Questions About Trifectas and Supers

[Dave Schwartz]

I have need of some conversation.

I am writing some very deep A.I. and am having some issues with structuring trifecta and superfecta bets.

Understand that I am not designing how the A.I. will bet these plays in a rigid manner. Rather, I am trying to design a way for the A.I. to be able to DISCOVER how it should bet.

What I have it doing now is (indirect) Topological Data Analysis.

WHAT THAT MEANS
==============
Image that there are a set of systems for handicapping a horse - 250 for previous starters, 100 for FTS, and 50 for foreign horses.
(When I say 250 "systems" I mean a set of weighted factors for producing a "score" for an individual horse.)

There are a set up rules - 100's of them (created by the A.I.) - that assign points to the 250 systems. The idea is that the A.I. can decide that this is the "type of horse" that should be handicapped using (say) System #17, while the 2nd horse in the same race might be a System #221 horse.

As an example, the A.I. could be able to UNDERSTAND the difference between an ES horse with a class edge who is likely to quit completely if he doesn't win, and an ES horse that is likely to persevere for 2nd or 3rd even if he is headed in the stretch.

HERE'S MY PROBLEM
==================
Creating an environment where the A.I. engine can UNDERSTAND how to bet the win pool is actually relatively easy. All it has to do is project the odds for each horse and the hit rate to determine profitability.

Even place and show bets are relatively easy to figure out.

The problems begin with exactas - because projecting exacta payoffs is not so easy. I think I have this one whipped.

The REAL PROBLEMS begin with trifectas because there really isn't much chance of projecting a trifecta or superfecta payoff.

Understand that the system is smart enough to know (or, hopefully will be) that a horse can be tossed from finishing 1st but still be left underneath.

So, I am hoping for some discussion from anyone who has a mathematical idea to put forth about how to address this situation.

I know that the bar is very high here. But if you have knowledge to share, please bring it forth, whether publicly or privately.


Regards,
Dave Schwartz

[Exotic1]

I have need of some conversation.

I am writing some very deep A.I. and am having some issues with structuring trifecta and superfecta bets.

Understand that I am not designing how the A.I. will bet these plays in a rigid manner. Rather, I am trying to design a way for the A.I. to be able to DISCOVER how it should bet.

What I have it doing now is (indirect) Topological Data Analysis.

WHAT THAT MEANS
==============
Image that there are a set of systems for handicapping a horse - 250 for previous starters, 100 for FTS, and 50 for foreign horses.
(When I say 250 "systems" I mean a set of weighted factors for producing a "score" for an individual horse.)

There are a set up rules - 100's of them (created by the A.I.) - that assign points to the 250 systems. The idea is that the A.I. can decide that this is the "type of horse" that should be handicapped using (say) System #17, while the 2nd horse in the same race might be a System #221 horse.

As an example, the A.I. could be able to UNDERSTAND the difference between an ES horse with a class edge who is likely to quit completely if he doesn't win, and an ES horse that is likely to persevere for 2nd or 3rd even if he is headed in the stretch.

HERE'S MY PROBLEM
==================
Creating an environment where the A.I. engine can UNDERSTAND how to bet the win pool is actually relatively easy. All it has to do is project the odds for each horse and the hit rate to determine profitability.

Even place and show bets are relatively easy to figure out.

The problems begin with exactas - because projecting exacta payoffs is not so easy. I think I have this one whipped.

The REAL PROBLEMS begin with trifectas because there really isn't much chance of projecting a trifecta or superfecta payoff.

Understand that the system is smart enough to know (or, hopefully will be) that a horse can be tossed from finishing 1st but still be left underneath.

So, I am hoping for some discussion from anyone who has a mathematical idea to put forth about how to address this situation.

I know that the bar is very high here. But if you have knowledge to share, please bring it forth, whether publicly or privately.


Regards
Dave Schwartz

Do trifecta payouts correlate to exacta probables by some factor of odds, field size and running style? Superfecta payouts are getting to be ridiculously low. It's almost at a point of trifecta * 2 or 3 regardless (or without an appreciation) of the odds on the 4th place finisher.

Sounds like you've mastered exacta probable payouts to their real probabilities. That alone sounds very difficult if one assumes that win odds vs. odds of running second w no chance of winning, are two different subjects. Not sure that assumption is correct.

[Robert Fischer]

Payout probables aren't necessary for tris and supers, at least not in a vacuum.
If I personally had to do a model, I'd do a value model, and just run it separately, provided it was profitable. I wouldn't care about payout, just that I was buying tickets that were more valuable than the price.

If you wanted those tris and supers to be part of a multi-layered dutch/hedge/leveraged 'maximum churn' algorithm, then you'd need (I think?) the payout probables.
I don't have a complete picture in my head.
Pool Size, Field size, a Win Odds distribution, Odds Ranking, and maybe the Under-the-Fav Exacta Probables would probably be nearly all you need in live data. Then generate a low estimate for payouts.

[Dave Schwartz]

Do trifecta payouts correlate to exacta probables by some factor of odds, field size and running style? Superfecta payouts are getting to be ridiculously low. It's almost at a point of trifecta * 2 or 3 regardless (or without an appreciation) of the odds on the 4th place finisher.

Sounds like you've mastered exacta probable payouts to their real probabilities. That alone sounds very difficult if one assumes that win odds vs. odds of running second w no chance of winning, are two different subjects. Not sure that assumption is correct.


I have no idea if trifectas correlate. Hence, my problem. LOL


Payout probables aren't necessary for tris and supers, at least not in a vacuum.
If I personally had to do a model, I'd do a value model, and just run it separately, provided it was profitable. I wouldn't care about payout, just that I was buying tickets that were more valuable than the price.

How would you do a value model without payoffs?

[Robert Fischer]

How would you do a value model without payoffs?

Horses that the public are significantly wrong about, are going to generate underlays and overlays.

[HalvOnHorseracing]

Conceptually I would think you are trying to calculate the percentage of the trifecta pool represented by different combinations without having the printout that the track gets. So if you have a fair pay calculation for the exacta, the trick may be to come up with a multiplier, perhaps even a weighted multiplier, for the third place horse in a trifecta.

I'm just stabbing here, but if you have 2-1 over 8-1 over 5/2 you might take the fair pay exacta (let's call it $50 for $2) and multiply it by 2.5 (the odds of the third place horse to one) for a trifecta pay of $125 for $2. The problem, is that payouts can be highly different depending on which place the price horse is in, or whether the trifecta is one of the high probability outcomes, and how many horses are in the race.

So you might want to have a multiplier for the number of horses, and one for the odds of the third place horse. This way if the longshot is first or second, you've accounted for it in the exacta pay, and you just have to figure out how to multiply that number by something representing the third place horse (with the longer the odds the greater the deviation from the third place finisher's odds to one) and an adjustment for the number of starters (because the fewer the starters, the fewer the combinations and vice versa.)

I'm not sure how much this would help, but it just struck me that if you can calculate the exacta pay, you should be able to use that as a starting point to calculate the trifecta.

The superfecta - I'm not sure the same sort of calculations would work. I don't have a concept on the super.

[Dave Schwartz]

Conceptually I would think you are trying to calculate the percentage of the trifecta pool represented by different combinations without having the printout that the track gets. So if you have a fair pay calculation for the exacta, the trick may be to come up with a multiplier, perhaps even a weighted multiplier, for the third place horse in a trifecta.

I'm just stabbing here, but if you have 2-1 over 8-1 over 5/2 you might take the fair pay exacta (let's call it $50 for $2) and multiply it by 2.5 (the odds of the third place horse to one) for a trifecta pay of $125 for $2. The problem, is that payouts can be highly different depending on which place the price horse is in, or whether the trifecta is one of the high probability outcomes, and how many horses are in the race.

So you might want to have a multiplier for the number of horses, and one for the odds of the third place horse. This way if the longshot is first or second, you've accounted for it in the exacta pay, and you just have to figure out how to multiply that number by something representing the third place horse (with the longer the odds the greater the deviation from the third place finisher's odds to one) and an adjustment for the number of starters (because the fewer the starters, the fewer the combinations and vice versa.)

I'm not sure how much this would help, but it just struck me that if you can calculate the exacta pay, you should be able to use that as a starting point to calculate the trifecta.

The superfecta - I'm not sure the same sort of calculations would work. I don't have a concept on the super.

Actually a remarkably good idea!

Maybe not the specifics, but what I am hearing you say is that I need an algorithm (formula) which considers:

BaseNumber + WinPool% * Wgt1 + WinPool% * Wgt2 + WinPool% * Wgt3.

This is a linear version of what the formula needs to look like. I could use a more robust type of formula. In fact, I do for the win pool. (The AI figures out what these variables should be for each horse.)

[Franco Santiago]

Dave, I think your question is rhetorical in nature, but just in case, here's my sincere answer:

Don't waste your time. No is ever going to throw out a horse in the 3rd slot based on the AI. Why? Fear of regret. If the first two came in at long odds and Joe Blow doesn't have the 3rd horse at 5-1 because Mr. Arti Intel hold him it was a bad bet, he'll promptly go to the strong box or safe, get his gun, chamber a round, and end it all LOL!!!

Seriously, people handicap much more loosely for the show spot and they toss in all kinds of crap. They ain't tossing one out based on AI...now they might toss one IN based in AI, but it won't be because of some value algorithm...it'll be because they want to HIT THE BET.

Now, come clean, you don't really expect a cogent answer from anyone, do you?

[HalvOnHorseracing]

Actually a remarkably good idea!

Maybe not the specifics, but what I am hearing you say is that I need an algorithm (formula) which considers:

BaseNumber + WinPool% * Wgt1 + WinPool% * Wgt2 + WinPool% * Wgt3.

This is a linear version of what the formula needs to look like. I could use a more robust type of formula. In fact, I do for the win pool. (The AI figures out what these variables should be for each horse.)

Yes, that is essentially where I was going. The variables are the odds of the top the three finishers, the order in which they finish, the amount of the pool, the number of horses and the distance of the three finishers from the crowd favorites.

Anecdotally, if the first two choices come in in that order, the odds of the third place horse. However, if the winner runs out, or the second and third probability horses run out, or longshots finish first and second or first and third or even second and third, you start to get proportionally higher payoffs. So if you can contstruct the algorithm to account for those variables, you may come close to predicting the more complex combinations.

[menifee]

I have never understood why you can't look up a tri price and get an accurate quote. There should be real time systems to tell you the will pay for a certain combination.

Other than the variables already mentioned (odds of third place horse, exacta price, size of field), pool size matters quite a bit. Larger pools more efficient pricing.

[whodoyoulike]

...
Maybe not the specifics, but what I am hearing you say is that I need an algorithm (formula) which considers:

BaseNumber + WinPool% * Wgt1 + WinPool% * Wgt2 + WinPool% * Wgt3.

This is a linear version of what the formula needs to look like. I could use a more robust type of formula. In fact, I do for the win pool. (The AI figures out what these variables should be for each horse.)

I saw the consideration of the WinPool also in another thread and the question I have when you do this is ..... each exacta, trifecta and superfecta pools are separate wagering pools.

Do these pools correlate to the WinPools?

Since we don't receive that info, I don't think we'll ever know.

Do the tracks even know (probably yes)?

[Dave Schwartz]

Dave, I think your question is rhetorical in nature, but just in case, here's my sincere answer:

Don't waste your time. No is ever going to throw out a horse in the 3rd slot based on the AI. Why? Fear of regret. If the first two came in at long odds and Joe Blow doesn't have the 3rd horse at 5-1 because Mr. Arti Intel hold him it was a bad bet, he'll promptly go to the strong box or safe, get his gun, chamber a round, and end it all LOL!!!

Seriously, people handicap much more loosely for the show spot and they toss in all kinds of crap. They ain't tossing one out based on AI...now they might toss one IN based in AI, but it won't be because of some value algorithm...it'll be because they want to HIT THE BET.

Now, come clean, you don't really expect a cogent answer from anyone, do you?


LOL - Apparently not from you.

This is not for "Joe Blow."

PS: Halvey did pretty well.

[Dave Schwartz]

I saw the consideration of the WinPool also in another thread and the question I have when you do this is ..... each exacta, trifecta and superfecta pools are separate wagering pools.

Do these pools correlate to the WinPools?

Since we don't receive that info, I don't think we'll ever know.

Do the tracks even know (probably yes)?

Since there is no trifecta or superfecta grid available, we cannot know.

[Jeff P]

Obviously you cannot know with precision what the pool amounts were on non winning individual tri and super combinations in individual races.

But you do have historical payoffs that can be scraped from chart data or read from a db that contains chart data.

And from that you can get a pretty good idea overall if you build a classification system that:

a. identifies like situations

- and

b. uses historical payoffs on winning combinations.



Dave, I think your question is rhetorical in nature, but just in case, here's my sincere answer:

Don't waste your time. No (one) is ever going to throw out a horse in the 3rd slot based on the AI. Why? Fear of regret. If the first two came in at long odds and Joe Blow doesn't have the 3rd horse at 5-1 because Mr. Arti Intel hold him it was a bad bet, he'll promptly go to the strong box or safe, get his gun, chamber a round, and end it all LOL!!!

Seriously, people handicap much more loosely for the show spot and they toss in all kinds of crap. They ain't tossing one out based on AI...now they might toss one IN based in AI, but it won't be because of some value algorithm...it'll be because they want to HIT THE BET.

Now, come clean, you don't really expect a cogent answer from anyone, do you?

Actually there's quite a bit of good info in the above post -- especially the bolded part.

People do handicap much more loosely and toss all kinds of crap into the underneath slots.

Personally - I've found that instead of modeling to get a Y/N decision whether or not to include a horse underneath in tri and super tickets:

You'll do much better by modeling the likelihood of individual horses to achieve an exact finish position.

And from there base ticket structure around likelihood estimates of exact finish positions for individual horses.



-jp

.

[AltonKelsey]

I think it's safe to say there is a tremendous amount of variance in the payoffs.

So even if you had a respectable 'model', there will still be a lot of room for error in any given race payoff. Chalky finishes might be less susceptible to that , but anything involving longshots, expect wild differences in price.


You had a finish at BEL yest , where the EX paid 170 or so. 10-1 over 20-1. 5 horse field , but still. Good luck trying to bang out long term profits with payouts like that .

[HalvOnHorseracing]

Seriously, people handicap much more loosely for the show spot and they toss in all kinds of crap. They ain't tossing one out based on AI...now they might toss one IN based in AI, but it won't be because of some value algorithm...it'll be because they want to HIT THE BET.

I've opined in the past that people have been trained to select the winner, but they have not been trained to accurately evaluate any other finish position. So if you followed the exacta probability discussion, second place is simply an extension of the win odds, with the assumption that the second best horse has the highest probability of finishing second, the third best third, and so on. Of course, this is not always true. "Sucker" horses (or horses with seconditis) may have a 2% chance of winning and a 50% chance of finishing second, and an exacta system based on win probability might underestimate this horse. The other issue is that as you place horses in slots below win, their win odds become less and less valuable as a predictor. Which is why longshots regularly occupy the backholes, and people use "all kinds of crap" in the third and fourth slots. I actually did some graphs of different types of runners (high win probability, low win probability, mid-level probability, in-the-money type, all or nothing type) and suggested even the low win types had probabilities of finishing third or fourth much higher than the win probability, and you had to draw the line farther down the list of starters than you might just based on win odds.

[thaskalos]

I've opined in the past that people have been trained to select the winner, but they have not been trained to accurately evaluate any other finish position. So if you followed the exacta probability discussion, second place is simply an extension of the win odds, with the assumption that the second best horse has the highest probability of finishing second, the third best third, and so on. Of course, this is not always true. "Sucker" horses (or horses with seconditis) may have a 2% chance of winning and a 50% chance of finishing second, and an exacta system based on win probability might underestimate this horse. The other issue is that as you place horses in slots below win, their win odds become less and less valuable as a predictor. Which is why longshots regularly occupy the backholes, and people use "all kinds of crap" in the third and fourth slots. I actually did some graphs of different types of runners (high win probability, low win probability, mid-level probability, in-the-money type, all or nothing type) and suggested even the low win types had probabilities of finishing third or fourth much higher than the win probability, and you had to draw the line farther down the list of starters than you might just based on win odds.

IMO...the second-best horse will NOT finish second if it shares the same running style as the race's BEST horse.

[ReplayRandall]

IMO...the second-best horse will NOT finish second if it shares the same running style as the race's BEST horse.
What crappy book did you read that from??....I got two authors in mind, who might even post here at PA.

[DeltaLover]

IMO...the second-best horse will NOT finish second if it shares the same running style as the race's BEST horse.

Whether this is true or false will be shown by the related data.

How the data are presented to the decision algorithm is where the real challenge lies while the deepness of the machine learning is of a secondary importance, in the sense that it is relatively easy to switch network topologies and rely on a trial and error process which although it might become time consuming it can still be automated and run automatically in the background.

This thread is representative of the most interesting topic that I can think about horse betting but it is not presented in a sufficiently detailed level that can allow for concrete conversations and exchange of ideas and methodologies. I feel that to proceed in a right direction we certainly need training and testing data otherwise the thread will become another "opinion battlefield" as usually.

[HalvOnHorseracing]

IMO...the second-best horse will NOT finish second if it shares the same running style as the race's BEST horse.

It was an assumption made for the purposes of calculation using a generic example with no handicapping consideration. Clearly in the real world you don't ignore handicapping.

[HalvOnHorseracing]

Whether this is true or false will be shown by the related data.

How the data are presented to the decision algorithm is where the real challenge lies while the deepness of the machine learning is of a secondary importance, in the sense that it is relatively easy to switch network topologies and rely on a trial and error process which although it might become time consuming it can still be automated and run automatically in the background.

This thread is representative of the most interesting topic that I can think about horse betting but it is not presented in a sufficiently detailed level that can allow for concrete conversations and exchange of ideas and methodologies. I feel that to proceed in a right direction we certainly need training and testing data otherwise the thread will become another "opinion battlefield" as usually.

The theory behind calculating fair pay on the exacta is sound enough. You handicap the probabilities of each respective horse finishing first. The tricky part is making a calculation for the probability of any respective horse finishing second. Simplistically you can adjust the place odds after putting a horse in the win slot by redistributing the percentage from the win horse and redistributing it to the remaining runners.

Obviously that is a formulaic method with no handicapping beyond the win hole.

After putting a respective horse in first, you can also readjust the place percentages by rehandicapping the race. That is a subset of Thask's point. When Horse A finishes first, that may compromise the probability of Horse B finishing second. But you might also argue that you took that into account when you handicapped Horse A on top, in which case Horse B wouldn't have been the second best horse in the first place.

If I'm writing for the average horseplayer, I'm going to provide an easy to use method that is sustainable. I think that is

(win probability * place probability) = exacta probability (converted to odds)

exacta odds + margin of safety = acceptable exacta pay

A valuable discussion would be how to calculate win and place odds, because the rest of it is mechanical.

And I'd argue that once you have a useful exacta price, you can extrapolate that to trifectas and supers.

[DeltaLover]

The theory behind calculating fair pay on the exacta is sound enough. You handicap the probabilities of each respective horse finishing first. The tricky part is making a calculation for the probability of any respective horse finishing second. Simplistically you can adjust the place odds after putting a horse in the win slot by redistributing the percentage from the win horse and redistributing it to the remaining runners.

Obviously that is a formulaic method with no handicapping beyond the win hole.

After putting a respective horse in first, you can also readjust the place percentages by rehandicapping the race. That is a subset of Thask's point. When Horse A finishes first, that may compromise the probability of Horse B finishing second. But you might also argue that you took that into account when you handicapped Horse A on top, in which case Horse B wouldn't have been the second best horse in the first place.

If I'm writing for the average horseplayer, I'm going to provide an easy to use method that is sustainable. I think that is

(win probability * place probability) = exacta probability (converted to odds)

exacta odds + margin of safety = acceptable exacta pay

A valuable discussion would be how to calculate win and place odds, because the rest of it is mechanical.

And I'd argue that once you have a useful exacta price, you can extrapolate that to trifectas and supers.

The objective is not to discover the winning probabilities but the error that is embedded in the crowd’s betting.

The winning probability of a horse is not only useless for betting purposes (since we are not betting into fixed lines) but, given the chaotic nature of the game it is also impossible to calculate with the necessary accuracy. We might end up with a valid way to order the horses based in their winning probabilities (similar to the behavior of the final odds as formed by the crowd) but such a model will not spawn precise enough numbers to allow for successful betting (which again is getting worse given the fact that you never know in advance your expected payout).

What is possible though, is to build a probability distribution expressing the accuracy of the crowd in a cumulative way. Having this kind of information available it is valid to expect that your potential advantage (of disadvantage if the crowd will prove your model wrong) to materialize after a relative long sequence of races. In other words and to paraphrase a common saying: “You can beat the races but you cannot beat the race:

[thaskalos]

The objective is not to discover the winning probabilities but the error that is embedded in the crowd’s betting.

The winning probability of a horse is not only useless for betting purposes (since we are not betting into fixed lines) but, given the chaotic nature of the game it is also impossible to calculate with the necessary accuracy. We might end up with a valid way to order the horses based in their winning probabilities (similar to the behavior of the final odds as formed by the crowd) but such a model will not spawn precise enough numbers to allow for successful betting (which again is getting worse given the fact that you never know in advance your expected payout).

What is possible though, is to build a probability distribution expressing the accuracy of the crowd in a cumulative way. Having this kind of information available it is valid to expect that your potential advantage (of disadvantage if the crowd will prove your model wrong) to materialize after a relative long sequence of races. In other words and to paraphrase a common saying: “You can beat the races but you cannot beat the race:

Yes...but is the crowd exhibiting the same wagering "characteristics" across different time periods...or does the crowd "get smarter" as time goes on...thus gradually negating whatever "advantage" we might think we have over the game? If it is indeed "getting tougher to win in this game", as most of us seem to think, then the assumption must be made that any predetermined advantage that we might currently have can hardly be used as an "expectation" of things to come.

[traynor]

The objective is not to discover the winning probabilities but the error that is embedded in the crowd’s betting.

The winning probability of a horse is not only useless for betting purposes (since we are not betting into fixed lines) but, given the chaotic nature of the game it is also impossible to calculate with the necessary accuracy. We might end up with a valid way to order the horses based in their winning probabilities (similar to the behavior of the final odds as formed by the crowd) but such a model will not spawn precise enough numbers to allow for successful betting (which again is getting worse given the fact that you never know in advance your expected payout).

What is possible though, is to build a probability distribution expressing the accuracy of the crowd in a cumulative way. Having this kind of information available it is valid to expect that your potential advantage (of disadvantage if the crowd will prove your model wrong) to materialize after a relative long sequence of races. In other words and to paraphrase a common saying: “You can beat the races but you cannot beat the race:

Exactly. The probability of a horse winning--if only calculated as a factor of particular attributes of that individual horse--is, at best, no more than a rough approximation. To be accurate enough to be worth betting on, it would have to be the result of calculating that individual horse's win probability as a factor of the combined attributes of this specific mix of horses in this specific race. And even then would still be only an approximation.

You did a really good job of explaining your perspective in another thread a few days ago, and I am finally beginning to understand what you have been writing all along.

Specifically (as you point out above) the useful thing is not necessarily the win probability (or place probability, or whatever probability), but what the other bettors believe those probabilities to be.

That makes the study of combinations of attributes in that specific group of horses more useful than study of the individual horses' attributes as predictors. A simplistic example would be trifecta/superfecta wagers "automatically" including class droppers, or some other favored attribute. Many of the inclusions (in particular those based on what some believe to be "pace analysis" and "preferred running styles") have little or nothing to do with the actual value of that particular attribute of that particular horse in that particular race.

To clarify, " the study of combinations of attributes in that specific group of horses," consider a race with two near-equal front runners. Bettors may (as a mindless knee-jerk reflex based on spurious premises) believe an "early speed duel" will "burn out" both front runners, and "set the race up for a strong closer"--and bet accordingly.

My understanding of what you are writing (which may be totally wrong, and if so, I apologize) is that you believe it is more useful (or at least as useful) to study what the bettors are likely to do as a result of seeing certain attributes/factors in a given group of horses as a predictor than to study (only) the certain attributes/factors as predictors.

And that knowledge is best derived (and perhaps can ONLY be derived) from studying trends and patterns in large groups of races, rather than trying to divine the future based on the few races of which one has "personal experience."

[AndyC]

The objective is not to discover the winning probabilities but the error that is embedded in the crowd’s betting.

Without the winning probabilities how would someone know if the public is making an error?

[thaskalos]

Exactly. The probability of a horse winning--if only calculated as a factor of particular attributes of that individual horse--is, at best, no more than a rough approximation. To be accurate enough to be worth betting on, it would have to be the result of calculating that individual horse's win probability as a factor of the combined attributes of this specific mix of horses in this specific race. And even then would still be only an approximation.

You did a really good job of explaining your perspective in another thread a few days ago, and I am finally beginning to understand what you have been writing all along.

Specifically (as you point out above) the useful thing is not necessarily the win probability (or place probability, or whatever probability), but what the other bettors believe those probabilities to be.

That makes the study of combinations of attributes in that specific group of horses more useful than study of the individual horses' attributes as predictors. A simplistic example would be trifecta/superfecta wagers "automatically" including class droppers, or some other favored attribute. Many of the inclusions (in particular those based on what some believe to be "pace analysis" and "preferred running styles") have little or nothing to do with the actual value of that particular attribute of that particular horse in that particular race.

To clarify, " the study of combinations of attributes in that specific group of horses," consider a race with two near-equal front runners. Bettors may (as a mindless knee-jerk reflex based on spurious premises) believe an "early speed duel" will "burn out" both front runners, and "set the race up for a strong closer"--and bet accordingly.

My understanding of what you are writing (which may be totally wrong, and if so, I apologize) is that you believe it is more useful (or at least as useful) to study what the bettors are likely to do as a result of seeing certain attributes/factors in a given group of horses as a predictor than to study (only) the certain attributes/factors as predictors.

And that knowledge is best derived (and perhaps can ONLY be derived) from studying trends and patterns in large groups of races, rather than trying to divine the future based on the few races of which one has "personal experience."

Has it been proven that these "studies of trends and patterns in large groups of races" have the PREDICTIVE effect that they pretend to have?

[traynor]

AI is interesting, but not magic. It can "learn" based on what it is fed. It can "project" based on what has happened in the past. Much the same as people do, but over more data, faster.

[Frost king]

Personally I find the Trifecta to be a bad bet. To me, it should pay out 10x or more than the exacta bet, and that does not happen to often. If you are going to spend $30-50 trying to catch a bottom horse, you are better off putting that same money into the exacta and trying to catch that. A perfect example was at GP with Uncle Runt. He paid $210.20 to win. The $2 exacta was $431.40, and the tri came back $1288.70. So it was only 3x the exacta. Check out the payouts, and the tri very rarely tops the exacta by 10x. So if you took all the time and money and played exactas, you would be further ahead, unless you have the uncanny ability to hit the thing straight up.

[traynor]

Has it been proven that these "studies of trends and patterns in large groups of races" have the PREDICTIVE effect that they pretend to have?

Of course. Basic data mining processes and procedures. Training sets, control sets, data cleaning, tossing outliers. Not rocket science. Basic stuff, that almost everyone "studying" horse races avoids like the plague, while busily chasing rainbows that don't exist.

[thaskalos]

Personally I find the Trifecta to be a bad bet. To me, it should pay out 10x or more than the exacta bet, and that does not happen to often. If you are going to spend $30-50 trying to catch a bottom horse, you are better off putting that same money into the exacta and trying to catch that. A perfect example was at GP with Uncle Runt. He paid $210.20 to win. The $2 exacta was $431.40, and the tri came back $1288.70. So it was only 3x the exacta. Check out the payouts, and the tri very rarely tops the exacta by 10x. So if you took all the time and money and played exactas, you would be further ahead, unless you have the uncanny ability to hit the thing straight up.

Why should the trifecta always pay 10X more than the exacta? Does the exacta always pay 10X more than the win-bet?

[thaskalos]

Of course. Basic data mining processes and procedures. Training sets, control sets, data cleaning, tossing outliers. Not rocket science. Basic stuff, that almost everyone "studying" horse races avoids like the plague, while busily chasing rainbows that don't exist.

And yet...I have seen you post selections here which have been based on such "proven trends"...but the "trends" REVERSED soon thereafter.

[traynor]

Without the winning probabilities how would someone know if the public is making an error?

For starters, the complexity of determining an accurate win probability for each horse in a given race in combination with all the other specific horses in that specific race is huge. Most "winning probabilities" are based on individual horses, not on that horse in the specific combination with the other horses in that race.

By "winning probability" I assume you are referring to the (relatively exact) chance that Horse A will win the race, given all the various factors and attributes of all the other entries in that race. Nothing to do with morning lines, odds, or other such.

[AndyC]

For starters, the complexity of determining an accurate win probability for each horse in a given race in combination with all the other specific horses in that specific race is huge. Most "winning probabilities" are based on individual horses, not on that horse in the specific combination with the other horses in that race.

By "winning probability" I assume you are referring to the (relatively exact) chance that Horse A will win the race, given all the various factors and attributes of all the other entries in that race. Nothing to do with morning lines, odds, or other such.

No doubt computing the probability of horse A winning and horse B placing is complex. My comment was in response to a suggestion that it would be wiser to focus on where the bettors make errors. Doesn't there need to be some way of measuring whether or not there is an error?

[traynor]

And yet...I have seen you post selections here which have been based on such "proven trends"...but the "trends" REVERSED soon thereafter.

Abso-bleeping-lutely. And I have spent most of this year (to date) re-writing my data mining and data analysis applications to avoid such anomalies in the future. It has been an interesting and illuminating experience. There is a bit more involved than clicking a few buttons on my handy-dandy software app and sipping coffee while I await the laundry list of today's winners (in advance).

I will resume that ill-fated endeavor in the near future, and intend to overcome the -$700 or so that was evident in the last posting (specifically, the "easy money" that I had enjoyed for many months by boxing exactas at Yonkers).

Trends do that sort of thing. The trick is to recognize such, mercilessly milk the upswings and neatly sidestep the downswings.

I find no difficulty whatsoever in the situation with boxed exactas at Yonkers, other than that I thought I was doing a generic favor by posting what I had found to be quite profitable bets. The worm turned. I stopped betting on box exactas at Yonkers. As I have indicated numerous times, I have "flags" built-in to my software that continually monitor results and indicate when they are not as I expect them to be. When that happens, I stop making that particular bet, and either sit out a downswing or search for some anomaly that explains it. Or both.

Not rocket science. Pretty basic data mining/data analysis.

[traynor]

No doubt computing the probability of horse A winning and horse B placing is complex. My comment was in response to a suggestion that it would be wiser to focus on where the bettors make errors. Doesn't there need to be some way of measuring whether or not there is an error?

Prior mutuel payoffs. Some factors/attributes tend to be overbet.

[thaskalos]

Abso-bleeping-lutely. And I have spent most of this year (to date) re-writing my data mining and data analysis applications to avoid such anomalies in the future. It has been an interesting and illuminating experience. There is a bit more involved than clicking a few buttons on my handy-dandy software app and sipping coffee while I await the laundry list of today's winners (in advance).

I will resume that ill-fated endeavor in the near future, and intend to overcome the -$700 or so that was evident in the last posting (specifically, the "easy money" that I had enjoyed for many months by boxing exactas at Yonkers).

Trends do that sort of thing. The trick is to recognize such, mercilessly milk the upswings and neatly sidestep the downswings.

I find no difficulty whatsoever in the situation with boxed exactas at Yonkers, other than that I thought I was doing a generic favor by posting what I had found to be quite profitable bets. The worm turned. I stopped betting on box exactas at Yonkers. As I have indicated numerous times, I have "flags" built-in to my software that continually monitor results and indicate when they are not as I expect them to be. When that happens, I stop making that particular bet, and either sit out a downswing or search for some anomaly that explains it. Or both.

Not rocket science. Pretty basic data mining/data analysis.

Now we agree. :ThmbUp:

But your post #29 sounded a bit more "certain" of the sustainability of "trends".

[traynor]

Now we agree. :ThmbUp:

But your post #29 sounded a bit more "certain" of the sustainability of "trends".

Trends are only broad brush strokes of tendencies--not guarantees of individual results. I apologize if I gave some other impression.

[AndyC]

Prior mutuel payoffs. Some factors/attributes tend to be overbet.

Maybe I am dense. But when I try to determine overbet or underbet I have a target payoff. Win bet should pay $5 but odds are 3-1. Exacta should pay $30 but comes back $40.

When someone says overbet, I ask how should it have been bet? Once a target is determined I then look for factors/attributes that lead to under or over betting.

[DeltaLover]

.
My understanding of what you are writing (which may be totally wrong, and if so, I apologize) is that you believe it is more useful (or at least as useful) to study what the bettors are likely to do as a result of seeing certain attributes/factors in a given group of horses as a predictor than to study (only) the certain attributes/factors as predictors.


Exactly this is what I am trying to express with my limited writing skills!

[DeltaLover]

Yes...but is the crowd exhibiting the same wagering "characteristics" across different time periods...or does the crowd "get smarter" as time goes on...thus gradually negating whatever "advantage" we might think we have over the game? If it is indeed "getting tougher to win in this game", as most of us seem to think, then the assumption must be made that any predetermined advantage that we might currently have can hardly be used as an "expectation" of things to come.


Definitely the crowd changes its selection criteria albeit not so quickly as to make it impossible to detect its direction.

I am not sure that the game becomes tougher as time goes by. What I believe it is happening the crowd is changing its 101 handicapping but still tends to overbet the obvious, so the value is still there.

[DeltaLover]

Without the winning probabilities how would someone know if the public is making an error?

This is exactly the point I am trying to say. We should not care about winning probability if somehow we know that the crowd will abuse a specific handicapping factor and systematically lower the odds of a matching horse.

[traynor]

Maybe I am dense. But when I try to determine overbet or underbet I have a target payoff. Win bet should pay $5 but odds are 3-1. Exacta should pay $30 but comes back $40.

When someone says overbet, I ask how should it have been bet? Once a target is determined I then look for factors/attributes that lead to under or over betting.

How do you set the target? If you look at a large number of races (that have a specific combination of factors) and the average/mean/typical exacta in those races is $30, the target is $30. The target is set by the attributes/factors of the horses involved (and the bettors perception of those attributes/factors). This set of attributes/factors in combination typically pays $30.

"Overbet" would be when (historically, over many races) if you bet all the matches, you would have lost buckets of money. "Underbet" would be when (historically, over many races) if you bet all the matches, you would have won buckets of money.

[DeltaLover]

How do you set the target? If you look at a large number of races (that have a specific combination of factors) and the average/mean/typical exacta in those races is $30, the target is $30. The target is set by the attributes/factors of the horses involved (and the bettors perception of those attributes/factors). This set of attributes/factors in combination typically pays $30.

"Overbet" would be when (historically, over many races) if you bet all the matches, you would have lost buckets of money. "Underbet" would be when (historically, over many races) if you bet all the matches, you would have won buckets of money.

Exactly! Of course I am not sure that making this kind of statements in a public forum is a smart move :confused:

[thaskalos]

Exactly! Of course I am not sure that making this kind of statements in a public forum is a smart move :confused:

Not to worry. Such comments have been made publicly for decades...without any side-effect.

[AndyC]

How do you set the target? If you look at a large number of races (that have a specific combination of factors) and the average/mean/typical exacta in those races is $30, the target is $30. The target is set by the attributes/factors of the horses involved (and the bettors perception of those attributes/factors). This set of attributes/factors in combination typically pays $30.

"Overbet" would be when (historically, over many races) if you bet all the matches, you would have lost buckets of money. "Underbet" would be when (historically, over many races) if you bet all the matches, you would have won buckets of money.

OK, the fog has lifted and I understand what you are describing. Thanks.

[DeltaLover]

Not to worry. Such comments have been made publicly for decades...without any side-effect.

For decades? I am not sure I have seen them, any references?

[whodoyoulike]

... A valuable discussion would be how to calculate win and place odds, because the rest of it is mechanical.

And I'd argue that once you have a useful exacta price, you can extrapolate that to trifectas and supers.

I've mentioned this before as a rough guide I've found that the Win payout is approximately and usually 2.x times the Place payout for the horse but you need to consider minimums. For example, A horse at 10/1 would pay $22 +/-, I would expect the place payout about $9 - $11 +/-.

You can verify this with any chart for any track. I don't understand why this occurs since Place calculations considers two horses but I've been aware of this for many years.

[AndyC]

This is exactly the point I am trying to say. We should not care about winning probability if somehow we know that the crowd will abuse a specific handicapping factor and systematically lower the odds of a matching horse.

Gotcha. Thanks.

[whodoyoulike]

... In other words and to paraphrase a common saying: “You can beat the races but you cannot beat the race:

I think you have this backwards.

[HalvOnHorseracing]

I've mentioned this before as a rough guide I've found that the Win payout is approximately and usually 2.x times the Place payout for the horse but you need to consider minimums. For example, A horse at 10/1 would pay $22 +/-, I would expect the place payout about $9 - $11 +/-.

You can verify this with any chart for any track. I don't understand why this occurs since Place calculations considers two horses but I've been aware of this for many years.

You sort of missed what I was saying. I don't care what the place payoff is. About 40 years ago someone told me figure that the place pay on a horse is equivalent to its odds. The number will obviously go up or down based on the amounts bet on the first two finishers.

This issue is the relationship between the first two finishers for the purpose of computing combination probabilities. If you don't have probabilities for Horse A finishing first and Horse B finishing second if Horse A does win, then your basis for computing the exacta pay is at best based on some study of averages based on payoffs associated with respective odds, pool size and number of horses in the race.

What you are looking for is the deviation of your opinion against the crowd opinion. If you can only do that qualitatively, or based on a macro-statistic derived from the public opinion, you limit your ability to exploit superior skill, when you have it.

[DeltaLover]

What you are looking for is the deviation of your opinion against the crowd opinion. If you can only do that qualitatively, or based on a macro-statistic derived from the public opinion, you limit your ability to exploit superior skill, when you have it.

What you need, before you get to the macro-statistic (which by the way is the easier part of the equation) is to build an AI layer to filter the database to discover "similar" races. "Similarity" must be decided upon the most dominant factors used by the crowd, meaning that an unsupervised learning layers is also needed to discover how the crowd is betting; what makes the problem more challenging is that as Thask mentioned already, the crowd changes preferences often and also across different circuits chances are that the factors must be used differently. After you solve this problem, the rest becomes a relatively simple and deterministic procedure with we very high potential for a huge (betting) success.

[MPRanger]

Since there is no trifecta or superfecta grid available, we cannot know.

Barry Meadow shows a Trifecta Estimator on p. 81 of Money Secrets at the Race Track.

Dick Mitchell shows a formula for making break even odds for trifectas in the chapter named The Trifecta starting on p.235 of Common Sense Betting.
Perhaps the very math you seek.

[whodoyoulike]

You sort of missed what I was saying. I don't care what the place payoff is. About 40 years ago someone told me figure that the place pay on a horse is equivalent to its odds. The number will obviously go up or down based on the amounts bet on the first two finishers. ...

Sorry, I thought I was following along just fine which is the reason I only included that portion of your entire post.

That person who told you 40 years ago the approximate Place payoff matches mine except I've never heard or read of it before.

[HalvOnHorseracing]

What you need, before you get to the macro-statistic (which by the way is the easier part of the equation) is to build an AI layer to filter the database to discover "similar" races. "Similarity" must be decided upon the most dominant factors used by the crowd, meaning that an unsupervised learning layers is also needed to discover how the crowd is betting; what makes the problem more challenging is that as Thask mentioned already, the crowd changes preferences often and also across different circuits chances are that the factors must be used differently. After you solve this problem, the rest becomes a relatively simple and deterministic procedure with we very high potential for a huge (betting) success.

The basic question is, if I make this bet (as one of a series of bets) do I have a long term expectation of profit? The generic answer is - yes, if the probable pay represents an overlay.

I would agree it makes sense that if the crowd is overvaluing certain factors and undervaluing others the potential for an overlay exists. I'm less concerned about how you get there than whether you have a reliable (and defensible) methodology. I would agree you could get there with a probabilistic model, although I'm can't be sure (until the model exists) that it is inherently superior to the sharp handicapper applying a more subjective probabilistic model - in essence his own skill, intuitively determining if the crowd has misapplied the dominant factors, or in fact failed to identify the correct dominant factor.

Of course the "skilled handicapper" subset would be very small, making the AI version quite valuable.

[traynor]

The basic question is, if I make this bet (as one of a series of bets) do I have a long term expectation of profit? The generic answer is - yes, if the probable pay represents an overlay.

I would agree it makes sense that if the crowd is overvaluing certain factors and undervaluing others the potential for an overlay exists. I'm less concerned about how you get there than whether you have a reliable (and defensible) methodology. I would agree you could get there with a probabilistic model, although I'm can't be sure (until the model exists) that it is inherently superior to the sharp handicapper applying a more subjective probabilistic model - in essence his own skill, intuitively determining if the crowd has misapplied the dominant factors, or in fact failed to identify the correct dominant factor.

Of course the "skilled handicapper" subset would be very small, making the AI version quite valuable.

It is really, really tough to try to make a profit betting one race at a time, and trying to out-think, out-smart the other bettors in this specific race (as in trying to find overlays based on this specific cluster of mutuel pools). It can be done, people do it, but it is tough.

Much easier is putting the time and effort into building predictive models that--if applied over time--generate a positive return. In that scenario, the sharp handicappers tend to be their own worst enemies--betting vigorously and enthusiastically (or passing the race) for all the wrong reasons.

The "wrong reasons" often include "confounding variables"--stuff that goes along with the model selections when it is right, and are absent when it is wrong, but are not factored in. The effect of those confounding variables is diminished when betting (relatively) large numbers of races according to models. Beating one specific race is a tough way to make a profit. Beating lots of races, over time, is much easier.

[HalvOnHorseracing]

Barry Meadow shows a Trifecta Estimator on p. 81 of Money Secrets at the Race Track.

Dick Mitchell shows a formula for making break even odds for trifectas in the chapter named The Trifecta starting on p.235 of Common Sense Betting.
Perhaps the very math you seek.

Mitchell was doing the same thing we've talked about at some length in the posts about the exacta (and trifecta). Assuming that the probability to win is representative for calculating the probability of a horse finishing in the second and third positions. This has been the traditional model for calculating combination bets.

All Mitchell does is provide the formulaic version of "If the three horses that come in are 2-1, 3-1 and 4-1 to win, the probability of that trifecta is according to the formula and the acceptable pay is ____."

Mitchell's thesis is that the trifecta is a bad bet, and demonstrates that boxing the top four horses in a trifecta is a 12% loser. He continues that if you are going to bet the trifecta, you need to invest in races where you have a false favorite or two.

What he also opines is what DeltaLover and Dave Schwartz have be talking about. Having a win profile for a respective track. And he tangentially references something I've written about - that horses with a low win probability can have a deceptively high show (or fourth place) probability, and that good trifecta payoffs are also dependent on these lower probability horses hitting the board.

Unfortunately, it is not the breakthrough math.

[Dave Schwartz]

Without the winning probabilities how would someone know if the public is making an error?

And how will you know that the public is making an error with the winning probabilities?

[DeltaLover]

And how will you know that the public is making an error with the winning probabilities?

I think Traynor explains it:

It is really, really tough to try to make a profit betting one race at a time, and trying to out-think, out-smart the other bettors in this specific race (as in trying to find overlays based on this specific cluster of mutuel pools). It can be done, people do it, but it is tough.

Much easier is putting the time and effort into building predictive models that--if applied over time--generate a positive return. In that scenario, the sharp handicappers tend to be their own worst enemies--betting vigorously and enthusiastically (or passing the race) for all the wrong reasons.

The "wrong reasons" often include "confounding variables"--stuff that goes along with the model selections when it is right, and are absent when it is wrong, but are not factored in. The effect of those confounding variables is diminished when betting (relatively) large numbers of races according to models. Beating one specific race is a tough way to make a profit. Beating lots of races, over time, is much easier.

http://www.statisticshowto.com/design-of-experiments/confounding-variable/

[HalvOnHorseracing]

It is really, really tough to try to make a profit betting one race at a time, and trying to out-think, out-smart the other bettors in this specific race (as in trying to find overlays based on this specific cluster of mutuel pools). It can be done, people do it, but it is tough.

Much easier is putting the time and effort into building predictive models that--if applied over time--generate a positive return. In that scenario, the sharp handicappers tend to be their own worst enemies--betting vigorously and enthusiastically (or passing the race) for all the wrong reasons.

The "wrong reasons" often include "confounding variables"--stuff that goes along with the model selections when it is right, and are absent when it is wrong, but are not factored in. The effect of those confounding variables is diminished when betting (relatively) large numbers of races according to models. Beating one specific race is a tough way to make a profit. Beating lots of races, over time, is much easier.

I'm not sure I got your point. If you bet, you are betting a specific race, and you are, in fact, trying to outsmart the bettors in that race. And that is true whether you are using a digital probabilistic model or a subjective probabilistic model. The point is that even if you as the sharp handicapper are trying to outsmart the crowd, you are doing it probabilistically, regardless of how mechanistic you were in getting there. You can't come to the proper conclusion in any particular race until you actually see what the crowd does, because obviously it is possible the crowd is getting it right. And you only want to bet the races where you (or your model) believes the crowd is getting it wrong.

You can't be a sharp handicapper (by my definition) if you bet based on the wrong factors.

If I'm reading you right, you believe without the model a bettor is too idiosyncratic to get enough races right to be a winner, or else he sees each event as independent of all other events, in which case he gets misdirected by confounding factors. I certainly don't believe that. The idea of subjective probabilistic modeling is that the bettor is doing what the digital model has done, although admittedly perhaps not as accurately. In other words, they know which variables are relevant to win probability, and they can see if the crowd has missed them.

If the player is betting enthusiastically for all the wrong reasons, you can't use the word sharp in relation to that bettor. You're correct that the more events you have in your model, the less confounding factors will affect the model results, but that's true regardless of the nature of the model.

If you've watched 20,000 races and you learned nothing about where the crowd gets misdirected, that is definitely a problem, especially if you can't apply it to a respective event. But if the point is that you can't overvalue or undervalue a factor because of some inherent bias when you use a digital probabilistic model, I would just say, of course. I'm not a Luddite!

[Dave Schwartz]

Thank you all for the discussion. It has been enlightening.

To be clear, I was not expecting ANSWERS. I asked for, and received, DISCUSSION, and what I got will (I hope) turn out to be worth its weight in gold.

Just like with a prediction market, the wide spectrum of education, experience, and opinions is where the profit is to be found.

Usually, when I need a discussion, I head to the whiteboard with Beth. She does a lot of listening, and every once in a while, offers a suggestion. It was actually her idea which got me to finding the missing piece.

[Dave Schwartz]

I think Traynor explains it:


He did. It was a rhetorical question. The point is that only by looking at a large sample can you validate your probabilities, and even those cannot be validated simply by counting the winners versus expected winners.

[MPRanger]

Mitchell was doing the same thing we've talked about at some length in the posts about the exacta (and trifecta). Assuming that the probability to win is representative for calculating the probability of a horse finishing in the second and third positions. This has been the traditional model for calculating combination bets.

All Mitchell does is provide the formulaic version of "If the three horses that come in are 2-1, 3-1 and 4-1 to win, the probability of that trifecta is according to the formula and the acceptable pay is ____."

Mitchell's thesis is that the trifecta is a bad bet, and demonstrates that boxing the top four horses in a trifecta is a 12% loser. He continues that if you are going to bet the trifecta, you need to invest in races where you have a false favorite or two.

What he also opines is what DeltaLover and Dave Schwartz have be talking about. Having a win profile for a respective track. And he tangentially references something I've written about - that horses with a low win probability can have a deceptively high show (or fourth place) probability, and that good trifecta payoffs are also dependent on these lower probability horses hitting the board.

Unfortunately, it is not the breakthrough math.


Wow, very well developed post. Thanks for your response. Sounds like you're up on it. I think any mathematical model is a static thing. In reality the order of finish is random. (line 'em up. Run 'em again and you'll get a different order of finish). Still, a mathematical model is a useful tool. No?

[traynor]

I'm not sure I got your point. If you bet, you are betting a specific race, and you are, in fact, trying to outsmart the bettors in that race. And that is true whether you are using a digital probabilistic model or a subjective probabilistic model. The point is that even if you as the sharp handicapper are trying to outsmart the crowd, you are doing it probabilistically, regardless of how mechanistic you were in getting there. You can't come to the proper conclusion in any particular race until you actually see what the crowd does, because obviously it is possible the crowd is getting it right. And you only want to bet the races where you (or your model) believes the crowd is getting it wrong.

The individual races are almost irrelevant. Specifically, it is not useful to combine wagering with models and "seeking value." If the models are predictive and relatively accurate, the value will be there at the end of the day (or week, or whatever). It is not useful to pass a race (that your model tells you is part of a profitable pattern) because the tote board numbers are bobbling up and down.



You can't be a sharp handicapper (by my definition) if you bet based on the wrong factors.

If I'm reading you right, you believe without the model a bettor is too idiosyncratic to get enough races right to be a winner, or else he sees each event as independent of all other events, in which case he gets misdirected by confounding factors.

Not exactly. Difficult, not impossible.


I certainly don't believe that. The idea of subjective probabilistic modeling is that the bettor is doing what the digital model has done, although admittedly perhaps not as accurately. In other words, they know which variables are relevant to win probability, and they can see if the crowd has missed them.


I would tend to disagree. That disagreement would be more related to the ability of people to detect patterns based on personal experience, as opposed to building a model from a wider base of races, analyzed objectively. Human experience--in racing as elsewhere--tends to remember successes more strongly than failures. Specifically, the tendency to be swayed by heuristics--that may or may not be appropriately applied in a specific situation.

If the player is betting enthusiastically for all the wrong reasons, you can't use the word sharp in relation to that bettor. You're correct that the more events you have in your model, the less confounding factors will affect the model results, but that's true regardless of the nature of the model.

If you've watched 20,000 races and you learned nothing about where the crowd gets misdirected, that is definitely a problem, especially if you can't apply it to a respective event. But if the point is that you can't overvalue or undervalue a factor because of some inherent bias when you use a digital probabilistic model, I would just say, of course. I'm not a Luddite!

Most of the factors that have been learned after watching 20,000 races have also been learned by many others who also watched 20,000 races. The money is in the stuff they don't see/consider/evaluate or that they see/consider/evaluate inappropriately.

I have no idea what a Luddite, is but I'll look it up.

[AndyC]

....
Mitchell's thesis is that the trifecta is a bad bet, and demonstrates that boxing the top four horses in a trifecta is a 12% loser. He continues that if you are going to bet the trifecta, you need to invest in races where you have a false favorite or two........

You could put the name of any bet at the track in the above statement.

[HalvOnHorseracing]

I have no idea what a Luddite, is but I'll look it up.

A Luddite is a member of any of the bands of English workers who destroyed machinery, especially in cotton and woolen mills, that they believed was threatening their jobs. It's sort of a generic term for people who believe mechanization hurts the individual.

[DeltaLover]

Most of the factors that have been learned after watching 20,000 races have also been learned by many others who also watched 20,000 races. The money is in the stuff they don't see/consider/evaluate or that they see/consider/evaluate inappropriately.


:ThmbUp::ThmbUp:


It cannot be stated any better!

If you do not understand what is said here and continue to believe that the game is about picking winners, you better give it up as you have no chance of beating it in the long run.

[HalvOnHorseracing]

:ThmbUp::ThmbUp:


It cannot be stated any better!

If you do not understand what is said here and continue to believe that the game is about picking winners, you better give it up as you have no chance of beating it in the long run.

You've obviously never read my stuff. I've said on numerous occasions that too many bettors are selection oriented, not value oriented. And I've had articles published on how to identify value. How you could have interpreted what I wrote as simply "picking winners" is another thing that puzzles me. At best, I'd say we're talking past each other.

What I believe neither of you understood is that there is more than one type of probabilistic model, and regardless of the probabilistic model you apply, the trick is still identifying value in any respective event. I don't care what data you input into your Univac. You are looking for races where the crowd made a mistake by mis-assigning a probability to an entrant. And in races where the crowd has not made a mistake there is no bet.

You can beat the hell out of a race, and if you do it often enough, you can beat the races. Again I'm puzzled that anyone would argue with the idea that you beat the races in total by beating enough of the right individual races. The game is not about picking winners, but about betting the right winners. At some point you have to have a strategy for putting money through the windows, and as I said above, I don't care what sort of model you use to identify value as long as you correctly bet value.

What Traynor said,

The individual races are almost irrelevant. Specifically, it is not useful to combine wagering with models and "seeking value." If the models are predictive and relatively accurate, the value will be there at the end of the day (or week, or whatever). It is not useful to pass a race (that your model tells you is part of a profitable pattern) because the tote board numbers are bobbling up and down.


If the subjective model is predictive and relatively accurate, the value is equally there. Obviously I didn't make that clear.

I don't know where anyone came up with the statement that it is not useful to pass a race because the tote board numbers are bobbling up and down. I said, there are races that the crowd gets it right, and those races offer no edge, no value. Those races are a pass for me. One wants to bet races where the crowd hasn't gotten it right, and I really don't care how you come to that conclusion. The idea that you would bet those races because they are part of some pattern remains a head scratcher to me.

On the day it is not useful to determine value and bet it, I'm out of the game. Every successful gambler knows that winning is dependent on identifying his edge and investing the correct amount, and that's true whether you are playing poker, blackjack or the horses. You ever hear the old saw, never draw to an inside straight? Well if your probability of catching your card is 14% and your call would get you a return of 30%, the bet is value. Thorp's book on blackjack is totally about value. The theme of the book is that if you can identify when you have an edge and bet accordingly you can beat the game in the long run. But it still comes down to betting a series of individual events. If you and Traynor are saying something different than this, then we'll simply have to disagree.

[DeltaLover]

I don't know where anyone came up with the statement that it is not useful to pass a race because the tote board numbers are bobbling up and down. I said, there are races that the crowd gets it right, and those races offer no edge, no value. Those races are a pass for me. One wants to bet races where the crowd hasn't gotten it right, and I really don't care how you come to that conclusion. The idea that you would bet those races because they are part of some pattern remains a head scratcher to me.*


This is the core of our disagreement! What I am saying is that you have no way to know whether the crowd has it right or wrong in any individual race; instead you should try to discover “value” in a large set of races following a generic approach and without relying on the tote as an indicator of over / under values.

I will repeat that:

(1) It is impossible to estimate the winning probabilities of a specific horse with sufficient accuracy.

(2) Even if you had some “magic spell” to discover the “real” probabilities, this would not allow you to win in the long run, for the simple reason that you do not know your payoff in advance.

Instead:

You need to discover systematic errors in the crowd's opinion and constantly bet against them without paying attention to the price of the individual race. Note that, I do not imply that the errors you are going to bet against are always the same, as they change through time and from circuit to circuit but they are always discoverable and developing the required procedures consist the main challenge for the bettor.



Thorp's book on blackjack is totally about value. The theme of the book is that if you can identify when you have an edge and bet accordingly you can beat the game in the long run. But it still comes down to betting a series of individual events. If you and Traynor are saying something different than this, then we'll simply have to disagree.


The difference is that, in BJ you have all the necessary input to derive the mathematically correct move for every specific hand you play while the same does not apply to horse racing where payoffs are decided by human opinions and also the related probabilities are impossible to define analytically.

[Poindexter]

Delta, can you provide 1 example (don't give away the keys to the mint)of where the crowd continues to make a mistake (either overlooking 1 variable or overbetting another). Also what is your evidence.

I really have a hard time with your and Traynors thinking that making a somewhat accurate oddsline is near impossible, yet somehow there are some magic bullets you can find pouring over the data of thousands of races that will continue to overcome the enormous takeout. I believe the opposite.

[DeltaLover]

Delta, can you provide 1 example (don't give away the keys to the mint)of where the crowd continues to make a mistake (either overlooking 1 variable or overbetting another). Also what is your evidence.

I really have a hard time with your and Traynors thinking that making a somewhat accurate oddsline is near impossible, yet somehow there are some magic bullets you can find pouring over the data of thousands of races that will continue to overcome the enormous takeout. I believe the opposite.

A "toy" example that is indicative of how to discover systematic crowd errors:

Claiming races, SANTA ANITA, 6.5 f TURF, FINAL ODDS <= 6.0, ROI BY POST POSITION:


http://www.paceadvantage.com/forum/attachment.php?attachmentid=20118&stc=1&d=1495049624

What this statistic suggests it that it is possible that the crowd overbets the two inner post positions shifting some value to the outers.

This info alone will not allow you to show any profit of course but an automated system with the ability to discover hundreds of similar betting anomalies and combine them in many ways has the potential to discover a profitable macro-handicapping strategy.

[Poindexter]

A "toy" example that is indicative of how to discover systematic crowd errors:

Claiming races, SANTA ANITA, 6.5 f TURF, FINAL ODDS <= 6.0, ROI BY POST POSITION:


http://www.paceadvantage.com/forum/attachment.php?attachmentid=20118&stc=1&d=1495049624

What this statistic suggests it that it is possible that the crowd overbets the two inner post positions shifting some value to the outers.

This info alone will not allow you to show any profit of course but an automated system with the ability to discover hundreds of similar betting anomalies and combine them in many ways has the potential to discover a profitable macro-handicapping strategy.


Or it tells you that in a whopping 36 races that the horses that ran in posts 1 and 2 at 6-1 or less just happened have a poor roi. Maybe in the next 36 races they will have the best roi. I thought you were dealing with thousands of races.

[DeltaLover]

Or it tells you that in a whopping 36 races that the horses that ran in posts 1 and 2 at 6-1 or less just happened have a poor roi. Maybe in the next 36 races they will have the best roi. I thought you were dealing with thousands of races.

Sure, what you saying is very possible and the specific example can very well be a random anomaly that will slowly regress to the mean. In racing (as in science or anything kind of empirical domain) induction does not prove anything but serves as an indicator of the validity of a proposition.

As far as the size of the db, I keep a universe or approximately 200K races in my active data set but this does not mean that I am trying to analyze all of them at once. In contrary I believe that betting anomalies can change based on circuit and time and my systems are taking this is consideration trying to discover how the crowd is changing its opinions thus my active data sets can very well become so small as the one I presented, assuming of course that it was selected from a very large and complete universe of races as a potential point of betting irregularity.

[Poindexter]

Let's go back to the data you provided about post 1 and 2 in 6 1/2 furlongs. Unless it has something to do with a bias then you are dealing with something about the course that punishes the inside post position. But why would you expect that to change?. In other words if posts 1-2 are a detriment down the hill in 2017 they should have been a detriment in 2016 and 2014 and 2013.........So imo, wouldn't back data shed light on whether the recent poor roi means anything or not.

I would love to have a database like you have and find negative factors and positive factors but they are only pieces of the puzzle to make a more informed determination of each horses chances. Just like a workout report might be, or paddock inspection might be or trip notes are or bias notes are.... Ultimately you factor that information into determining each horses chance of winning. But in the end you have to have an idea how much chance each horse if you want to have a chance of winning. No? At least that is the way I see it.

[DeltaLover]

Let's go back to the data you provided about post 1 and 2 in 6 1/2 furlongs. Unless it has something to do with a bias then you are dealing with something about the course that punishes the inside post position. But why would you expect that to change?. In other words if posts 1-2 are a detriment down the hill in 2017 they should have been a detriment in 2016 and 2014 and 2013.........So imo, wouldn't back data shed light on whether the recent poor roi means anything or not.

I would love to have a database like you have and find negative factors and positive factors but they are only pieces of the puzzle to make a more informed determination of each horses chances. Just like a workout report might be, or paddock inspection might be or trip notes are or bias notes are.... Ultimately you factor that information into determining each horses chance of winning. But in the end you have to have an idea how much chance each horse if you want to have a chance of winning. No? At least that is the way I see it.

What might change (and what I really care about as a bettor) is the way the crowd is perceiving the chances of each of the 101 handicapping factors.

A factor that is completely neutral can very well be a gold mine assuming that the crowd believes in something that is very wrong.

AGAIN: It is not the winning probability of a horse that matters! What matters is the probability of the crowd to miscalculate the probability(!) of the horse.

[Poindexter]

So while most horse players are capping horses, you are capping the betting public. You feel that they are consistent enough in their actions that they will continue with certain betting biases that will open up a world of betting opportunity for you. I don't see it but whatever works for you.

But I am a little confused. Don't you have home grown variants and such. Why would you bother with such if all you are looking for is where the public continually makes mistakes. What does one have to do with the other.

[HalvOnHorseracing]

What might change (and what I really care about as a bettor) is the way the crowd is perceiving the chances of each of the 101 handicapping factors.

A factor that is completely neutral can very well be a gold mine assuming that the crowd believes in something that is very wrong.

AGAIN: It is not the winning probability of a horse that matters! What matters is the probability of the crowd to miscalculate the probability(!) of the horse.
I'll say two things. One, it is not impossible to calculate the probabiiity of a respective horse winning a race. Two, nobody, not even the machine, gets it right most of the time, but it isn't necessary to get it right most of the time. If you bet a horse at 5-1 and you hit 1 in 4 races, long term you are a winner. THAT is how you make profit. It is only necessary to get it right enough to overcome the times you get it wrong, and in forty years of keeping personal records, I can guarantee it is possible.

If anyone had a program that was right most of the time, they could sell it for prices most of us couldn't afford and and retire in a year.

Stunningly, saying what matters is the probability of the crowd to miscalculate the probability of the horse is no different than saying you bet when the crowd makes a mistake and you have an edge. And calculating the amount of your edge with data is no different than applying subjective probability to calculate the amount of the edge. No, it is not the winning probability of the horse that matters, and I would never suggest it is. It is the difference between the winning probability of the horse and the crowd assigned probability that is the critical metric, and you can't discern that without knowing both probabilities.

[DeltaLover]

So while most horse players are capping horses, you are capping the betting public. You feel that they are consistent enough in their actions that they will continue with certain betting biases that will open up a world of betting opportunity for you. I don't see it but whatever works for you.

But I am a little confused. Don't you have home grown variants and such. Why would you bother with such if all you are looking for is where the public continually makes mistakes. What does one have to do with the other.

Correct. I am using the term meta-handicapping to describe what you (successfully) are referring as "capping the betting public".

As far as my own metrics, track variants, figures etc. , remember Traynor's posting:

The money is in the stuff they don't see/consider/evaluate or that they see/consider/evaluate inappropriately.


so the custom "numbers" stand for stuff they don't see/consider/evaluate while the meta-handicapping for what hat they see/consider/evaluate inappropriately.

[Poindexter]

Correct. I am using the term meta-handicapping to describe what you (successfully) are referring as "capping the betting public".

As far as my own metrics, track variants, figures etc. , remember Traynor's posting:



so the custom "numbers" stand for stuff they don't see/consider/evaluate while the meta-handicapping for what hat they see/consider/evaluate inappropriately.

What you say makes sense. I get the general gist of what you are doing. How you put it all together into betting decisions seems complex. But lets go with hypothetical example. Let's say that there are 3 main contenders

1)8/5

2) 3-1

4 6-1


The recent Beyers look like this

1) 90, 89, 84

2) 87, 85, 89

3) 81, 76, 78


Now there Deltas(your numbers on the same scale)

1) 84, 88, 86
2) 89, 90, 85
3) 87, 86, 87

Now anyone doing their own speed figures as you do is going to look at the tote board and see the public is betting the race consistent with the Beyers, but their own speed figures tell a little different story and they will focus their energies on the 2 and 3 who clearly offer better value. Now this is obviously simplistic (we have to look at the other handicapping variables), but would you not do the same and likely focus your betting on the 2 and 3 over the 1? If not why not (assuming that the other handicapping factors didn't sway you any other direction)? If so, then you are just seeking value no different than had you created an oddsline and looked for value. You are just finding it another way.

[DeltaLover]

What you say makes sense. I get the general gist of what you are doing. How you put it all together into betting decisions seems complex. But lets go with hypothetical example. Let's say that there are 3 main contenders

1)8/5

2) 3-1

4 6-1


The recent Beyers look like this

1) 90, 89, 84

2) 87, 85, 89

3) 81, 76, 78


Now there Deltas(your numbers on the same scale)

1) 84, 88, 86
2) 89, 90, 85
3) 87, 86, 87

Now anyone doing their own speed figures as you do is going to look at the tote board and see the public is betting the race consistent with the Beyers, but their own speed figures tell a little different story and they will focus their energies on the 2 and 3 who clearly offer better value. Now this is obviously simplistic (we have to look at the other handicapping variables), but would you not do the same and likely focus your betting on the 2 and 3 over the 1? If not why not (assuming that the other handicapping factors didn't sway you any other direction)? If so, then you are just seeking value no different than had you created an oddsline and looked for value. You are just finding it another way.

Instead of relying on my intuition, I will feed my model with a representation of the past performances (using Deltas, fractional positions and some other data) to a black-box (for example a NN) that will decide which horse to bet. The challenges are:
(1) How to create contrarian (enough) Deltas.
(2) How to decide about the combination of data to create the input.
(3) How to build and train the black-box.

[ReplayRandall]

(2) How to decide about the combination of data to create the input.
How many years have you been working on this? Guessing at least 3 or more?
.D
.I
.F
.F
.I
.C
.U
.L
.T
.Y
(3) The Black Box must seem like light-years away

[DeltaLover]

How many years have you been working on this? Guessing at least 3 or more?
.D
.I
.F
.F
.I
.C
.U
.L
.T
.Y
(3) The Black Box must seem like light-years away

Many years of piecemeal progress!

[ReplayRandall]

Many years of piecemeal progress!
It's good to see you posting again like this, I enjoy what you share...:ThmbUp:

[DeltaLover]

It's good to see you posting again like this, I enjoy what you share...:ThmbUp:

I am always posting, assuming that there is some topic that I am interested and have something to say about!

[ReplayRandall]

I am always posting, assuming that there is some topic that I am interested and have something to say about!
Well then, just post under DeltaLover ONLY from now on......:lol:

[DeltaLover]

Well then, just post under DeltaLover ONLY from now on......:lol:

:p

[PaceMaker]

So I have been tracking tracks by day and scoring the races that pay say over $60 on the Tri and give it a score of .5 and over $100 and give it a score of 1 and so on. You begin to see a pattern of tracks and days that pay better.

Then I try to figure out if the race for today may pay out more with certain factors:
# horses
Lowest MLO
Ave MLO
% of MLO over 8 to 1
the speed difference between horses ( less is better)

I weight each of these and come up with a score and if over a certain number it has a better chance of paying off better
But not always!And sometimes a 6 horse race pays out a lot!

Any other thoughts would be apprciated

[whodoyoulike]

So I have been tracking tracks by day and scoring the races that pay say over $60 on the Tri and give it a score of .5 and over $100 and give it a score of 1 and so on. You begin to see a pattern of tracks and days that pay better.

Then I try to figure out if the race for today may pay out more with certain factors:
# horses
Lowest MLO
Ave MLO
% of MLO over 8 to 1
the speed difference between horses ( less is better)

I weight each of these and come up with a score and if over a certain number it has a better chance of paying off better
But not always!And sometimes a 6 horse race pays out a lot!

Any other thoughts would be apprciated

Have you been keeping track of the trifecta pools at each of these tracks on a daily basis by race?

Do they really vary, which is what you're implying to me from your pattern comment?

The only pattern I've noticed is the tri pool gets larger as the day goes along until the feature then declines. Actually, I think it's the same pattern with the other pools.

[traynor]

So I have been tracking tracks by day and scoring the races that pay say over $60 on the Tri and give it a score of .5 and over $100 and give it a score of 1 and so on. You begin to see a pattern of tracks and days that pay better.

Then I try to figure out if the race for today may pay out more with certain factors:
# horses
Lowest MLO
Ave MLO
% of MLO over 8 to 1
the speed difference between horses ( less is better)

I weight each of these and come up with a score and if over a certain number it has a better chance of paying off better
But not always!And sometimes a 6 horse race pays out a lot!

Any other thoughts would be apprciated

You might consider:
Grouping tris by size is great, if you truncate outliers. If you have Pattern A that is for tris in the 50-80 range and a tri that fits Pattern A pays 200, it belongs in the Pattern A pile, not the 200 and up pile. Specifically, the mutuels should not be the determining factor of categorizing. That should be the pattern (however you define it).

MLO are flaky, at best, especially if you are studying multiple tracks. They are (supposedly) a prediction of how the crowd will bet, not the probabilities of the horse winning. Grouping by final odds (in ranges) is more accurate.

[PICSIX]

So while most horse players are capping horses, you are capping the betting public. You feel that they are consistent enough in their actions that they will continue with certain betting biases that will open up a world of betting opportunity for you. I don't see it but whatever works for you.

But I am a little confused. Don't you have home grown variants and such. Why would you bother with such if all you are looking for is where the public continually makes mistakes. What does one have to do with the other.

I think Michael Pizzola says it best:

We should bet on horses that we like that the Public should not like.

To do this we need a way/method to handicap the handicappers.

[traynor]

I think Michael Pizzola says it best:

We should bet on horses that we like that the Public should not like.

To do this we need a way/method to handicap the handicappers.

As with most things in life, it is faster, easier, and simpler to take shortcuts. It is pretty simple to find the "bad stuff"--pointers to factors/attributes coincident with losers. It is pretty simple to locate those which will be apparent to most bettors (for example, a trainer with very low win percentage), and cause them to be reluctant to bet on that particular horse, or to "eliminate" that horse from most of their wagers.

Then (and ONLY then) it becomes useful to study the factors/attributes coincident with winners (or keys for tris and supers). In particular the "overlapping areas" (like a Venn diagram)--instances in which the other factor/attributes coincident with winners trump the (possible) existence of (one or more, with more being better) factors/attributes considered "negative" by most bettors.

In that subset of race scenarios--relatively simple to locate with a decent data mining app and a bit of competent data analysis--one has BOTH a key with solid win probability AND a key that many (if not most) bettors will be reluctant to bet.

Which is all a long-winded way to say what you wrote above, citing Mikey.
"We should bet on horses that we like that the Public should not like."

[Poindexter]

I think Michael Pizzola says it best:

We should bet on horses that we like that the Public should not like.

To do this we need a way/method to handicap the handicappers.

I agree 100% with this quote The best bets are horse we like for reasons that, the public doesn't latch onto, whether it is bias, trip, trainer, pace,projected development, profile, distance/surface switches...................I also think this amplifies in the pick 4's/pick 5's especially when they come in the later legs. It is much harder to make money when everyone sees the same thing we do.

I would say most experienced horseplayers have a pretty good pulse as to what the public is probably going to do-and even if we can't project it we certainly can recognize what seems reasonable and what seems out of the ordinary. What I find curious isn't that Delta Lover doesn't want to use his database to make a better line (better estimate of a horses probability-since he has a better understanding of the variables that work and do not work over thousands of races), he want to use it to choose horses. My point is even if 10,000 races tell him that the public went overboard on "Beyers" for instance, that is likely not enough to overcome the take, and no law says it will repeat again to the same degree over the next 10,000 races. I am not sayihg he is wrong to do what he does (he obviously is very intelligent and puts a lot of thought into what he does and perhaps for a very good reason-same obviously applies to traynor as well)., I just would never choose to go that route. Different strokes for different folks as they used to say back in the day.

The way I see racing is that every race is a menu of horses with a menu of variables(positibe and negative) that need to be weighed to assess the likelihood of each horse winning (whether you create an oddsline or not-the oddline is just a tool). Then you need to factor in the toteboard(who is live, who is overbet, who appears dead and does it matter). Then you have to make a betting decision. Then it is on to the next race. This is the game I play.


I believe every facet of the above can be improved upon until the game becomes profitable to you the player. If that never happens, well you had a lot of fun on the way and you have a lot of company. If you turn the corner then you rise through the ranks of sea creatures until you become a whale yourself.