What would you want to have done...

[TrifectaMike]

Originally Posted by Dave Schwartz

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]

Originally Posted by sjk

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]

Originally Posted by TrifectaMike

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]

Originally Posted by yak merchant

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]

Originally Posted by TrifectaMike

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]

Originally Posted by Jeff P

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]

Originally Posted by 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.

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

[Cratos]

Originally Posted by TrifectaMike

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]

Originally Posted by 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]

Originally Posted by Cratos

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]

Originally Posted by TrifectaMike

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]

Originally Posted by Cratos

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]

Originally Posted by TrifectaMike

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.