likelihood need help

[wkcalvin]

Hi All,

I am using SPSS developing a MLR model of racing, I have ask some very STUPID questions at here before, but I need to say great thanks to you all for your helpful replies.

Now my problem is MAXIMUM LIKELIHOOD for my model, I have checked many paper but still do not know to start it.

Since BENTER and BOLTON mentioned the max likelihood is important to measure the accuraces of model..................

Can any one teach me how to calculate it with a life sample of Horse Racing ???

I can not undertand the symbols of the papers, can any one use a simple example of horse racing to tell me how ????

Many thanks in advance............

CW

[robert99]

MW,

You may find that the link below explains things in plainer English

http://statgen.iop.kcl.ac.uk/bgim/mle/sslike_4.html

Actually it is far easier to say these things than carry them out in real life.

For horse racing, MLR is essentially trying to confirm the best estimate of true race odds based on the weightings of variables you have used in your model. Unless you are talking about coins and dice, true odds are never known but you do have a reasonable estimate in what the public bet the odds down to.

The numerical way (there are others) of finding MLR is by iteration - this means that you change one variable weighting at a time by tiny amounts and see what the result is on your prediction (better or worse). It is a suck it and see method. If you are using the public odds, then you will see what is the "best" value of your variable that fits in with the odds. It might never fit exactly but has a value that fits as best it can. Repeat that with all variables. You now have a better set of weightings. Actually it has only averaged things out for the test data you used so don't be fooled.

The next thing to do is to change weightings for two variables at a time and from a 3 dimensional plot see where the peak lies i.e. there will be a point where variable one and variable two give the best (nearest) result. Excel etc can do this for you. For 5 variables that will be 10 plots. You then adjust your weightings for combination pairs.

You can take it into higher dimensions (3, 4, .. 120 variables at a time) but you quickly lose sense of what is going on. It is more important to find a few strong forecasting variables that are independent of each other than use hundreds which all say the same thing. Suggest you use your best 5 and take it from there - more variables confuse things more than they help. Only add more variables, one at a time, if they really help things.

You will not reach perfection as each horse race is an independent event, never repeated, but you may get close. The real objective is to see if you can predict true odds better than the market and for that just test your tuned model on a few weeks / months race results. If it is better than the public's odds then either accept things or combine this test set with your original data and find a "better" fitting weightings average. The more you do this the nearer you will get to the average public's odds (self defeating) so then it might be better to have a specific model for specific race types.

[wkcalvin]

HI Rob,

Thanks to your reply very much

I agree your points on MLR, BUT.......beside MLR, I can't find another way to beat the races.

And do you have any new insight ?

[traynor]

MW,

You may find that the link below explains things in plainer English

http://statgen.iop.kcl.ac.uk/bgim/mle/sslike_4.html

Actually it is far easier to say these things than carry them out in real life.

For horse racing, MLR is essentially trying to confirm the best estimate of true race odds based on the weightings of variables you have used in your model. Unless you are talking about coins and dice, true odds are never known but you do have a reasonable estimate in what the public bet the odds down to.

The numerical way (there are others) of finding MLR is by iteration - this means that you change one variable weighting at a time by tiny amounts and see what the result is on your prediction (better or worse). It is a suck it and see method. If you are using the public odds, then you will see what is the "best" value of your variable that fits in with the odds. It might never fit exactly but has a value that fits as best it can. Repeat that with all variables. You now have a better set of weightings. Actually it has only averaged things out for the test data you used so don't be fooled.

The next thing to do is to change weightings for two variables at a time and from a 3 dimensional plot see where the peak lies i.e. there will be a point where variable one and variable two give the best (nearest) result. Excel etc can do this for you. For 5 variables that will be 10 plots. You then adjust your weightings for combination pairs.

You can take it into higher dimensions (3, 4, .. 120 variables at a time) but you quickly lose sense of what is going on. It is more important to find a few strong forecasting variables that are independent of each other than use hundreds which all say the same thing. Suggest you use your best 5 and take it from there - more variables confuse things more than they help. Only add more variables, one at a time, if they really help things.

You will not reach perfection as each horse race is an independent event, never repeated, but you may get close. The real objective is to see if you can predict true odds better than the market and for that just test your tuned model on a few weeks / months race results. If it is better than the public's odds then either accept things or combine this test set with your original data and find a "better" fitting weightings average. The more you do this the nearer you will get to the average public's odds (self defeating) so then it might be better to have a specific model for specific race types.

Handicappers would do well to take heed of your comments about the number of variables to use in forecasting. One of the oddest quirks in decision making is the increased confidence/reduced accuracy of decisions produced using multiple factors.

The effect is similar to the psychological effects called "Bandwagon Effect" or "social proof." The basic idea is that if a lot of people are doing something, if the individual does it, there is less chance of looking foolish when it doesn't work. Similarly, the use of numerous "ratings" in software applications creates the illusion of consensus, often resulting in the same increase in confidence with the same associated decrease in accuracy.
Good Luck

[traynor]

HI Rob,

Thanks to your reply very much

I agree your points on MLR, BUT.......beside MLR, I can't find another way to beat the races.

And do you have any new insight ?

Read up on the Efficient Market Hypothesis. Then realize that almost everyone is using the same basic data (often from the same source) and tweaking it in various ways to "beat the races."

Just as the EMH posits the deficiencies in a closed information system, so should you realize that unless new elements are introduced into the system, pari-mutuel racing has exactly the same deficiencies.

Good Luck

[robert99]

Handicappers would do well to take heed of your comments about the number of variables to use in forecasting. One of the oddest quirks in decision making is the increased confidence/reduced accuracy of decisions produced using multiple factors.

The effect is similar to the psychological effects called "Bandwagon Effect" or "social proof." The basic idea is that if a lot of people are doing something, if the individual does it, there is less chance of looking foolish when it doesn't work. Similarly, the use of numerous "ratings" in software applications creates the illusion of consensus, often resulting in the same increase in confidence with the same associated decrease in accuracy.
Good Luck

Agree on that - possibly it is also the Puritan ethic that hard work (more and more data) is the only way. In reality, work smart rather than work hard can be more successful in this field. Start off simple and only add in extra data if it proves essential. That way you get to understand what you are doing and why - start off complex and you stay confused. Perhaps that's why so many smart economists continue to make such dumb financial forecasts.

[robert99]

HI Rob,

Thanks to your reply very much

I agree your points on MLR, BUT.......beside MLR, I can't find another way to beat the races.

And do you have any new insight ?

MW,

I only have old insights ;)

Keep it simple
Specialise in race types that appeal - forget the rest
Once you have thoroughly checked your model, believe in it, use it where the odds are in your favour
Above all, your model should be using factors that best predict which horse will run the fastest on the day (not yesterday).

[arkansasman]

Calvin,

I don't know if this is any consolation, but building a Benter type model is extremely hard and time consuming. I would estimate that there are only a few people in the world who can correctly build a Benter type logit horse racing model. If you have any universities or colleges that are close to you, I would suggest talking to the professors at those colleges.

John

[PriceAnProbability]

Handicappers would do well to take heed of your comments about the number of variables to use in forecasting. One of the oddest quirks in decision making is the increased confidence/reduced accuracy of decisions produced using multiple factors.

The effect is similar to the psychological effects called "Bandwagon Effect" or "social proof." The basic idea is that if a lot of people are doing something, if the individual does it, there is less chance of looking foolish when it doesn't work. Similarly, the use of numerous "ratings" in software applications creates the illusion of consensus, often resulting in the same increase in confidence with the same associated decrease in accuracy.
Good Luck

Profitability as a horseplayer is based on finding one key mistake, usually a large one rather than a small one, made by the public.

The lack of speed figures in the pre-Beyer era is a good example. The lack of power ratings in this era is another. Power ratings were not necessary to win until the speed figures were made available to the masses, as there is no need to work on getting two steps ahead of the public when one is already ahead by one and doesn't want to risk losing that edge.

[GameTheory]

The lack of power ratings in this era is another.What is your definition of a power rating? Isn't a power rating just a one-number encapsulation of all the handicapping factors that go into it? As such, isn't a power rating just a simple expression of handicapping done rather than a measurement of something new (like speed figures were because they involved calculating variants in a new way to better measure times)?

[traynor]

Agree on that - possibly it is also the Puritan ethic that hard work (more and more data) is the only way. In reality, work smart rather than work hard can be more successful in this field. Start off simple and only add in extra data if it proves essential. That way you get to understand what you are doing and why - start off complex and you stay confused. Perhaps that's why so many smart economists continue to make such dumb financial forecasts.

Calvinism is alive and well in the handicapping world. There must be some underlying characteristic that makes people either believe they don't deserve to win, or that winning is so easy it has to made complex and difficult to be acceptable.

One of the most profitable episodes of my life was while performing "comprehensive handicapping" of harness races, after I discovered that the majority of my betting choices could be made in three or four seconds a race, by focusing on one key factor and literally ignoring everything else.

Even better, because a number of the selections were "counter-intuitive" to more comprehensive methods, the mutuel prices were often substantially more than those of the more "intuitive" selections.
Good Luck

[singunner]

Calvin,

I don't know if this is any consolation, but building a Benter type model is extremely hard and time consuming. I would estimate that there are only a few people in the world who can correctly build a Benter type logit horse racing model. If you have any universities or colleges that are close to you, I would suggest talking to the professors at those colleges.

John
What exactly is a Benter-type model? I tried Google and Wiki but got nothing.

[arkansasman]

What exactly is a Benter-type model? I tried Google and Wiki but got nothing.

Sin,

Try googling "bill benter" or "william benter".

He wrote a chapter in "Efficiency of Racetrack Betting Markets" in which he gave an overview of a multinomial logit horse racing model that he was using in Hong Kong.

The book "EORBM" is sellling for $2000 and $2100 on Amazon the last time I looked. But you can get copies of the book for $200 at http://hypernormal.com

Hope this helps.

John

[PlanB]

Calvinism is alive and well in the handicapping world. There must be some underlying characteristic that makes people either believe they don't deserve to win, or that winning is so easy it has to made complex and difficult to be acceptable.

One of the most profitable episodes of my life was while performing "comprehensive handicapping" of harness races, after I discovered that the majority of my betting choices could be made in three or four seconds a race, by focusing on one key factor and literally ignoring everything else.

Even better, because a number of the selections were "counter-intuitive" to more comprehensive methods, the mutuel prices were often substantially more than those of the more "intuitive" selections.
Good Luck

Sir, I will nominate you for the Derek's "Consumate Humble Pie" award, given only after careful thought & a coin toss; sorry, Big Red, you got tails.

[robert99]

Sin,

Try googling "bill benter" or "william benter".

He wrote a chapter in "Efficiency of Racetrack Betting Markets" in which he gave an overview of a multinomial logit horse racing model that he was using in Hong Kong.

The book "EORBM" is sellling for $2000 and $2100 on Amazon the last time I looked. But you can get copies of the book for $200 at http://hypernormal.com

Hope this helps.

John

Bill Benter (and others) describe the details of the logit method in the Appendix of Nick Mordin's "Winning Without Thinking" (hardback only) - cost £25, or less.
The method is very simple to follow, with basic maths.

[arkansasman]

Bill Benter (and others) describe the details of the logit method in the Appendix of Nick Mordin's "Winning Without Thinking" (hardback only) - cost £25, or less.
The method is very simple to follow, with basic maths.

Robert

The best I remember, Mordin had both chapters by Chapman, Benter's chapter and John Kelly's chapter.

There are other chapters in "Efficiency of Racetrack Betting Markets" that I have found helpful and they are:

1)Estimating the Probabilities of the Outcomes of a Horse Race (Alternatives to the Harville Formulas) by H.S. Stern
2)Post Position Bias by S. Betton
3)Subjective Information and Market Efficiency in a Betting Market by Stephen Figlewski

John

[wkcalvin]

Calvinism is alive and well in the handicapping world. There must be some underlying characteristic that makes people either believe they don't deserve to win, or that winning is so easy it has to made complex and difficult to be acceptable.



One of the most profitable episodes of my life was while performing "comprehensive handicapping" of harness races, after I discovered that the majority of my betting choices could be made in three or four seconds a race, by focusing on one key factor and literally ignoring everything else.



Even better, because a number of the selections were "counter-intuitive" to more comprehensive methods, the mutuel prices were often substantially more than those of the more "intuitive" selections.

Good Luck





Calvinism--- a very interesting word.





I have work for MLR model for 2 years, no $ I gain from this approach, BUT, I have found something during development.



May be I said 'Criteria Betting'





If the horse met the following criteria, bet.......





1. Career starts less than 20

2. Not at extreme distance such as 1000/2000/22000, focus on 6F and 7F

3. Not rest more that 42 days

4. Regular track work with reasonable speed -- thin horse don't work too much, for horse over 1150 LB, more is better

5. NOT a favor



......you can add more.





I have used these criteria, not being rich, but gained some short vacations.

[hubcapchaser]

My friend there is no way to explain in simple terms what is found in the papers you mentioned, especially the letters you are reffering to.

The Maximum Likely hood portion of the paper refers to formulas that can only be worked out using statistal analysis software.

There are many steps you should take in your region of the world before you can even consider using such software, and many of them involve collating the right data using the right factors. This task alone is a huge undertaking that will take more than 12 months, even if you had a winning model in another region and were trying to duplicate the winning results elsewhere.

hope this helps

hub