Vote Me Out ?

[MySportsAI]

I have voted for you to leave. If you find starting threads like this necessary you are effectively self-reporting that it isn't the best place for you. Generally I've found your posting style to be quite terse and edgy which isn't in keeping with the vast majority of what I read on here and what had made the place such a valuable offering in the past.

I may be reading a small sample size and doing you a disservice as I'm not contributing like I was earlier in the year but I appreciate that others could find great value in your contributions so please don't take my viewpoint as representative of anyone else.

There is absolutely no need for this garbage, nobody else has projected themselves like this and nobody needs to. I hope alcohol hasn't played some part in your starting of this thread.

Great summary above DuckandDive probably puts it better than I can.

I appreciate your honesty

 

[MySportsAI]

As Austin said you’re an alcoholic who needs help, but in all seriousness add attention seeker to the poll we will all click that and then move on.

That is a very serious accusation from someone who clearly knows nothing about alcoholism, click on the vote and shut up

 

[Graeme]

I think you need to reflect on your interactions and be grateful I did not report some of them

OK - you have your right to respond and I will not respond further to you in any way hence forward - good luck with what ever you choose to do!

 

[Marshall]

That is a very serious accusation from someone who clearly knows nothing about alcoholism, click on the vote and shut up

Voted stay for the banter, cheers.

 

[ARAZI91]

Looking at your trainer model, observation bearing in mind B at GSCE is my maths level, the highest trainer model % is 30% so equating to 9/4 ish. Could your model be improved with expanding the range of the trainer model % to therefore trainers model odds for comparison with b365, are they too compacted ? Or am I talking nonsense ?

What your describing here O Outlander was once seen as "state of the art" in the syndicate betting world but has now become par for the course.
The mathematics at best are high school level and involve combining a) the logit transformed probabilities of a fundamental model with b) the betting public's estimation or "market" (again transformed as logits). Optimising both for the correct weighting is usually done by maximising the log likelihood function over a set of past and also hold out races. Benter discusses the process in his seminal paper but the origins of the idea is as old as the hills and has it's roots in physics and statistical mechanics and is known as the "softmax function" - combining two or more probability distributions is nothing new.
Was probably an elementary step for Benter given that the choice of model by those HK pioneers was logistic regression, later upgrading to probit type regression. The mainly Bayesian "state space" models used mostly nowadays by syndicates would make the Benter's/Woods & Bob Moore type models seem very rudimentary. That's not to say Benter was not a genius - it's rumoured that his latter probit models in the early 90's in HK had 14 different "jockey" auxiallary models each with up to 15 variables. I also have a rare Benter paper where he puts forward a very elegant solution to modelling "distance preferance" in thoroughbreds using a least-squares parabolic curve fitting method involving Hill-Keller's "optimal running" theory and the use of mathematical Tack Points and Biezer Curves.

The basic mathematics

logit transformation

then becomes



combined using the exponential distribution and normalised



Log Likelihood function then maximised on past/hold out races
at that time the model of choice was logistic regression where 1 = winner and 0 = otherwise



Benter moving to Probit meant that he could condition on say "normalised finishing position" or "target speed rating" rather than the binary outcome of winner/otherwise.


some rare(ish) Benter

 

[MySportsAI]

What your describing here O Outlander was once seen as "state of the art" in the syndicate betting world but has now become par for the course.
The mathematics at best are high school level and involve combining a) the logit transformed probabilities of a fundamental model with b) the betting public's estimation or "market" (again transformed as logits). Optimising both for the correct weighting is usually done by maximising the log likelihood function over a set of past and also hold out races. Benter discusses the process in his seminal paper but the origins of the idea is as old as the hills and has it's roots in physics and statistical mechanics and is known as the "softmax function" - combining two or more probability distributions is nothing new.
Was probably an elementary step for Benter given that the choice of model by those HK pioneers was logistic regression, later upgrading to probit type regression. The mainly Bayesian "state space" models used mostly nowadays by syndicates would make the Benter's/Woods & Bob Moore type models seem very rudimentary. That's not to say Benter was not a genius - it's rumoured that his latter probit models in the early 90's in HK had 14 different "jockey" auxiallary models each with up to 15 variables. I also have a rare Benter paper where he puts forward a very elegant solution to modelling "distance preferance" in thoroughbreds using a least-squares parabolic curve fitting method involving Hill-Keller's "optimal running" theory and the use of mathematical Tack Points and Biezer Curves.

The basic mathematics

logit transformation
View attachment 172111
then becomes

View attachment 172112

combined using the exponential distribution and normalised

View attachment 172113

Log Likelihood function then maximised on past/hold out races
at that time the model of choice was logistic regression where 1 = winner and 0 = otherwise

View attachment 172114

Benter moving to Probit meant that he could condition on say "normalised finishing position" or "target speed rating" rather than the binary outcome of winner/otherwise.


some rare(ish) Benter

View attachment 172115
View attachment 172116

View attachment 172117

Still works today

 

[ArkRoyal]

I usually only ban spammers when they attempt to join or those that appear to want more then one username.
As has been suggested earlier in this thread the forum has an ignore feature which can be useful if you don't want to see what a particular member is posting. If reading other members posts wind you up then "ignore" them and enjoy the rest of the forum.

 

[MySportsAI]

I usually only ban spammers when they attempt to join or those that appear to want more then one username.
As has been suggested earlier in this thread the forum has an ignore feature which can be useful if you don't want to see what a particular member is posting. If reading other members posts wind you up then "ignore" them and enjoy the rest of the forum.

and what about abusive language

 

[tony spencer]

What your describing here O Outlander was once seen as "state of the art" in the syndicate betting world but has now become par for the course.
The mathematics at best are high school level and involve combining a) the logit transformed probabilities of a fundamental model with b) the betting public's estimation or "market" (again transformed as logits). Optimising both for the correct weighting is usually done by maximising the log likelihood function over a set of past and also hold out races. Benter discusses the process in his seminal paper but the origins of the idea is as old as the hills and has it's roots in physics and statistical mechanics and is known as the "softmax function" - combining two or more probability distributions is nothing new.
Was probably an elementary step for Benter given that the choice of model by those HK pioneers was logistic regression, later upgrading to probit type regression. The mainly Bayesian "state space" models used mostly nowadays by syndicates would make the Benter's/Woods & Bob Moore type models seem very rudimentary. That's not to say Benter was not a genius - it's rumoured that his latter probit models in the early 90's in HK had 14 different "jockey" auxiallary models each with up to 15 variables. I also have a rare Benter paper where he puts forward a very elegant solution to modelling "distance preferance" in thoroughbreds using a least-squares parabolic curve fitting method involving Hill-Keller's "optimal running" theory and the use of mathematical Tack Points and Biezer Curves.

The basic mathematics

logit transformation
View attachment 172111
then becomes

View attachment 172112

combined using the exponential distribution and normalised

View attachment 172113

Log Likelihood function then maximised on past/hold out races
at that time the model of choice was logistic regression where 1 = winner and 0 = otherwise

View attachment 172114

Benter moving to Probit meant that he could condition on say "normalised finishing position" or "target speed rating" rather than the binary outcome of winner/otherwise.


some rare(ish) Benter

View attachment 172115
View attachment 172116

View attachment 172117

What your describing here O Outlander was once seen as "state of the art" in the syndicate betting world but has now become par for the course.
The mathematics at best are high school level and involve combining a) the logit transformed probabilities of a fundamental model with b) the betting public's estimation or "market" (again transformed as logits). Optimising both for the correct weighting is usually done by maximising the log likelihood function over a set of past and also hold out races. Benter discusses the process in his seminal paper but the origins of the idea is as old as the hills and has it's roots in physics and statistical mechanics and is known as the "softmax function" - combining two or more probability distributions is nothing new.
Was probably an elementary step for Benter given that the choice of model by those HK pioneers was logistic regression, later upgrading to probit type regression. The mainly Bayesian "state space" models used mostly nowadays by syndicates would make the Benter's/Woods & Bob Moore type models seem very rudimentary. That's not to say Benter was not a genius - it's rumoured that his latter probit models in the early 90's in HK had 14 different "jockey" auxiallary models each with up to 15 variables. I also have a rare Benter paper where he puts forward a very elegant solution to modelling "distance preferance" in thoroughbreds using a least-squares parabolic curve fitting method involving Hill-Keller's "optimal running" theory and the use of mathematical Tack Points and Biezer Curves.

The basic mathematics

logit transformation
View attachment 172111
then becomes

View attachment 172112

combined using the exponential distribution and normalised

View attachment 172113

Log Likelihood function then maximised on past/hold out races
at that time the model of choice was logistic regression where 1 = winner and 0 = otherwise

View attachment 172114

Benter moving to Probit meant that he could condition on say "normalised finishing position" or "target speed rating" rather than the binary outcome of winner/otherwise.


some rare(ish) Benter

View attachment 172115
View attachment 172116

View attachment 172117

Thanks, that’s really helpful.
I’m definitely not claiming to have invented anything new or state of the art. I’m building this in stages and trying to make sure each part stands up before I combine everything.
At the moment I’m treating the trainer percentage as a trainer-strength signal, not as the final tissue price for the horse. The next stage I had in mind was to test it on the 2026 results and see whether the percentages are calibrated, too compact, or too stretched.
So I’ll be looking at trainer model rank, trainer percentage bands, actual win/place rates, SP/BFSP returns, and how the Bet365 price moves from evening to morning. That should tell me whether the market already has most of it covered, or whether there is still something useful in the trainer signal.
Your point about combining my model with the market properly, rather than just using a rough weighting, makes sense. I’m not at that stage yet, but I can see that is the right direction once the individual signals have been tested.
So thanks — I appreciate the explanation. The aim is to prove the trainer/form/pace signals separately first, then look at the proper way of blending them with the market.

---

 

[MySportsAI]

Com

Thanks, that’s really helpful.
I’m definitely not claiming to have invented anything new or state of the art. I’m building this in stages and trying to make sure each part stands up before I combine everything.
At the moment I’m treating the trainer percentage as a trainer-strength signal, not as the final tissue price for the horse. The next stage I had in mind was to test it on the 2026 results and see whether the percentages are calibrated, too compact, or too stretched.
So I’ll be looking at trainer model rank, trainer percentage bands, actual win/place rates, SP/BFSP returns, and how the Bet365 price moves from evening to morning. That should tell me whether the market already has most of it covered, or whether there is still something useful in the trainer signal.
Your point about combining my model with the market properly, rather than just using a rough weighting, makes sense. I’m not at that stage yet, but I can see that is the right direction once the individual signals have been tested.
So thanks — I appreciate the explanation. The aim is to prove the trainer/form/pace signals separately first, then look at the proper way of blending them with the market.

---

Combine them using regression but dont normalize the fundamental (ie your trainer prob's) at stage 1

 

[tony spencer]

Com

Combine them using regression but dont normalize the fundamental (ie your trainer prob's) at stage 1

Thanks, that makes sense.

So rather than treating the trainer percentage as a finished tissue price, I should keep it as a raw trainer-strength signal at the first stage.

Then the next step is to combine that signal with the other factors — form rating, pace/draw/setup, market price and price movement — and let the regression work out how much weight each part deserves.

That also avoids me manually stretching or compressing the trainer percentages too early. The important test is whether the trainer signal adds value once it is combined properly with the market and the other race factors.

So I take your point as:

Stage 1: keep the trainer/fundamental signal raw
Stage 2: combine it with the market and other signals using regression
Stage 3: calibrate/normalise the final race probabilities after that
That is very helpful and gives me a clearer route for the 2026 testing.

 

[MySportsAI]

Thanks, that makes sense.

So rather than treating the trainer percentage as a finished tissue price, I should keep it as a raw trainer-strength signal at the first stage.

Then the next step is to combine that signal with the other factors — form rating, pace/draw/setup, market price and price movement — and let the regression work out how much weight each part deserves.

That also avoids me manually stretching or compressing the trainer percentages too early. The important test is whether the trainer signal adds value once it is combined properly with the market and the other race factors.

So I take your point as:

Stage 1: keep the trainer/fundamental signal raw
Stage 2: combine it with the market and other signals using regression
Stage 3: calibrate/normalise the final race probabilities after that
That is very helpful and gives me a clearer route for the 2026 testing.

Yes

 

[MySportsAI]

Thanks, that makes sense.

So rather than treating the trainer percentage as a finished tissue price, I should keep it as a raw trainer-strength signal at the first stage.

Then the next step is to combine that signal with the other factors — form rating, pace/draw/setup, market price and price movement — and let the regression work out how much weight each part deserves.

That also avoids me manually stretching or compressing the trainer percentages too early. The important test is whether the trainer signal adds value once it is combined properly with the market and the other race factors.

So I take your point as:

Stage 1: keep the trainer/fundamental signal raw
Stage 2: combine it with the market and other signals using regression
Stage 3: calibrate/normalise the final race probabilities after that
That is very helpful and gives me a clearer route for the 2026 testing.

Dont forget to natural log them before stage 2