What would you want to have done...

[TrifectaMike]

If you had access to a mathematician, an expert horse player, the finances, the desire, and commitment, what type of horse racing selection process would you like to develop?

For example, would you want a Benter Logistic or Probit approach, or a Bayesian Belief Network approach, Jcapper Approach, an actual simulation of the races, an expert system, a genetic algorithm, neural network, etc.

Many may say Benter's approach because of the success in Hong Kong racing, but I'd like to hear your ideas. At some point in this thread I'll give my approach.

Mike (Dr. Beav)

[sjk]

Many years ago there was lots of interest in talking about how to develop handicapping software. These days very few on the board seem to be on that path.

[TrifectaMike]

Many years ago there was lots of interest in talking about how to develop handicapping software. These days very few on the board seem to be on that path.

If that is true, it is sad.

About a year ago or maybe two years ago, Jeff Platt wrote something in a thread about seeing a professor to get a better understanding on a mathematical concept. At the time, I was shocked that he had considered the approach. It was one of the wisest things I've read on this board.

I don't know if he went any further than just being curious. I'm sure he knows what I am speaking of. If he wants to comment, that is his choice. I will not.

Mike (Dr Beav)

[BIG49010]

I did some programing for a guy, about 10 years ago, that was trying to make the Neural network ideas work. He just couldn't seem to make it work, so it will be interesting to hear your comments.

[Freightliner]

I am working with a statistician and we are working on the theory of really big numbers.

[Cratos]

If you had access to a mathematician, an expert horse player, the finances, the desire, and commitment, what type of horse racing selection process would you like to develop?

For example, would you want a Benter Logistic or Probit approach, or a Bayesian Belief Network approach, Jcapper Approach, an actual simulation of the races, an expert system, a genetic algorithm, neural network, etc.

Many may say Benter's approach because of the success in Hong Kong racing, but I'd like to hear your ideas. At some point in this thread I'll give my approach.

Mike (Dr. Beav)

None of the approaches are attractive to me, but before I give my reason why in a later post I will suggest that you have left out the most important brain, the physicist from your think tank.

[Stillriledup]

If you had access to a mathematician, an expert horse player, the finances, the desire, and commitment, what type of horse racing selection process would you like to develop?

For example, would you want a Benter Logistic or Probit approach, or a Bayesian Belief Network approach, Jcapper Approach, an actual simulation of the races, an expert system, a genetic algorithm, neural network, etc.

Many may say Benter's approach because of the success in Hong Kong racing, but I'd like to hear your ideas. At some point in this thread I'll give my approach.

Mike (Dr. Beav)

I have everything except the computer nerd to write me a pick 6 program so i can efficiently invest into america's greatest bet.

[thaskalos]

If you had access to a mathematician, an expert horse player, the finances, the desire, and commitment, what type of horse racing selection process would you like to develop?


I would request that they spare no expense in order to create a way of constructing an accurate odds line...which would, at last, transform the gambler into an investor...

The late Dick Mitchell - ALSO a mathematician and expert horse player - promised me as much when he sold me a SHARP programmable pocket calculator for $399...25 years ago...but, regrettably, he overestimated the device's effectiveness.

[Freightliner]

The late Dick Mitchell - ALSO a mathematician and expert horse player - promised me as much when he sold me a SHARP programmable pocket calculator for $399...25 years ago...but, regrettably, he overestimated the device's effectiveness.
Does this mean I'm screwed?

[Greyfox]

IThe late Dick Mitchell - ALSO a mathematician and expert horse player - promised me as much when he sold me a SHARP programmable pocket calculator for $399...25 years ago...but, regrettably, he overestimated the device's effectiveness.

Mitchell knew his math but he wasn't a mathetmatician.
Regrettably, you overestimated the device's effectiveness when you purchased it, for $399 ...25 years ago...which would be a lot more today.

[thaskalos]

Mitchell knew his math but he wasn't a mathetmatician.
Regrettably, you overestimated the device's effectiveness when you purchased it, for $399 ...25 years ago...which would be a lot more today. Greyfox...

If you are going to pick a fight with me...at least make the effort to get your facts straight.

I would like to refer you to the back of Mitchell's book "Commonsense Handicapping"...where the famed handicapping author James Quinn has posted the following quote:

"Only a decade ago, Dick Mitchell was a better MATHEMATICIAN and computer scientist than he was a handicapper. Now he is better-than-good at all three."

It is common knowledge that Dick Michell was indeed a mathematician...and he returned to teaching math toward the end of his life...after his "retirement" from thoroughbred handicapping.

[Greyfox]

Greyfox...

If you are going to pick a fight with me...at least make the effort to get your facts straight.

.

Geez. Who's picking a fight with you?? :rolleyes: Commenting on what you are saying is not "picking a fight."
I thought Mitchell's expertise was Economics.
You are correct. I was wrong. Here is his obituary:
http://www.cynthiapublishing.com/dickmitchell.html

Never the less, you missed the main point I was making.
You thaskalos overestimated the device's effectiveness and paid $399. I wasn't wrong about that.

[PhantomOnTour]

What I really would like is someone with a keen eye in judging t'breds in the flesh to accompany me to the paddock before each race. Just about any trainer will do (and I mean ANY trainer...sat next to one of the worst trainers I've ever seen at FG for a few years. He couldn't train 'em, but he could see which horse was off or sore just by watching them in the paddock. Good info imo)

[thelyingthief]

I am working with a statistician and we are working on the theory of really big numbers.

wow, I didn't know that really big numbers was a theory. In fact, I can spin off a really big number without theorizing whatsoever.

fer instance: 12312515565756351656998796546161549984977700078411 11000_
00000000000000000000000000000000000000000000000000 0000000000000000_
0000000000000000000000006468497897979.

This is a considerably large number. I could have made it larger.

tlt-

[Stillriledup]

What I really would like is someone with a keen eye in judging t'breds in the flesh to accompany me to the paddock before each race. Just about any trainer will do (and I mean ANY trainer...sat next to one of the worst trainers I've ever seen at FG for a few years. He couldn't train 'em, but he could see which horse was off or sore just by watching them in the paddock. Good info imo)

The actual warmup on the track will tell you more than a paddock inspection. At least it does for me. If a horse is sore, he won't 'unravel' in the paddock as much as he might unravel with 115 lbs on his back and being aggressively warmed up. Once that jock gets aboard and starts to warmup the horse, you might be able to tell more than if the horse is just standing there calmly being held by a shank.

[sjk]

Your individual sounds a lot like me 15 years ago. I agree with the others that the goal needs to be an actionable odds line which amounts to a set of probablitites.

A probit model sounds like something that would be helpful in developing a set of probabilities. I don't know that it is necessary to know the details of the methods you enumerated or to stick to a single method so long as the end product is an actionable odds line.

[fast4522]

Very interesting post that should spark many follow up posts. For my take on such software would be that it have several aspects including a good odds line that is better than out of the can morning line. There should be a limit of five people using it, what value would there be if everyone has the same data. Overall the majority of winners should park inside the top four rankings on a card. The software should crunch all the good cards on any given day when you start the program. The program should have a very rich in depth modeling export to Microsoft Access. And data should not cost $130 per month period.

[TrifectaMike]

Your individual sounds a lot like me 15 years ago. I agree with the others that the goal needs to be an actionable odds line which amounts to a set of probablitites.

A probit model sounds like something that would be helpful in developing a set of probabilities. I don't know that it is necessary to know the details of the methods you enumerated or to stick to a single method so long as the end product is an actionable odds line.

I agree with your assertion of an actionable odds line. I don't believe logistic or probit model probabilities will solve the problem. It can be part of a solution, but not the solution.

Benter stumbled on the solution by trial and error, but not necessarily a complete solution. He actually solved the problem in reverse order.

Mike (Dr Beav)

[gm10]

I agree with your assertion of an actionable odds line. I don't believe logistic or probit model probabilities will solve the problem. It can be part of a solution, but not the solution.

Benter stumbled on the solution by trial and error, but not necessarily a complete solution. He actually solved the problem in reverse order.

Mike (Dr Beav)

What does 'actionable' mean in this context?

[sjk]

I used the word actionable to mean that you have a set of rules that tells you which horses to play:e.g. "bet any horse which has run in the past 120 days and whose odds offer a 2x overlay based on the odds line".

It is implied that you get a significant number of plays using the rule and that over the long haul these generate a profit.

[TrifectaMike]

What does 'actionable' mean in this context?

An oddsline which is well calibrated as the tote odds, but sharper.

Mike (Dr Beav)

[Dave Schwartz]

I agree with your assertion of an actionable odds line. I don't believe logistic or probit model probabilities will solve the problem. It can be part of a solution, but not the solution.

Benter stumbled on the solution by trial and error, but not necessarily a complete solution. He actually solved the problem in reverse order.

Such solutions seem to be working for the five or six top whales.

[gm10]

An oddsline which is well calibrated as the tote odds, but sharper.

Mike (Dr Beav)

I've got one of those. It's based on a multinomial logit model (see Benter). The model is just slightly less efficient than the tote, but the ROI of the "best value horse that is also in the top 2 of winning probabilities" is around 5% (make that 10% using Betfair).

It's boring, but it makes a little bit of money. The key is getting fixed odds. You can just place your bets, go and watch TV and check your balance 6 hours later.

[TrifectaMike]

Such solutions seem to be working for the five or six top whales.

Now, Dave I'm surprised at your comment. Did I say it isn't profitable? I said it is part of the solution, but it is not necessarily the complete solution.

You well know that without the tote odds and associated probabilities, the logistic and probit models aren't sufficient.

So, I ask what is leading the path to a solution the logistic or probit or it might it just possibly be the tote odds?

Mike (Dr Beav)

[Cratos]

An oddsline which is well calibrated as the tote odds, but sharper.

Mike (Dr Beav)


The odds line should be a function of the handicapping model and should be consistent with the wager’s risk profile. I am quite sure you know and understand Arrow’s Paradox when it comes to risk.

The relevance of the odds line to the tote board is that the odds line should exploit the tote board; in general the tote board is a probabilistic representation of collusion between bettors.

Also I never answered your initial question and the answer is none of them. While they all have benefit to some degree I would rather start by developing my own predictive model and I would start with one built with a logarithmic curve model and drive it with laws of physics as they relate to the measurable parameters of horseracing.

[Dave Schwartz]

You well know that without the tote odds and associated probabilities, the logistic and probit models aren't sufficient.

Mike,

So what you meant by "aren't sufficient" was that the tote odds were missing?


Dave

[yak merchant]

logarithmic curve model and drive it with laws of physics as they relate to the measurable parameters of horseracing.

This. The hard part is the "pace predictor" to understand the distribution of possible outcomes depending on the pace scenarios. Wouldn't hurt to throw in a data mining app for trainer patterns. While I'm sure I could learn a lot from a mathematician and an expert horse player, I'd want a team of people in this order:

Computer Programmer
Statistician
Trainer/Physical Appearance Handicapper
Track Clocker

I have the programmer part covered, but even if I had the money, the other 3 probably would cost more than I could reasonable expect to extract from the game. If you had a crystal ball and could tell me that horse racing in 10 to 20 years is thriving, and takeouts have been lowered then maybe. But going full retard on a 'business model" that is precarious at best, and directly tied to the whims of an incompetent and corrupt bureaucracy sounds like a bad plan. It works alot better if you are a statistician/trainer that just inherited a bunch of money and all you need is to find/hire a programmer.

[sjk]

Now, Dave I'm surprised at your comment. Did I say it isn't profitable? I said it is part of the solution, but it is not necessarily the complete solution.

You well know that without the tote odds and associated probabilities, the logistic and probit models aren't sufficient.

So, I ask what is leading the path to a solution the logistic or probit or it might it just possibly be the tote odds?

Mike (Dr Beav)

Anyone who uses their program derived probabilities to look for bets which are overlaid by some factor to tote odds are using tote odds in their decision process.

It is arithmetically equivalent to using probabilities which are a linear combination of derived and tote probabilities.

So far as I know I have never achieved probabilities which are sharper than the tote (this is really hard to do) but as long as they are nearly as good and substantially different I can make a good profit.

Lots of good discussion of these types of things 6-8 years ago for anyone who cares to search.

[TrifectaMike]

Also I never answered your initial question and the answer is none of them. While they all have benefit to some degree I would rather start by developing my own predictive model and I would start with one built with a logarithmic curve model and drive it with laws of physics as they relate to the measurable parameters of horseracing.

I like the idea. This is exactly what I'm looking for. If you can get a set of derived probabilities, I would incorporate your predictive model into my ideal overall model.

Mike (Dr Beav)

[TrifectaMike]

This. The hard part is the "pace predictor" to understand the distribution of possible outcomes depending on the pace scenarios.

This is the most difficult part of a race to quantify. Why? Because it has to be answered with the application of Chaos Theory (change the initial conditions and the results are widely different).

Mike (Dr Beav)

[TrifectaMike]

Mike,

So what you meant by "aren't sufficient" was that the tote odds were missing?


Dave

I'm saying that and more.

Let's look at this as a curve. The totes odds is the basic curve. We want to apply "additional" information to bend and reshape the curve.

Mike (Dr Beav)

[TrifectaMike]

It is arithmetically equivalent to using probabilities which are a linear combination of derived and tote probabilities.

Bingo! Therein lies the problem.

Mike (Dr Beav)

[superfecta]

Id just want to be able to get accurate payoffs for exotic bets and a probability scale for said bet.
so if I want to bet a super,I want to know the expected payoff for the horses I bet according to the amount of horses I bet.
Since it is a finite pool,I dont want to bet a dollar super if my selections will take down the pool for a dime.
And I dont want to invest too much that even if I win,I bet too much to win the bet.
I want to maximize value.
Course this takes records,other bettors selections and access to pools,all in real time
Simple huh :(

[yak merchant]

This is the most difficult part of a race to quantify. Why? Because it has to be answered with the application of Chaos Theory (change the initial conditions and the results are widely different).

Mike (Dr Beav)

Why? Human intent. Leading to almost instant chaos and skewed data. (Struggling through an Applied Chaos book now). If I do the whole Randy Giles pace profiling thing, and the data suggests there are 6 Early speed horses are in a race, I'm not the only one with that knowledge. All the trainers see this also and instruct their jocks to strangle their mount turning a seemingly fast paced race to a crawler. I then take this information and put into my database and now the 6 Early Speed horses aren't early speed horses anymore?. Then you have trainers doing in race "workouts" where they send a closer on a burnout, etc, etc.

Anyway, I use brute force Monte Carlo (not smart enough to understand Markov Chains well enough to implement), for my stuff (a baseball sim), but it is definitely a lot harder trying to model the physics of a horse race. I'm really just trying to get to a fancy energy and pace simulator that takes into account deviations in ground loss and the expected pace according to the tendencies of the horses, trainers, and jockeys. Easier said than done.

The problem (and genius) I see with the "logarithmic curve" model is it's crazy dependence on pace. Small changes in fast paces can cause even greater chaos in the "system". I haven't finished modeling it, but my initial findings are that the difference between a 22.4 and 22.2 first split and a 23.4 and 23.2 split are light years apart. So in the grand scheme of things, for better or worse, if you are using a logrithmic curve, basically the pace becomes the mother of all inputs. This is why I think a lot of the pace software out there, makes the user pick the paceline, and the results of people "picking" those pacelines can be so varied.

[TrifectaMike]

Why? Human intent. Leading to almost instant chaos and skewed data. (Struggling through an Applied Chaos book now). If I do the whole Randy Giles pace profiling thing, and the data suggests there are 6 Early speed horses are in a race, I'm not the only one with that knowledge. All the trainers see this also and instruct their jocks to strangle their mount turning a seemingly fast paced race to a crawler. I then take this information and put into my database and now the 6 Early Speed horses aren't early speed horses anymore?. Then you have trainers doing in race "workouts" where they send a closer on a burnout, etc, etc.

Anyway, I use brute force Monte Carlo (not smart enough to understand Markov Chains well enough to implement), for my stuff (a baseball sim), but it is definitely a lot harder trying to model the physics of a horse race. I'm really just trying to get to a fancy energy and pace simulator that takes into account deviations in ground loss and the expected pace according to the tendencies of the horses, trainers, and jockeys. Easier said than done.

The problem (and genius) I see with the "logarithmic curve" model is it's crazy dependence on pace. Small changes in fast paces can cause even greater chaos in the "system". I haven't finished modeling it, but my initial findings are that the difference between a 22.4 and 22.2 first split and a 23.4 and 23.2 split are light years apart. So in the grand scheme of things, for better or worse, if you are using a logrithmic curve, basically the pace becomes the mother of all inputs. This is why I think a lot of the pace software out there, makes the user pick the paceline, and the results of people "picking" those pacelines can be so varied.

I must admit. I enjoyed reading all three paragraphs. Well thought out and insightful.

Mike(Dr Beav)

[Jeff P]

If that is true, it is sad.

About a year ago or maybe two years ago, Jeff Platt wrote something in a thread about seeing a professor to get a better understanding on a mathematical concept. At the time, I was shocked that he had considered the approach. It was one of the wisest things I've read on this board.

I don't know if he went any further than just being curious. I'm sure he knows what I am speaking of. If he wants to comment, that is his choice. I will not.

Mike (Dr Beav)

I know exactly which post you are talking about.

Yes, I was able to take ideas I talked about in that post and work them into JCapper Platinum.

Was that effort successful?

I like to think so. :) :)

-jp

.

[TrifectaMike]

I know exactly which post you are talking about.

Yes, I was able to take ideas I talked about in that post and work them into JCapper Platinum.

Was that effort successful?

I like to think so. :) :)

-jp

.

Congratulations. Was the effort successful? Can I say for you that properly implemented it was EXTREMELY successful.

That said, I believe you might be the only person on this board with insight into what I am alluding to in this thread.

Great success,
Mike (DR Beav)

[gm10]

Not entirely sure it this point needs to be made, but let me just say that there is a big difference between a model that adds value, and input variables that add value to that model.

A model that adds value will have been tested and re-tested against half a million races or more. The estimation of parameters will be based on reasonable assumptions and proven theorems, and probably be done through an algorithm that has been academically described in the context of the model. The theoretical framework of the model will be sound, and its implications well understood.

The multinomial logit and probit model are the two most obvious examples for this approach. I don't think the probit model can be implemented by anyone who doesn't get the help of professional statistical programmers, however. The multinomial logit model is relatively straightforward to implement - the seminal paper by McFadden is probably all that is required.


The input variables for any model, variables such as metrics resulting from pace scenario analysis/form analysis/class analysis/etc, are really quite separated from the model itself. It's the part where you need to implement, test (and maybe reject) conventional handicapping 'truths' - and probably add a bit of your own creativity and wisdom.

Models are great. They give you an objective odds-line that can compete with the public's odds-line, and often even outperform it. They are, however, also expensive and a lot, a lot, a lot of work, infinitely tweakable, and definitely not a substitute for intuition or common sense. If you are really clever enough to come up with really good input variables, you may not even need an expensive model that translates them into winning probabilities.

That's my opinion on it.

[Cratos]

Not entirely sure it this point needs to be made, but let me just say that there is a big difference between a model that adds value, and input variables that add value to that model.

A model that adds value will have been tested and re-tested against half a million races or more. The estimation of parameters will be based on reasonable assumptions and proven theorems, and probably be done through an algorithm that has been academically described in the context of the model. The theoretical framework of the model will be sound, and its implications well understood.

The multinomial logit and probit model are the two most obvious examples for this approach. I don't think the probit model can be implemented by anyone who doesn't get the help of professional statistical programmers, however. The multinomial logit model is relatively straightforward to implement - the seminal paper by McFadden is probably all that is required.


The input variables for any model, variables such as metrics resulting from pace scenario analysis/form analysis/class analysis/etc, are really quite separated from the model itself. It's the part where you need to implement, test (and maybe reject) conventional handicapping 'truths' - and probably add a bit of your own creativity and wisdom.

Models are great. They give you an objective odds-line that can compete with the public's odds-line, and often even outperform it. They are, however, also expensive and a lot, a lot, a lot of work, infinitely tweakable, and definitely not a substitute for intuition or common sense. If you are really clever enough to come up with really good input variables, you may not even need an expensive model that translates them into winning probabilities.

That's my opinion on it.


GM10, would you not say that models are great because in general horseplayers are linear algorithmic minded people.

[Cratos]

I like the idea. This is exactly what I'm looking for. If you can get a set of derived probabilities, I would incorporate your predictive model into my ideal overall model.

Mike (Dr Beav)

I have developed a predictive model that is based on the algorithm that will hit about 40% winners at odds above 3-1. However the sample size is not very large; about 300 actual wagers over the last 5 years, but I do believe it would hold up to a much larger sample size. The strength of the model is combining class and speed in a unique way to determine a performance final time and not a speed figure.

[TrifectaMike]

An oddsline which is well calibrated as the tote odds, but sharper.

Mike (Dr Beav)

Someone asked in a pm if I could explain what I mean by the terms calibrated and sharpness. So, I'll try.

Calibration
Suppose a horse player gives the probability of a horse winning in an up coming race, in advance. Assume that we have collected information on past predictions and actual results. We can know look at all those horses when the probability forecast was close to 30%. We would expect that the proportion of winners in those races when the forecast was 30% is close to 0.3. If this is true, we say that the horse player is calibrated at 0.3.

If the horse player is calibrated at all probabilities they forecast, then he/she is calibrated.

Sharpness
Calibration is an important criteria for evaluating the players probabilities. However, it is not enough to have a calibrated probability. Let's assume a player assigns a probability of .35 to every favorite running in the last 3 years. Let's also assume the the .35 holds up, the player is going to be calibrated. But this horse player is not very useful in making decisions. We need more than calibration. We need the probabilities to be as extreme as possible. A probability which is close to 0 or 1 is more informative than one that falls toward the middle of the interval.

So, a sharp forecast is one that approaches 0 or 1. To put calibration and sharpness in context, the goal is to maximize sharpness subject to calibration.

Mike (Dr Beav)

[TrifectaMike]

I have developed a predictive model that is based on the algorithm that will hit about 40% winners at odds above 3-1. However the sample size is not very large; about 300 actual wagers over the last 5 years, but I do believe it would hold up to a much larger sample size. The strength of the model is combining class and speed in a unique way to determine a performance final time and not a speed figure.

Why so few plays? Is this a spot play? Would not all horses to subject to the same metric?

Mike (Dr Beav)

[Cratos]

Why so few plays? Is this a spot play? Would not all horses to subject to the same metric?

Mike (Dr Beav)

I am not an exotic player and bet win only at $250 to $2500 per wager depending the expected win payoff and win probability from my model. To do this it takes both patience and money

[gm10]

GM10, would you not say that models are great because in general horseplayers are linear algorithmic minded people.

Maybe you have a point, yes. It's not easy being non-linear for us humans, even
though often reality is.

[gm10]

Someone asked in a pm if I could explain what I mean by the terms calibrated and sharpness. So, I'll try.

Calibration
Suppose a horse player gives the probability of a horse winning in an up coming race, in advance. Assume that we have collected information on past predictions and actual results. We can know look at all those horses when the probability forecast was close to 30%. We would expect that the proportion of winners in those races when the forecast was 30% is close to 0.3. If this is true, we say that the horse player is calibrated at 0.3.

If the horse player is calibrated at all probabilities they forecast, then he/she is calibrated.

Sharpness
Calibration is an important criteria for evaluating the players probabilities. However, it is not enough to have a calibrated probability. Let's assume a player assigns a probability of .35 to every favorite running in the last 3 years. Let's also assume the the .35 holds up, the player is going to be calibrated. But this horse player is not very useful in making decisions. We need more than calibration. We need the probabilities to be as extreme as possible. A probability which is close to 0 or 1 is more informative than one that falls toward the middle of the interval.

So, a sharp forecast is one that approaches 0 or 1. To put calibration and sharpness in context, the goal is to maximize sharpness subject to calibration.

Mike (Dr Beav)

I think I know where you're coming from but I am not sure about the second part of your post. You are going to struggle to have both 'calibrated' and 'sharp' odds. It sounds almost like a contradiction.

IMO what you want instead is calibrated odds that are different from the public's odds. The overlap between the two has to be minimized.

[TrifectaMike]

I think I know where you're coming from but I am not sure about the second part of your post. You are going to struggle to have both 'calibrated' and 'sharp' odds. It sounds almost like a contradiction.


I can assure you there is no contradiction. The goal is to maximize sharpness subject to calibration. It is doable. It is provable.

Think of a distribution which is concave (parabola shape)

Mike (Dr Beav)

[gm10]

I can assure you there is no contradiction. The goal is to maximize sharpness subject to calibration. It is doable. It is provable.

Think of a distribution which is concave (parabola shape)

Mike (Dr Beav)

What distribution do you mean, Mike? The one that governs the winning probabilities?

How would you do/prove this?

It seems quite hard to me, since on the one hand you want probabilities that reflect true winning chances, and on the other hand you want (unreflective) probabilities that are close to 0 or 1.

Firstly, probabilities close to 0 and 1 are by definition not calibrated, and secondly, I think horse racing is too random to have a lot of very low and very high probabilities.

I agree with your maximizing/minimizing approach, but I really think you ought to minimize the overlap between your calibrated odds and the public odds instead. The calibration of your odds is the first step in the process and should be independent of the next steps in the process.

[GameTheory]

I've tried to explain this concept a few times before (without using the terms "calibration" or "sharpness"). This was probably the most exhaustive attempt:

http://www.paceadvantage.com/forum/showpost.php?p=319872&postcount=183

(This was in the "ValueBetting" thread where the guy that wrote that book was trying to convince everyone that he had reached handicapping nirvana because his probabilities were well-calibrated.)

[TrifectaMike]

What distribution do you mean, Mike? The one that governs the winning probabilities? Yes, the winning probabilities - concave (parabolic)

How would you do/prove this? You really expect an answer.

It seems quite hard to me, since on the one hand you want probabilities that reflect true winning chances, and on the other hand you want (unreflective) probabilities that are close to 0 or 1. The probabilities are initially calibrated. Only after calibration are the probabilities transformed for sharpness( Actually it's accomplished simultaneously...I am purposely leaving out a major step here. The process is not linear.)

Firstly, probabilities close to 0 and 1 are by definition not calibrated, and secondly, I think horse racing is too random to have a lot of very low and very high probabilities. They aren't 0 or 1. They are pushed in that direction

I agree with your maximizing/minimizing approach, but I really think you ought to minimize the overlap between your calibrated odds and the public odds instead. The calibration of your odds is the first step in the process and should be independent of the next steps in the process. Start with a well calibrated personal oddsline introduce some nonlinear techniques, maximize sharpness and reshape the probabilities associated with the tote odds.

I know this may be frustrating for you, but this as far as I am will to go. As I said in an earlier post, Jeff Platt is probably the only person on this board with an understanding of the framework (at minimum a partial framework)

Let me simply say that the approach includes both a probabilistic framework as well as a statistic approach. It is a complex methodology.

Mike (Dr. Beav)

[TrifectaMike]

I've tried to explain this concept a few times before (without using the terms "calibration" or "sharpness"). This was probably the most exhaustive attempt:

http://www.paceadvantage.com/forum/showpost.php?p=319872&postcount=183

(This was in the "ValueBetting" thread where the guy that wrote that book was trying to convince everyone that he had reached handicapping nirvana because his probabilities were well-calibrated.)

Thanks for the link. I'll read it.

Mike (Dr Beav)

[sjk]

I've tried to explain this concept a few times before (without using the terms "calibration" or "sharpness"). This was probably the most exhaustive attempt:

http://www.paceadvantage.com/forum/showpost.php?p=319872&postcount=183

(This was in the "ValueBetting" thread where the guy that wrote that book was trying to convince everyone that he had reached handicapping nirvana because his probabilities were well-calibrated.)

A pleasure to re-read this thoughtful and cogent post.

[gm10]

I know this may be frustrating for you, but this as far as I am will to go. As I said in an earlier post, Jeff Platt is probably the only person on this board with an understanding of the framework (at minimum a partial framework)

Let me simply say that the approach includes both a probabilistic framework as well as a statistic approach. It is a complex methodology.

Mike (Dr. Beav)


You are clearly not willing to share much of your methodology, but tell me, are you in the end not looking for an unbiased estimator?

[TrifectaMike]

You are clearly not willing to share much of your methodology, but tell me, are you in the end not looking for an unbiased estimator?


No.

Mike (Dr Beav)

[gm10]

No.

Mike (Dr Beav)

So you are not looking to minimize the error between your own odds and the true odds?

That's kinky.

The purpose of an odds-line is usually to estimate each horse's true estimate of winning the race.

[TrifectaMike]

I've tried to explain this concept a few times before (without using the terms "calibration" or "sharpness"). This was probably the most exhaustive attempt:

http://www.paceadvantage.com/forum/showpost.php?p=319872&postcount=183

(This was in the "ValueBetting" thread where the guy that wrote that book was trying to convince everyone that he had reached handicapping nirvana because his probabilities were well-calibrated.)

Insightful, interesting, well reasoned, and I enjoyed reading it.

Mike (Dr Beav)

[Cratos]

I know this may be frustrating for you, but this as far as I am will to go. As I said in an earlier post, Jeff Platt is probably the only person on this board with an understanding of the framework (at minimum a partial framework)

Let me simply say that the approach includes both a probabilistic framework as well as a statistic approach. It is a complex methodology.

Mike (Dr. Beav)

Not really, the probability (odds) of a horse winning is based on an accurate assessment of its ability to win a given race with respect to the other horses in the race and that probability is not necessary the wagering odds by the handicapper.

The handicapper's wagering odds are determined by his/her risk profile which might be very different from the horse's probability of winning.

Therefore the attempt to converge the two is virtually impossible since one is calculated based on a set of assumptions and the other is genetically innate to the handicapper.

[TrifectaMike]

Not really, the probability (odds) of a horse winning is based on an accurate assessment of its ability to win a given race with respect to the other horses in the race and that probability is not necessary the wagering odds by the handicapper.

The handicapper's wagering odds are determined by his/her risk profile which might be very different from the horse's probability of winning.

Therefore the attempt to converge the two is virtually impossible since one is calculated based on a set of assumptions and the other is genetically innate to the handicapper.

After careful consideration I decided a response is necessary.

In view of what you described, let me say the following:

Calibration refers to the statistical compatibility of the predictive probability distributions (oddsline) and the observations (race results).

Sharpness refers to the spread of the predictive probability distributions and is a property of the individual.

So, what you have stated is inline with these definitions.

I might regret answering gm10's unbiased estimator question.

Mike (Dr Beav)

[riskman]

I've tried to explain this concept a few times before (without using the terms "calibration" or "sharpness"). This was probably the most exhaustive attempt:

http://www.paceadvantage.com/forum/showpost.php?p=319872&postcount=183

(This was in the "ValueBetting" thread where the guy that wrote that book was trying to convince everyone that he had reached handicapping nirvana because his probabilities were well-calibrated.)
Thanks for providing this perceptive post. I am finally starting to get tuned into what is being discussed here.
GameTheory said:
"But if you do "comprehensive" handicapping using all the conventional factors with the idea in mind that you are simply going to do it better than the public (like ValueBetting, presumably), then minimizing your error is generally just going to lead you to mirror the public odds, and you'll either find no overlays, or the overlays you do find won't make any money. So now you've got to strike a balance between your performance and your correlation to the public. You'll find that you can actually make more money by REDUCING your performance if you can decorrelate your line relative to the public's by some amount."
This statement appears to have merit.

[TrifectaMike]

I've tried to explain this concept a few times before (without using the terms "calibration" or "sharpness"). This was probably the most exhaustive attempt:

http://www.paceadvantage.com/forum/showpost.php?p=319872&postcount=183

(This was in the "ValueBetting" thread where the guy that wrote that book was trying to convince everyone that he had reached handicapping nirvana because his probabilities were well-calibrated.)

Happy Easter

From GT's post and additions:

FinPos AP_A AP_B AP_C AP_D AP_E
1 0.40 0.30 0.00 1.0 1.0
2 0.10 0.25 0.00 0.00 0.00
3 0.05 0.20 0.00 0.00 0.00
4 0.25 0.15 1.0 0.00 0.00
5 0.20 0.10 0.00 0.00 0.00

FinPos = Finish Position
AP_A = Assigned Probability Line A
AP_B = Assigned Probability Line B
AP_C = Assigned Probability Line C
AP_D = Assigned Probability Line D
AP_E = Assigned Probability Line E

Weights????? Recalibrate-How??????? Bingo!!!!

Mike (Dr Beav)

[Cratos]

After careful consideration I decided a response is necessary.

In view of what you described, let me say the following:

Calibration refers to the statistical compatibility of the predictive probability distributions (oddsline) and the observations (race results).

Sharpness refers to the spread of the predictive probability distributions and is a property of the individual.

So, what you have stated is inline with these definitions.

I might regret answering gm10's unbiased estimator question.

Mike (Dr Beav)

Okay, I agree with what you are saying

[Fastracehorse]

After careful consideration I decided a response is necessary.

In view of what you described, let me say the following:

Calibration refers to the statistical compatibility of the predictive probability distributions (oddsline) and the observations (race results).

Sharpness refers to the spread of the predictive probability distributions and is a property of the individual.

So, what you have stated is inline with these definitions.

I might regret answering gm10's unbiased estimator question.

Mike (Dr Beav)

........to engage in the relationship between an oddsline and the probability of winning per se.

I cannot control the oddsline, nor do I wish to set one - but I can increase my skill level at picking winners at all prices. The higher the better.

I do not bet short prices ( say 2-1 ) to win - so I'm not interested to try and uncover if 2-1 is fair odds - I don't care - but I do want to know If the 2-1 is going to win.

It takes alot of skill to determine the winner of the race - it also takes alot of time - and the time between parade to post and the off time takes handicapping energy that for me is best spent on uncovering the best bet. Again, I don't want to waste time on an oddsline - I already have set comfort levels on win wagers based on the likeliness of the horse winning which I don't quantify either. It makes no sense to me.

The game is about picking winners only. Then an effort is made how to best exploit your skill level at turning a profit - not fair value; who cares what fair value on a horse is if he doesn't run?? Why would I care about how good a line I could set - a good handicapper knows the liekelihood of favorites - it's just natural. It's perplexing when players don't realize it's not a big deal to predict favorites.

In essence, I create a rudimentary oddsline anyhow, just by the nature of the handicapping process, but completely relax on the tote board's #'s - I don't care - I just need to know the winner.

fffastt

[GameTheory]

who cares what fair value on a horse is if he doesn't run??The chance of the horse not running is exactly why knowing fair value is...well, valuable. And if you are so sure of the winner, why do you not bet short prices? We all have to deal with the problem of value whether we like it or not...

[Fastracehorse]

The chance of the horse not running is exactly why knowing fair value is...well, valuable. And if you are so sure of the winner, why do you not bet short prices? We all have to deal with the problem of value whether we like it or not...

.......of bets. I saw one race today where there were two 2-1's and one 8:5.

The 8:5 was the best horse in the race IMO, but for some reason was not meant and one of the 2-1's took a ton of late action. That horse made a strong rally, beat the 8:5 but just missed to the other 2-1.

I don't bet short prices to win but bet them in gimmicks.

I don't need to know a fair value line - I need to know chances of winning - the tote dicatates how I structure my bets - I don't structure my wagers before I handicap.

I deal with value. I love value. I don't care to limit or determine it prior to the wagering event; but attempt to exploit the tote during the event.

fffastt

[gm10]

After careful consideration I decided a response is necessary.

In view of what you described, let me say the following:

Calibration refers to the statistical compatibility of the predictive probability distributions (oddsline) and the observations (race results).

Sharpness refers to the spread of the predictive probability distributions and is a property of the individual.

So, what you have stated is inline with these definitions.

I might regret answering gm10's unbiased estimator question.

Mike (Dr Beav)

Mike, I am personally having difficulties with your posts because I can see that you are applying statistical analysis, which I am familiar with, but at the same time the terms you use are confusing me.

What do you mean with 'compatibility of the predictive probability distributions'? My first thoughts were that you are trying to find an unbiased estimator but you earlier suggested that you aren't?

Is the 'sharpness' part about having enough observations in the low and high odds categories?

[TrifectaMike]

Mike, I am personally having difficulties with your posts because I can see that you are applying statistical analysis, which I am familiar with, but at the same time the terms you use are confusing me.

What do you mean with 'compatibility of the predictive probability distributions'? My first thoughts were that you are trying to find an unbiased estimator but you earlier suggested that you aren't?

Is the 'sharpness' part about having enough observations in the low and high odds categories?

Look at my Happy Easter post and observe the new lines I added (degenerate distributions...constant random variable)... all biased, then read the last line

Mike (Dr Beav)

[TrifectaMike]

In a past thread about conditional wagering, GT made a comment that he would like to have the ability to wager on horses going off at odds less than tote odds. No one bothered ta ask why he would want this capability. Is it possible that he is introducing a degenerate distribution to his odds...maybe not directly. Maybe he wants the tote to reflect a probability closer to 1 than his own probability assigned to that particular horse... just maybe.

Mike (Dr Beav)

[TrifectaMike]

[QUOTE=Fastracehorse].I do not bet short prices ( say 2-1 ) to win - so I'm not interested to try and uncover if 2-1 is fair odds - I don't care - but I do want to know If the 2-1 is going to win./QUOTE]

If you have determine that the 2-1 will win with such certainty, why not bet it? Better yet tell us when these 2-1 will win, so others can bet them.

Mike (Dr Beav)

[Robert Fischer]

my taxes :D

[Robert Fischer]

seriously i would use my method.

the mathematician and programmer could help to automate some of the grunt work into an even easier "press-of-a-button" format access and download setting system. The horseplayer(s) would be welcome to bounce ideas off to me or put in work at the track in the mornings.

[GameTheory]

In a past thread about conditional wagering, GT made a comment that he would like to have the ability to wager on horses going off at odds less than tote odds. No one bothered ta ask why he would want this capability. Is it possible that he is introducing a degenerate distribution to his odds...maybe not directly. Maybe he wants the tote to reflect a probability closer to 1 than his own probability assigned to that particular horse... just maybe.I did? I was probably saying that I wanted to the ability to set a maximum odds as well as a minimum odds for conditional wagers. The only conditional wagering ability offered so far by some ADWs is a minimum odds requirement for your selection (at X minutes to post, your choice of X). I wasn't thinking about anything other than being able to eliminate horses that were going off at 40-1 for some reason when I would have expected to have them go off at 8-1 or so.

If I was blending the odds with the tote in real-time (which I have done in the past), then I have no use for conditional wagering on the ADW side since my program is calculating all that in real-time anyway (and making the bets) -- i.e. I have my own conditional wagering. But on spot plays, the ability to define more conditions would be useful, that's all. (So I don't need a computer monitoring it all the time.)

Like others here, I am having trouble following what you are getting at here, although I think I am somewhat familiar with the concepts, but not with the terms you are using since I am not formally trained in these things. (I had to look up "degenerate distribution" just now to find out that it meant "an oddsline with a single 1.0 and the rest 0.0".) I need dumbed-down explanations with worked-through examples that I can follow along, which is why my own posts on such subjects are usually so long-winded. Plus I think you are holding back the good parts, eh?

[Fastracehorse]

[QUOTE=Fastracehorse].I do not bet short prices ( say 2-1 ) to win - so I'm not interested to try and uncover if 2-1 is fair odds - I don't care - but I do want to know If the 2-1 is going to win./QUOTE]

If you have determine that the 2-1 will win with such certainty, why not bet it? Better yet tell us when these 2-1 will win, so others can bet them.

Mike (Dr Beav)

2-1 is spinning my wheels in the mud as far as MOST win wagers go. People are mature enough to make their own financial decisions, good or bad.

I'm a big boy. So are you. Focusing on picking winners is my creed. What could possibly be more important?

fffastt

[Light]

You can never achieve true odds nor come close. The reason is because handicapping software is 2 dimensional and the realities of racing are 3 dimensional.

[TrifectaMike]

[QUOTE=TrifectaMike]

2-1 is spinning my wheels in the mud as far as MOST win wagers go. People are mature enough to make their own financial decisions, good or bad.

I'm a big boy. So are you. Focusing on picking winners is my creed. What could possibly be more important? Making lots of money!

fffastt

Whatever works for you.

Mike (Dr Beav)

[TrifectaMike]

You can never achieve true odds nor come close. The reason is because handicapping software is 2 dimensional and the realities of racing are 3 dimensional.

Is that third dimension a light phase?

Mike (Dr Beav)

[Jeff P]

[QUOTE=TrifectaMike]

2-1 is spinning my wheels in the mud as far as MOST win wagers go. People are mature enough to make their own financial decisions, good or bad.

I'm a big boy. So are you. Focusing on picking winners is my creed. What could possibly be more important?

fffastt

I'd argue that the ability to make intelligent play or pass decisions has far more effect on the player's long term expectation than the selection process itself. In the right hands, a good oddsline can assist in that. :)

Suppose for the sake of argument you are a good handicapper with a proven track record. Through your own good "handicapping" you select a horse. Let's also say that you feel pretty strongly about your selection and like its chances in today's race.

You look up at the tote board and you see that the public doesn't agree with you. Some other horse that you don't like is the favorite. Your selection is 4-1.

Do you bet? (Most likely yes... no oddsline required.)

What if the same horse in the same race is 2-5? Do you still make the bet? Or do you sit on your hands?

What if someone sitting next to you offers 6-1 on your horse? Do you bet then?

Or what if you looked at the exotics and saw a way to get 8-1 on your selection?

To my way of thinking, price is everything because it drives value - which I'd argue is more important than the selection process or picking winners.


-jp

.

[Light]

Is that third dimension a light phase?


Its the holy grail of racing that no handicapping philosophy nor software has attained or probably ever will. The negative probability of this game is very strong due to this 3d reality vs. our 2d paper handicapping. And yes you will need at least something like a "phaser" to overcome the lack of an entire dimension.

One thing comes to mind in this area is beginners luck. I had it when I started and it is a phenomena that might encroach on this sorely needed 3rd dimension. Beginners do not think at all like seasoned handicappers. They don't have all the rules in place and are not as analytical,yet many have tremendous "luck" when they first start out and even embarass their hosts who try to explain the game from 30 years of "experience" .What's happening is these "lucky" newbies are thinking outside the box and don't even realize it. Of course it doesn't last,but during that "innocent" phase it can be uncanny. Now program that part and I'll buy it.

[Fastracehorse]

[QUOTE=Fastracehorse]

Whatever works for you.

Mike (Dr Beav)

,,,,,whatever works. ;)

fffastt

[Fastracehorse]

[QUOTE=Fastracehorse]

I'd argue that the ability to make intelligent play or pass decisions has far more effect on the player's long term expectation than the selection process itself. In the right hands, a good oddsline can assist in that. :)

Suppose for the sake of argument you are a good handicapper with a proven track record. Through your own good "handicapping" you select a horse. Let's also say that you feel pretty strongly about your selection and like its chances in today's race.

You look up at the tote board and you see that the public doesn't agree with you. Some other horse that you don't like is the favorite. Your selection is 4-1.

Do you bet? (Most likely yes... no oddsline required.)

What if the same horse in the same race is 2-5? Do you still make the bet? Or do you sit on your hands?

What if someone sitting next to you offers 6-1 on your horse? Do you bet then?

Or what if you looked at the exotics and saw a way to get 8-1 on your selection?

To my way of thinking, price is everything because it drives value - which I'd argue is more important than the selection process or picking winners.


-jp

.

.........that I do a rudimentary oddsline naturally through the handicapping process. For eg., I generally know what the public will like; and I know if I horse I like will be a $. It's impossible to avoid. My point was that it is not neccessary to consciously formulate fair value for horses that are not going to run anyways - it's too time consuming.

The final tote is what it is.

Interesting point you make: value trumps the selection process.

I love longshots but I have faith in the selection process. I do think it is important to play lower % creative angles; that are good $'s - to try and get some of that value.

fffastt

[Fastracehorse]

Its the holy grail of racing that no handicapping philosophy nor software has attained or probably ever will. The negative probability of this game is very strong due to this 3d reality vs. our 2d paper handicapping. And yes you will need at least something like a "phaser" to overcome the lack of an entire dimension.

One thing comes to mind in this area is beginners luck. I had it when I started and it is a phenomena that might encroach on this sorely needed 3rd dimension. Beginners do not think at all like seasoned handicappers. They don't have all the rules in place and are not as analytical,yet many have tremendous "luck" when they first start out and even embarass their hosts who try to explain the game from 30 years of "experience" .What's happening is these "lucky" newbies are thinking outside the box and don't even realize it. Of course it doesn't last,but during that "innocent" phase it can be uncanny. Now program that part and I'll buy it.

......I would enjoy the read.

fffastt

[TrifectaMike]

Like others here, I am having trouble following what you are getting at here, although I think I am somewhat familiar with the concepts, but not with the terms you are using since I am not formally trained in these things. (I had to look up "degenerate distribution" just now to find out that it meant "an oddsline with a single 1.0 and the rest 0.0".) I need dumbed-down explanations with worked-through examples that I can follow along, which is why my own posts on such subjects are usually so long-winded. Plus I think you are holding back the good parts, eh?

Am I holding back the good parts? Absolutely!

The purpose of using the term degenerate distribution was purposeful. If you have an interest in this subject, you would think of the Probability mass function as a Dirac Delta function, and the Cumulative distribution function as a translated Heaviside step function.

Let's assume our intent is to simplify the choice of the prospective winners in a horse race.

One way to accomplish this task is to stretch the distances between the higher probability horses and the lower probability horses. Transforming probabilities toward the direction of preferred probabilities is to change from an arithmetic scale to a geometric scale.

Another way to accomplish the same task is instead of transformation, a probability distribution (degenerate distribution(s)) is directly introduced.

Mike (Dr Beav)

[GameTheory]

Am I holding back the good parts? Absolutely!

The purpose of using the term degenerate distribution was purposeful. If you have an interest in this subject, you would think of the Probability mass function as a Dirac Delta function, and the Cumulative distribution function as a translated Heaviside step function.

Let's assume our intent is to simplify the choice of the prospective winners in a horse race.

One way to accomplish this task is to stretch the distances between the higher probability horses and the lower probability horses. Transforming probabilities toward the direction of preferred probabilities is to change from an arithmetic scale to a geometric scale.

Another way to accomplish the same task is instead of transformation, a probability distribution (degenerate distribution(s)) is directly introduced.

I do have an interest in the subject, but it is just over my head. I have no idea what you are saying. I understand the goal -- spreading the probabilities. Introducing a degenerate distribution? What does that mean, i.e. how do you do that? No clue.

[gm10]

Am I holding back the good parts? Absolutely!

The purpose of using the term degenerate distribution was purposeful. If you have an interest in this subject, you would think of the Probability mass function as a Dirac Delta function, and the Cumulative distribution function as a translated Heaviside step function.

Let's assume our intent is to simplify the choice of the prospective winners in a horse race.

One way to accomplish this task is to stretch the distances between the higher probability horses and the lower probability horses. Transforming probabilities toward the direction of preferred probabilities is to change from an arithmetic scale to a geometric scale.

Another way to accomplish the same task is instead of transformation, a probability distribution (degenerate distribution(s)) is directly introduced.

Mike (Dr Beav)

Mike, are you adding the degenerate RV

a) because your base probabilities are not very good and you want to improve them?

or

b) regardless of how good the base probabilities are, you are trying to make a intentionally inaccurate 'odds-line' whose only purpose it is to make money?

In either case, can you show us, without revealing your good parts, what the effect is of inserting this Dirac guy into your system? It sounds like it's working for you.

[yak merchant]

I definitely can't keep up with all this, one minute something seems to make sense, the next I'm not sure if I'm doing significance calcs correctly. I have a sneaking suspicion it will end up with a discussion on Laplace and/or Z transforms and like always I'll have no clue. Nevertheless maybe I'll learn something, so I appreciate those taking the time to post.

[TrifectaMike]

Mike, are you adding the degenerate RV

a) because your base probabilities are not very good and you want to improve them?

or

b) regardless of how good the base probabilities are, you are trying to make a intentionally inaccurate 'odds-line' whose only purpose it is to make money?

In either case, can you show us, without revealing your good parts, what the effect is of inserting this Dirac guy into your system? It sounds like it's working for you.

b) is close to truth.

Without getting too technical or complex, and still maintain some distance to what I'm actually doing, let me attempt to describe some simple ideas.

Let's assume I have an extremely well calibrated oddsline (unbiased). Now, this oddsline will look good theoretically according to all the statistical tests. However, will it differ sufficiently from the tote odds derived probabilities to make it long term profitable? Maybe or maybe not.

We spoke of calibration (what old-timers called reliability) in this thread. We spoke of averaging being at the heart of most techniques.

Let me speak of averaging in a different sense. Let's assume we have a great oddsline and has fidelity with the tote-odds. Not very helpful as a decision tool, but a good tool nonetheless. The objective probability is most likely in the neighborhood of the probability line. It could be below or above the probability value for each horse.

Okay, I hope so far.

Now, let me introduce (this for explanation only, but you'll get the idea) some subjectivity (Bayesian stuff on a single race basis), so that it can some how bound the objective probability from below and above. So, we have have this well calibrated oddsline, which we believe is within a distance X of the objective probability, and we have some subjectivity, which we believe contains bounds on the objective probabilities. So, if we can strategically introduce a single race subjective distribution (s) to "merge" with the oddsline, it will appear to be less calibrated, but sharper.

However, appearance isn't always reality. It only appears less claibrated according to the objective function we defined in determining the probability curve for the oddsline.

Let me try a very simple example (this might help, I hope).

Let's take two handicappers. Each produces an oddsline according their methodology (one can be a logistic regression, the other based on a scoring system) and both are well calibrated according to a scoring rule.Suppose we know the objective probability of Horse A to be .3. Handicapper 1 says it is .4, and Handicapper 2 says it is .2. A simple average would be closer to the truth, because the REAL value is bounded by their estimates.

If both handicappers are both below or above, then averaging will drive us further from the objective probability. So, it would be better to accept one of the probabilities over the other (so we have choice over averaging in this case.

If you can understand what I said, it can be helpful, otherwise it'll make absolutely no sense to you.

Mike (Dr. Beav)

[Cratos]

I do have an interest in the subject, but it is just over my head. I have no idea what you are saying. I understand the goal -- spreading the probabilities. Introducing a degenerate distribution? What does that mean, i.e. how do you do that? No clue.

If I understand Mike, he means by a degenerate random variable which is called X; X has only a single possible value with probability 1. If this is true Mike, what is X or is that something that cannot be revealed.

[GameTheory]

If I understand Mike, he means by a degenerate random variable which is called X; X has only a single possible value with probability 1. If this is true Mike, what is X or is that something that cannot be revealed.I'm wondering more at what stage of the process is it "introduced" and what does "introduced" mean exactly?

[gm10]

If I understand Mike, he means by a degenerate random variable which is called X; X has only a single possible value with probability 1. If this is true Mike, what is X or is that something that cannot be revealed.

That is the crux of the matter.
Maybe it's some sort of elimination variable. For example. If the horse has won over the distance, the variable takes value 1, otherwise 0.

[TrifectaMike]

I'm wondering more at what stage of the process is it "introduced" and what does "introduced" mean exactly?

I kinda left this hanging. I've given it some thought, and I believe I have a way of explaining in understandable terms without giving up too much info.

I'm going to present the statements below without proof or justification, nor will I use a Bayesian framework. In spite of this the information I'll present can be used in a crude fashion...but usable.

When there are n horses in a race, your beliefs about the horses are represented by the probability vector

p = (p1,........,pn)

(where 0<= pi <= 1 for i = 1,......n, and the sum of the p i's = 1)

The question I'll ask is a follows.

If uncle Guido has some "additional" information about a particular horse, i, how do I reassign (or reshape) the probability vector?

Uncle Guido may operate as an angle player and has a terrific angle. Or he might be tuned into some inside information, or he maybe a great handicapper and is totally selection oriented.

As I said without proof or justification, here is how to update the probability vector:

Pnew = (1 - a)p + aei (This doesn't have a name, so let's call it Mike's update)

where ei = (0,.....0, 1 , .......0) is a degenerate distribution assigning all probability mass to the ith horse (Guido' s horse) and

0< a <= 1. a represents the confidence in uncle Guido's horse.

This should help. I won't go any further, but I will answer questions on the above equation.This sets the groundwork for getting expected winners toward 1 and the other horses toward 0.

Mike (Dr Beav)

[pondman]

If you had access to a mathematician, an expert horse player, the finances, the desire, and commitment, what type of horse racing selection process would you like to develop?
Mike (Dr. Beav)

Because I single out horses for win bets, I take the Black Swan (Nassim Taleb) approach to explain why my methods work. I make a profit because I make bets on horses when they are not expected to win by the crowd. And the profit come as the result of a few longer shots, meaning I limit a large number of bets on short priced horses. It's not difficult.

I don't see any way of quantifying this, other than employing some gimmicky math learned in fourth grade. I'd have to add a few circles for boolean routines (if the horse)-- but I think that was covered in 6th grade. I do use trend lines provided by excell, but all of you will point out future profits can't be predicted (I probably would agree.) But it's fun to see if my profits are above or below the trend for the year.

I think pace and speed knowledge is required by people searching for contenders in a pick 4 or 6. But I want the single, which over time will crush the field (at high odds.) And so most of my thinking is based on class and on why a trainer would place a horse in a race. I absolutely ignore beyers numbers or any other figures. And with the exception of horses off a maiden win (I want to know it won) or maidens on grass in S. California, I ignore the past performances entirely.

I'm the expert! I've analyzed racing at a number of tracks and have the belief each is unique, has it own structure, and gems and thorns. Therefore handicapping methods from one track should never be utilized for another.

I'm only going to step to the window when I know a horse belongs and has an advantage at a certain condition. Again that's some unknown quantity defined as class. I don't have a quantitative method of ranking class, other than it ran against better.

My financing is manageable. I'm comfortable in the $100 -300 range. Because I'm betting on California I could bump it up another few hundred, but again this would cause some discomfort when I go on a 7 horse losing streak.

I write this all for only one reason:

You are making it too difficult on yourself. You've probably heard it before, but watch the game until you can pick a winner at a good price.

[TrifectaMike]

I write this all for only one reason:

You are making it too difficult on yourself. You've probably heard it before, but watch the game until you can pick a winner at a good price.

One would believe that after 45 years at this I would have have learned that lesson.

Thanks for the advice.

Mike (Dr Beav)

[davew]

maybe what you are trying to do is make a value line?


many people have found ways to make a fairly accurate probability line
what is more difficult is making a closing odds line
(I just had a horse I bet at 4/1 while loading - after it won it closed at 9/5)


if you make an accurate probability line and an accurate closing odds line
you can then decide which entries/bets represent good value bets

a selection process can never include some racing problems we see every day
such as crappy riders cutting off or blocking your horse

or a longshot/no shot horse breaking for the lead for a 3f workout, messong up the pace for the entire race

[TrifectaMike]

I am in the twilight zone!!!

Mike (Dr Beav)

[Dave Schwartz]

many people have found ways to make a fairly accurate probability line what is more difficult is making a closing odds line
(I just had a horse I bet at 4/1 while loading - after it won it closed at 9/5)

I must seriously disagree with this statement.

I find it far easier to predict what odds a horse will go off at than it is to produce an accurate odds like.

Oddly enough, one can use the predicted line as a betting line and do reasonably well.

Regards,
Dave Schwartz

[Cratos]

I kinda left this hanging. I've given it some thought, and I believe I have a way of explaining in understandable terms without giving up too much info.

I'm going to present the statements below without proof or justification, nor will I use a Bayesian framework. In spite of this the information I'll present can be used in a crude fashion...but usable.

When there are n horses in a race, your beliefs about the horses are represented by the probability vector

p = (p1,........,pn)

(where 0<= pi <= 1 for i = 1,......n, and the sum of the p i's = 1)

The question I'll ask is a follows.

If uncle Guido has some "additional" information about a particular horse, i, how do I reassign (or reshape) the probability vector?

Uncle Guido may operate as an angle player and has a terrific angle. Or he might be tuned into some inside information, or he maybe a great handicapper and is totally selection oriented.

As I said without proof or justification, here is how to update the probability vector:

Pnew = (1 - a)p + aei (This doesn't have a name, so let's call it Mike's update)

where ei = (0,.....0, 1 , .......0) is a degenerate distribution assigning all probability mass to the ith horse (Guido' s horse) and

0< a <= 1. a represents the confidence in uncle Guido's horse.

This should help. I won't go any further, but I will answer questions on the above equation.This sets the groundwork for getting expected winners toward 1 and the other horses toward 0.

Mike (Dr Beav)

Admittedly you have me puzzle and I leaning toward the concept of the confidence interval or do you have a “smart” algorithm that you make “learn” with additional information and that phenomenon shifts the probability of the horse under scrutiny closer to 1 while the rest of the horses in the race move toward zero.

If the second part is true you have stumble into something very good because you will be exploiting all other bettors because they will be moving away from your horse giving it higher odds. This wouldn’t always be true because sometimes the two will converge and you will have the prohibitive 1-9 favorite

[Pell Mell]

Doesn't this thread belong in the General section under "Need Puzzle Help"?

[TrifectaMike]

If the second part is true you have stumble into something very good because you will be exploiting all other bettors because they will be moving away from your horse giving it higher odds. This wouldn’t always be true because sometimes the two will converge and you will have the prohibitive 1-9 favorite

Cratos, my uncle Guido are Bayesian belief networks.

Mike (Dr Beav)

P.S. There is no stumbling here.

[TrifectaMike]

Doesn't this thread belong in the General section under "Need Puzzle Help"?

No real puzzle here. I gave a very good equation (Mike's update), which can be put to good use. Why doesn't someone put into an excel spreadsheet with some data. It is good stuff.

Mike (Dr Beav)

[TrifectaMike]

This wouldn’t always be true because sometimes the two will converge and you will have the prohibitive 1-9 favorite

That is true. There is the occasional convergence. It is expected. It's a no bet.

[GameTheory]

I get it, thank you. Again, something I am familiar with, just with a less formal conceptualization of it.

[TrifectaMike]

Admittedly you have me puzzle and I leaning toward the concept of the confidence interval

Confidence intervals are not probabilities. It is a frequentist concept, which I prefer not to use. In fact, when most people use confidence intervals they erroneous interpret them as tolerance intervals, which are probabilities.

Mike (Dr Beav)

[Robert Goren]

Confidence intervals are not probabilities. It is a frequentist concept, which I prefer not to use. In fact, when most people use confidence intervals they erroneous interpret them as tolerance intervals, which are probabilities.

Mike (Dr Beav)People who try to use confidence levels on horse racing learn how useless they are. Generally after a great deal of work. They find out that they need more than if something effects the out come, they need to know by how much. If your method can't tell you that, then you are using the wrong method.

[m001001]

No real puzzle here. I gave a very good equation (Mike's update), which can be put to good use. Why doesn't someone put into an excel spreadsheet with some data. It is good stuff.

Mike (Dr Beav)

1. If someone is to test the idea, they need to know what's ei and how to generate/identify the horse with 1.

2. if the Pi in the probability vector is from a multi-factor regression, why can't ei be one of the factors? Specifically an indicator factor where the value is 1 or 0.

3. in any multinomial model, you can always have a nested models inside a model, how is this approach different from your Mike's Update?

[TrifectaMike]

1. If someone is to test the idea, they need to know what's ei and how to generate/identify the horse with 1.

2. if the Pi in the probability vector is from a multi-factor regression, why can't ei be one of the factors? Specifically an indicator factor where the value is 1 or 0.

3. in any multinomial model, you can always have a nested models inside a model, how is this approach different from your Mike's Update?

1. You can easily test the idea. If uou can generate a probability vector and then introduce some new information...it can be unusual tote activity, an angle for which you have stats for, a friend who you have confidence in their selection, etc, etc.

2. ei is NOT an indicator variable. As I said it it is a degenerate distribution assigning all probability mass to the ith horse.

3. This has absolutely nothing in common with nested models. Once again it a degenerate distribution, not a variable.

Mike (Dr Beav)

[Robert Goren]

I am too old and too sick to this kind of labor intensive research anymore, but if I were young and in better health, I would start out with track odds. Then I would look for something that beats the track odds or lose to the track odds. Say like raced 22 days ago. Then I look to see if the "edge" increased as the odds went lower. There are a zillion things to at and most off the time the track odds have it right. If I found two things, then I want to check how they worked together. In these days of good computers and large data bases, it could be done. You are only limited by what your mind can come up with to check. I think starting with the tracks odds is the way to go rather than trying figure out what the horses chances of winning is and decide if the track odds warrant a bet.

[Cratos]

Confidence intervals are not probabilities. It is a frequentist concept, which I prefer not to use. In fact, when most people use confidence intervals they erroneous interpret them as tolerance intervals, which are probabilities.

Mike (Dr Beav)

Mike, I think that you are on to something good and I undertsand confidence intervals, but I made an incorrect interpretation of what you are doing. I hope I am correct this time when I say the gist of what you are you are doing is conditional probability.

[TrifectaMike]

Mike, I think that you are on to something good and I undertsand confidence intervals, but I made an incorrect interpretation of what you are doing. I hope I am correct this time when I say the gist of what you are you are doing is conditional probability.


Part 1 is a logistic model.

Part 2 Is a conditional probability

Part 3 is putting it all together.

Mike (Dr Beav)

[Cratos]

Part 1 is a logistic model.

Part 2 Is a conditional probability

Part 3 is putting it all together.

Mike (Dr Beav)

From reading your posts I believe that you have both the mathematical and statistical skill set beyond most posters on this forum and when I reviewed the concepts of the logistic model and conditional probability; I concluded that what your model is doing is fascinating and if you have put this in practice in which you call “putting it all together,” you have moved a long way toward beating this game.

[TrifectaMike]

From reading your posts I believe that you have both the mathematical and statistical skill set beyond most posters on this forum and when I reviewed the concepts of the logistic model and conditional probability; I concluded that what your model is doing is fascinating and if you have put this in practice in which you call “putting it all together,” you have moved a long way toward beating this game.

Thank you for the kind words.

I wish more people here and in the academic world would be more Bayesian oriented in their analysis (either objectivist or subjectivist view).

Mike (Dr Beav)

[GameTheory]

I wish more people here and in the academic world would be more Bayesian oriented in their analysis (either objectivist or subjectivist view).I agree, but I would tend to put it in my own way: "Frequentists can suck it!"

[TrifectaMike]

I agree, but I would tend to put it in my own way: "Frequentists can suck it!"

For many years there has been a controversy over "frequentist" vs "Bayesian" methods of inference. Obviously I am a partisan on the Bayesian side. In the past there was a strong tendency on both sides to argue on the level of philosophy or ideology. However, due to recent work there is no reason to appeal to those arguments.

There is an abundance of theorems and numerous worked out numerical examples. As a result, the superiority of Bayesian methods is now a thoroughly demonstrated fact in hundreds of different areas.

Probability theory in actuality is an extension of logic.

Frequentist methods which use only sampling distributions are usable and useful in many simple, idealized problems. The problem is that they only represent special cases of probability theory, because they presuppose conditions( independent repetitions of a "random" experiment, but no prior information) that are hardly ever met in the real world.

Lacking the necessary theoretical principles, they choose a statistic from intuition rather from probability theory, and then invent ad hoc devices such as unbiased estimators, confidence intervals, tail-area significance, which are not contained in probability theory.

All these defects are corrected by Bayesian methods.

In fact all the capability to solve complex problems is contained in the simple product and sum rules of probability theory interpreted as extended logic, with no need for ad hoc devices.

I agree frequentists suck.

Mike (Dr Beav)

[GameTheory]

For many years there has been a controversy over "frequentist" vs "Bayesian" methods of inference. Obviously I am a partisan on the Bayesian side.Are you also a "Jayesian"? (E T Jaynes)

[TrifectaMike]

Are you also a "Jayesian"? (E T Jaynes)

Absolutely!!!!!!!!!!!!!!!!!!

He lives on, even in death.


Mike (Dr Beav)

[TrifectaMike]

Are you also a "Jayesian"? (E T Jaynes)

Now do you understand, why I gave Speculus such a hard time on his beaten lengths thing to validate subjective probabilities for unique events.

Once he used the words the horse won AND by x lengths, he was on shaky ground.

Mike (DR Beav)

[Fastracehorse]

Thank you for the kind words.

I wish more people here and in the academic world would be more Bayesian oriented in their analysis (either objectivist or subjectivist view).

Mike (Dr Beav)

......like that fig!

fffastt

[Fastracehorse]

Now do you understand, why I gave Speculus such a hard time on his beaten lengths thing to validate subjective probabilities for unique events.

Once he used the words the horse won AND by x lengths, he was on shaky ground.

Mike (DR Beav)

....was trying to evaluate a horse race, lol

fffastt

[TrifectaMike]

......like that fig!

fffastt

With over 2,200 post...can you direct me to any one of those over 2,200 that made any sense?

Mike (Dr Beav)

[GameTheory]

Now do you understand, why I gave Speculus such a hard time on his beaten lengths thing to validate subjective probabilities for unique events.

Once he used the words the horse won AND by x lengths, he was on shaky ground.
I don't exactly remember, but I think it was just a semantics thing. He may have co-opted terminology that he shouldn't have, but I still followed his point. I think you were saying he was trying to validate subjective probabilities and doing it wrong, but he really wasn't trying to do that at all -- he just used words to that effect. So you were trying to lock him into the framework that his words indicated, and I was just following his actual meaning (I think) and saying he used the wrong words to describe it.

[Fastracehorse]

With over 2,200 post...can you direct me to any one of those over 2,200 that made any sense?

Mike (Dr Beav)

....I meant the speed figure the Beyer.

fffastt

[BCOURTNEY]

My research has shown me that observations are best used to falsify possible candidate models and not to be used to produce candidate models. The remainder models after all the falsifications are performed or carried out is the solution, which might be a set of models.

[DeltaLover]

A multinomial logit model following the Bolton – Chapman approach does not appear to solve the game, at least in real world conditions. After several implementations of 'base models' starting from the one described in original Searching for Positive Returns paper and extending it to cover many more handicapping factors including their impact values, ROI and winning percentage the best models were close to break even applied to a universe of approximately 40,000 races.

I have also experiment with a Neural Networks and Genetic Algorithms to estimate the weights for a MLM with any more success than beating the take out....

Changing direction to a model that instead of showing ultimate profitability tries to predict the most frequent winner led me to somehow better results having the best models being very close to the crowd's opinion. The following published approaches are doing the same pretty much:

http://nguyendangbinh.org/Proceedings/IPCV08/Papers/ICA4043.pdf
http://www.wseas.us/e-library/conferences/2010/Iasi/NNECFS/NNECFS-21.pdf

Besides the failure of these approaches to solve the game I was able to derive a set of indexes that have statistical significance and I use regularly in my handicapping process which still remains a composite of both art and science and quite possibly it will remain like this forever....

[GameTheory]

Changing direction to a model that instead of showing ultimate profitability tries to predict the most frequent winner led me to somehow better results having the best models being very close to the crowd's opinion.Aren't all the models (logit and NN) just assigning probabilities (or equivalent) to the horses? Boltman/Chapman or Benter didn't make a model to directly predict profitability that I know of -- they all made models to predict winning chances, no?

I have not yet read the papers you posted, but I don't see the "change in direction" of the models? (You could change what you did with the output in terms of selecting bets, but the models would be the same.)

[DeltaLover]

Aren't all the models (logit and NN) just assigning probabilities (or equivalent) to the horses? Boltman/Chapman or Benter didn't make a model to directly predict profitability that I know of -- they all made models to predict winning chances, no?

I have not yet read the papers you posted, but I don't see the "change in direction" of the models? (You could change what you did with the output in terms of selecting bets, but the models would be the same.)


This is an extract from Boltman/Chapman 's paper:

[http://184.106.114.17/test_client/original_paper.jpg

The "change in direction" has to do with the fact that the original objective was to estimate the actual distribution of the winning probability of each starter in a race was replaced by a model focusing in the prediction of the (most frequent) winner which in most of the cases resembles the way the crowd will form the odds of each horse in the race.....

[GameTheory]

The "change in direction" has to do with the fact that the original objective was to estimate the actual distribution of the winning probability of each starter in a race was replaced by a model focusing in the prediction of the (most frequent) winner which in most of the cases resembles the way the crowd will form the odds of each horse in the race.....Not understanding the distinction. Let's talk about prediction of a single race -- there are no distributions or "most frequent" winners. There is one winner, and bunch of losers. Each of these models will assign a probability value to each horse in that race, and all the probabilities will sum to 1. Right?

So what's the difference? Are you simply saying that the (newer) model is calibrated to get the most winners out of the top-ranked horse in each race without regard to keeping the other probabilities in line?

[DeltaLover]

Are you simply saying that the (newer) model is calibrated to get the most winners out of the top-ranked horse in each race without regard to keeping the other probabilities in line?

This is correct.

While the first approach is ROI oriented and results to a betting strategy based in overlays with no respect to winning frequency the second is relaxing the ROI requirement trying to increase the hit rate. This second approach has no betting value since it usually picks the obvious horses that happen to be underlays but it can be used as an indicator that a specific method (MLM, NN, GA) can predict the way the public is betting....

[TrifectaMike]

This is correct.

While the first approach is ROI oriented and results to a betting strategy based in overlays with no respect to winning frequency the second is relaxing the ROI requirement trying to increase the hit rate. This second approach has no betting value since it usually picks the obvious horses that happen to be underlays but it can be used as an indicator that a specific method (MLM, NN, GA) can predict the way the public is betting....

In the first example the dependent variable is a the natural log of the odds of unknown binomial probabilities modeled as a linear function of the independent variables, and nothing more. ROI and betting strategy doesn't enter the picture. Now how you use the predicted binomial probabilities is another subject. Logistic and ordinary regression only differ by a transformation and the min/max method used to estimate the coefficients.

You might be confusing utilization with regression.

Mike (Dr Beav)

[GameTheory]

This is correct.

While the first approach is ROI oriented and results to a betting strategy based in overlays with no respect to winning frequency the second is relaxing the ROI requirement trying to increase the hit rate. This second approach has no betting value since it usually picks the obvious horses that happen to be underlays but it can be used as an indicator that a specific method (MLM, NN, GA) can predict the way the public is betting....Gotcha, although I have found that maximizing winners is also the best way to make a full betting line (with some calibration).

[DeltaLover]

In the first example the dependent variable is a the natural log of the odds of unknown binomial probabilities modeled as a linear function of the independent variables, and nothing more. ROI and betting strategy doesn't enter the picture.


That's true..

The point I am trying to make is that the proposed overlays based in their derived probabilities can not lead to a positive ROI betting strategy....

[lansdale]

This is an extract from Boltman/Chapman 's paper:

[http://184.106.114.17/test_client/original_paper.jpg

The "change in direction" has to do with the fact that the original objective was to estimate the actual distribution of the winning probability of each starter in a race was replaced by a model focusing in the prediction of the (most frequent) winner which in most of the cases resembles the way the crowd will form the odds of each horse in the race.....

Delta,

There's no indication here that you've read Benter's original article, which explains his methods in some detail. He never 'changed direction' and he never was trying to model 'most frequent' winners. This is your misinterpretation of what you've read. Game Theory (as usual) has it right. Benter's logit (later probit) model, suggested by the paper of Bolton and Chapman, is intended to predict the outcome of a race. He projects a win probability for each horse, and bets the overlays, exactly like any other bettor. His ROI was ca. 28% and the fortune he won is well documented.

Cheers,

lansdale

[dnlgfnk]

...for stating what I have been thinking, but was wary of offering due to my inferior mathematical background.
From all reports, the guy extracted data from public statistical records (except for 3% visual input) with an obsessive,all consuming drive and, what's most interesting to me, won millions with a comprehensive approach. I had always assumed a winner must be a "specialist", ala "Charlie" in Beyer's "My 50k Year...".
It seems he is trying to be improved upon with theory and mathematical terms. It's horseracing.

[GameTheory]

There's no indication here that you've read Benter's original article, which explains his methods in some detail. He never 'changed direction' and he never was trying to model 'most frequent' winners. This is your misinterpretation of what you've read. Game Theory (as usual) has it right. Benter's logit (later probit) model, suggested by the paper of Bolton and Chapman, is intended to predict the outcome of a race. He projects a win probability for each horse, and bets the overlays, exactly like any other bettor. His ROI was ca. 28% and the fortune he won is well documented.
Thanks for the compliments, but in all fairness Delta didn't even mention Benter or say that that was what Benter was doing. He said it was what he (Delta himself) was doing.

[DeltaLover]

Delta,

There's no indication here that you've read Benter's original article, which explains his methods in some detail. He never 'changed direction' and he never was trying to model 'most frequent' winners. This is your misinterpretation of what you've read. Game Theory (as usual) has it right. Benter's logit (later probit) model, suggested by the paper of Bolton and Chapman, is intended to predict the outcome of a race. He projects a win probability for each horse, and bets the overlays, exactly like any other bettor. His ROI was ca. 28% and the fortune he won is well documented.

Cheers,

lansdale

Sorry, but before you make such bold statements you have to be more careful. Please reread the thread and try to understand what I am talking about....

[lansdale]

Sorry, but before you make such bold statements you have to be more careful. Please reread the thread and try to understand what I am talking about....

Delta,

Let me caveat this by saying that, pace the benign poster above, I have no advanced training in math, but hopefully you can bear with me.

You claim in your first post that the MLM model per Chapman and Bolton, cannot 'solve the game' in the real world - I assume this means produce a profit. That does sound like you're implying not only that you were unable to make it work, but that no one could. My apologies, in that it was GT and not you who mentioned Benter, but doesn't his achievement and that of the other Hong Kong betting teams, as well as those now using MLM in this country contradict your claim?

My apologies also in misunderstanding that the use of 'change of direction' and 'frequency of winners' referred only to your own work. But you also seem to imply that these NN models were unable to produce a profit, which is very much in line with what I've heard about them for the past two decades.

Cheers,

lansdale

[Native Texan III]

Delta,

Let me caveat this by saying that, pace the benign poster above, I have no advanced training in math, but hopefully you can bear with me.

You claim in your first post that the MLM model per Chapman and Bolton, cannot 'solve the game' in the real world - I assume this means produce a profit. That does sound like you're implying not only that you were unable to make it work, but that no one could. My apologies, in that it was GT and not you who mentioned Benter, but doesn't his achievement and that of the other Hong Kong betting teams, as well as those now using MLM in this country contradict your claim?

My apologies also in misunderstanding that the use of 'change of direction' and 'frequency of winners' referred only to your own work. But you also seem to imply that these NN models were unable to produce a profit, which is very much in line with what I've heard about them for the past two decades.

Cheers,

lansdale

No mathematical averaging method can beat the races.
Racing prediction is about small differences in individual races, not averages from past data which even when printed in black and white no two people can quite agree upon - speed ratings for example. With the data for today's race, the missing or unknown data and its quality means the problem of difference and variability is further magnified.
Wrong data, wrongly weighted data and prediction and prices are wrong -GIGO.
Successful prediction only applies to today's individual event.
Each of today's events go to the long term - not the other way around.
You can only bet today's races not yesterday's.
Successful profit is being a bit better at that than the general public over the long term.

Benter, nor anyone else, can make averaging work for those very reasons.
The method does not fit the real problem to be solved and never did.
(The current financial crisis was largely caused by reliance on "sophisticated" models that did not model reality but the average of the past - no causal understanding, no understanding of difference).

Where the averaging methods have merit is for odds line usage where the average for the long term - a season of races say has some applicability to averaging.

Even then Benter found that despite the years of effort the "ignorant public" could come up with a far better predictive odds line than he could.

What he did have was a methodical monopoly in a huge betting market (with just 2 race meetings per week) of making the average right decision long term far better than the public flipping their prejudices from race to race. As other "me too" teams by the dozen do the same thing today, it means there is zilch advantage left and today's more successful bettors use different methods based more on difference.

[TrifectaMike]

No mathematical averaging method can beat the races.

I'm no great fan of averaging. However, I have to disagree with your statement. Averaging is not always a bad "thing".

Let me setup a toy race, and three players.

I am the racing god, and set the true probabilities of the horses winning their races. And in this toy racing game we have three players. Players A and B are excellent handicappers. Player C is agnostic in his selections, and not a handicapper.

All three players are aware of each others contenders and probability estimates.

Player A is always overconfident in his probability estimates. Player B is always more conservative in his probability estimates.

Player C is agnostic, but has some basic understanding of probability, so he assigns a probability of winning, 1/n, for each horse.

After several races, I tell Player C that Player A's probability estimates are nearly always higher than they should be on the contenders. And I also inform Player C that Player B's probability estimates are nearly always lower than they should be on the contenders.

Player C takes this information and drops his estimates of 1/n and instead averages Players A and Players B's probability estimates on their contenders.

I leave it to you to answer, who the big winner will be.

So, averaging in general is not helpful, but useful when properly utilized.

Racing prediction is about small differences in individual races, not averages from past data which even when printed in black and white no two people can quite agree upon - speed ratings for example. With the data for today's race, the missing or unknown data and its quality means the problem of difference and variability is further magnified.
Wrong data, wrongly weighted data and prediction and prices are wrong -GIGO.
Successful prediction only applies to today's individual event. And the next.
Each of today's events go to the long term - not the other way around.
You can only bet today's races not yesterday's. Do you believe that tomorrows race can give one information on a race run yesterday? I do.
Successful profit is being a bit better at that than the general public over the long term.

Benter, nor anyone else, can make averaging work for those very reasons.
The method does not fit the real problem to be solved and never did.
(The current financial crisis was largely caused by reliance on "sophisticated" models that did not model reality but the average of the past - no causal understanding, no understanding of difference).

Where the averaging methods have merit is for odds line usage where the average for the long term - a season of races say has some applicability to averaging.

Even then Benter found that despite the years of effort the "ignorant public" could come up with a far better predictive odds line than he could. I would not say the public had a better line than Benter. I would say that Benter's model Logistic or Probit had no ability to update his probabilities with new information, so he regressed on his line and the public's line to make use of additional information.

What he did have was a methodical monopoly in a huge betting market (with just 2 race meetings per week) of making the average right decision long term far better than the public flipping their prejudices from race to race. As other "me too" teams by the dozen do the same thing today, it means there is zilch advantage left and today's more successful bettors use different methods based more on difference.

Mike (Dr Beav)

[Cratos]

I'm no great fan of averaging. However, I have to disagree with your statement. Averaging is not always a bad "thing".

Let me setup a toy race, and three players.

I am the racing god, and set the true probabilities of the horses winning their races. And in this toy racing game we have three players. Players A and B are excellent handicappers. Player C is agnostic in his selections, and not a handicapper.

All three players are aware of each others contenders and probability estimates.

Player A is always overconfident in his probability estimates. Player B is always more conservative in his probability estimates.

Player C is agnostic, but has some basic understanding of probability, so he assigns a probability of winning, 1/n, for each horse.

After several races, I tell Player C that Player A's probability estimates are nearly always higher than they should be on the contenders. And I also inform Player C that Player B's probability estimates are nearly always lower than they should be on the contenders.

Player C takes this information and drops his estimates of 1/n and instead averages Players A and Players B's probability estimates on their contenders.

I leave it to you to answer, who the big winner will be.

So, averaging in general is not helpful, but useful when properly utilized.



Mike (Dr Beav)

Mike, what you gave is an example of Arrow’s Paradox where Player A is risk-preferred, Player B is risk-against, and Player C is risk neutral. It is in my opinion not the averages that changed the game, but the risk profiles of the players.

[TrifectaMike]

Mike, what you gave is an example of Arrow’s Paradox where Player A is risk-preferred, Player B is risk-against, and Player C is risk neutral. It is in my opinion not the averages that changed the game, but the risk profiles of the players.

I'm well aware of Arrow's Paradox, and this is not it. I set the game. I provided the information. Averaging is the difference in this toy game.

In any case, the point is averaging is not always bad, just most of the time.

Mike (Dr Beav)

[dnlgfnk]

Native...

Does the phrase "lightning in a bottle" then sum up Benter's accomplishments? If so, I'm comfortable with being back to the assumption that "no computer can beat the races", even if that's because of too many copycats.

He stated his doubts that successful computer handicapping could be profitable in the future, especially if one weren't in the front of the line at a particular time and place.

At a 24% ROI, I had always suspected it was more the massive pools than the technology. I've been following his story and the next generation's attempts to somewhat duplicate his success, but time to move on.

I wish Beyer would elaborate on his trip handicapping friend "Charlie", if he were able. He is the true, mysterious, prototypical professional that holds knowledge known by the few, assuming Andy hasn't romanticized Charlie's persona a bit.

[Cratos]

I'm well aware of Arrow's Paradox, and this is not it. I set the game. I provided the information. Averaging is the difference in this toy game.

In any case, the point is averaging is not always bad, just most of the time.

Mike (Dr Beav)
You cannot set the game against someone's else emotions in a fair game. I understand your assertion about averaging is not always bad and I agree, but I don't think its application is applicable in your example.