ability to see what is going on right now-this hour. I have never played FL or WO with so much success as the last couple of months.

[MtKen]

What thread did DS first post his 42% remarks? I couldn't find it. Used to post over there a long time ago as debuscher22 & ,MikeD, in the Indians Forum under the same.

Tried to re-register over there but got blocked for some reason. Under the HTR vs. HSH thread Dick Schmidt claimed HTR was a 'stable' or 'static' software vs. the ever-evolving HSH. I'd have like to have pointed out that HTR is evolving as well. I'm sure HSH users will be glad to know they are getting a manual soon anyway.

[Carl]

Mt Ken,

Dave first posted about his "42% winners across all tracks and races ant" on the old PA board.

Dave gives a very fair synopsis of the post in the "Questions for Dave Schwartz" thread currently under "software" there. To his credit, he's not backing away from his statement, and says the ant performed that well or close to "on every 5000 race db he tested it on". His explanation of what was posted when on the old board is correct, you can read the history of the ant and the post there.

As to the "HSH dynamic/HTR static" thing, I wouldn't concern myself too much. As Eldo puts it "Every knock is a boost". Interesting you can't get on though.

Later,
Carl

[Gramps]

No such animal as a 40% selection method across all races and all tracks over any reasonably sized sample: And I don't need higher mathematics to prove it. Common sense will.

Almost everybody uses the same basic data. Some adjust parts in one way or another, but it comes down to using the same basic raw data.

If one factor alone (say, Beyer speed figures) could produce winners, then a lot of people would have found it and beaten it into unprofitability. Which is basically what happens every day with Beyer Speed figures.

Now you could make your own speed figures and figure out a way to do it that would be different enough from everybody else making speed figures that you'd have better figures. But your best speed figure wouldn't win 40% of the time, because the best speed figure horse doesn't win 40% of the time. Not over all races at all tracks, because the difference isn't marked enough in all cases to produce 40%.

And there lies the rub. The best looking horse by any way you analyze doesn't win 40% of the time. There's no telling if that horse has regressed since its last race or if other horses have improved significantly. The horse can't tell a computer that, and if it could, the computer probably couldn't be programmed to understand it---in other words, it would be hard to quantify.

And a 'super factor' comprised of many many sound factors is doomed to failure in my opinion. It is doomed for the simple reason that the more sub-factors that you plug into it, the more you're going to be pointing towards conventional thinking, and the more you move towards conventional thinking, the closer your win percentage (And ROI) approaches conventional thinking. The parimutuel top choice wins somewhere around 32% of all races, and many of those times the favorite is so as the result of money bet by somebody who knows something that most people don't. And it's usually a different somebody for each instance of what you might call 'smart' or 'inside' money.

But let's assume that the public's accumulated conventional wisdom (minus smart money) will select about 27% winners over all races at all tracks. The more factors that you plug into your super-rating, the more your number will approach that wisdom, and the closer your winning percentage is going to be to that 27%. It's the law of regression towards the mean.

Now it's possible, albeit barely, that Dave Schwartz's trained ant did get 40% winners over that 5,000 race sample. But if it was run over a much larger sample, say 50,000 races, I doubt it's going to reach that plateau. It might break 30%, and if it were something innovative enough, it might approach 35%--although i'd bet against the latter--but 40% is just out of bounds as far as I'm concerned. There isn't enough innovation to overcome the accumulated conventional thinking of the public at large over every race at every track, and this conventional thinking probably picks about 25-28% winners on its own merits.

I found a wonderful query that picked 100% winners over all races at all tracks. Then I went to the help file and found out what nFIN=1 meant. Back to the drawing board.

[hurrikane]

Gramps...I remember I did the nFIN thing too. Real exciting eh.

As for Dave..with all do respect..I think his 'ants' are just a great example of backfitting. To his credit...he never once said it had any value or the he used the formula for anything but a test.

[fred4now]

Gee, and I thought I was the only one the knew how to get 100% winners:D

[BMeadow]

Gramps gave an excellent synopsis of why the "my method picks 42% winners across all tracks for all races" not only isn't true, but cannot be true.

Let's use a little logic. In the entire history of public handicapping, not a single handicapper has ever come close to that number picking every race. In the entire history of database mining, no one has ever isolated ANY factor that comes close to this number picking every race.

If one factor cannot do this, why should a combination of several factors do this? You can't add two negative numbers and get a positive number.

Now if you want to restrict your plays to five-horse fields or allowance races or races in which only one horse shows early speed or only one horse has ever exceeded par or...whatever filters you wish to add...you can certainly improve your win percentage. But this statement does not mean you can bet ALL races at EVERY track using whatever super combination you believe you have, and hit anything close to 42% winners.

Maybe over 100 races you can. But not over 5000 races, betting every race with no exceptions.

In 5000 races, the public on average picks 1640 winners. And someone claims he can pick 2100 winners from the same group? An extra 460 winners (and they're not favorites, either, since the public has already picked 1640 winning favorites)?

If you believe you can, or believe anyone else can, you might be ripe for plucking by those selling systems/software/selections who make such claims.

[MikeDee]

Gramps

I've often felt that the level of accuracy that handicappers try to get out of PP data (was/is) way out of line as the what can reasonalbly be expected but I couldn't express it as well or as clear as you did. Good post.

[Gramps]

Barry:

I'm available for ghost writing :D

The whole thing sounds a lot like a throwback to the old Sartin days of "let's pick 66% winners with the top two selections using Factor Whatever. And even Sartin admitted that that couldn't be done playing every race at every track. But he claimed that it could be done over about 70% of the races, and I don't buy that, either. To hit 66% winners from the top two choices, I'd imagine that, since the top choice should win more often than the second one (example: the favorite wins about 33% and the second favorite wins about 20%), we're looking, for 67% on the top two choices, at about 40% for the first choice and 27% or so for the second choice. Which is saying that Factor Whatever can do 14% better than the public can. Which puts us right back to square one: 40% winners for the top horse. And I doubt that can be done over 70% of all races, and I'm not sure it can be done over 50% of all races.

As Barry said, the parimutuel favorite wins about 1650 out of 5000 races. That's 33%. But out of those 1650, probably about 400 ended up the favorite from other means than what conventional handicapping (e.g. past performances and whatever else makes up the public's mindset). In other words, those 400 became favorite (where they would not have otherwise been so) as a result of money bet by somebody or a series of somebodies that knew something that the public as a whole didn't. That could be the stables, trip handicappers that saw something that the charmaker missed, guys who make their own speed/power/etc. numbers, or guys using information that the public as a whole doesn't have access to. So conventional handicapping would probably pick closer to 24-28% winners. And to up that to 40% over every race at every track just doesn't happen. Not over the long run.

And if somebody claims that they can do it, but won't document it forwardly or over a large enough sample to have a low standard deviation or margin of statistical error, then it's very unlikely to be true.