I am trying to work out a decent way of changing rating assessments to odds. Does anyone know a good way of doing this? Some programs do this, but I would like to understand how they go about it. I think the relation between ratings and odds is not linear, but logarithmic and is influenced by the number of runners. Anyone have ideas? Bookmakers do use ratings to assess horses,but i don't think they set markets this way.

Hi when you first structure your Variable point scores, you need to make sure you standardise the scores so that when you combine them they have the same impact from Variable to variable ie. that is if you use an Indexing score of 100 then each point should equate to 1 % point if you havn't thought of doing that, then your whole point scoreing method is skewed.
The other method if using only one Variable is by using the Mean & Standard deviation of each horses life to date history and then finding the Winners Mean score for todays Grade. if you look up Zscore tables on the net you will get a good idea to accuratly assess the Payoff for anything.
The idea is to find the performance envolope of each horse,if you think you find one score or total then thats wrong, look at how standard deviation works and you will get an idea on how a horses form bounces along..

I outlined DonScott's method in this topic:

http://www.smartgambler.com.au/foru...=3024&forum=2&1

It uses a table converting rating differences to relative chances. From that you can easily verify the relationship is not linear.

As for the number of runners, obviously if you scratch a runner with probability p, you merely divide the remaining probabilities by 1-p so that the total is again 1.

his weight for advantage tables were very inefficent..
If you test his method over a large sample it falls down badly..

http://www.smartgambler.com.au/foru...=2745&forum=2&8

Hopefully this time I have the link to the earlier discussion on ratings to odds right.

Woof I'll try to reply to you once I re-read your stuff, as I still don't understand what you're trying to say.

dear woof,I am considering using a single rating number. Even when these are standardised to a number out of 100 this number is not necessarily the percentage chance the horse has of winning. this is precisely the problem. Many different ratings are available. Don Scott appears to use a weight rating which is converted to a probability using a particular scale. If I wanted to, for example change techform ratings to a probability or a price, it is not obvious how to do this. The idea would be to use this price to determine value-Value bets would occur if the available price was significantly greater than the calculated price. Most punters do this by intuition- I want to use maths.

If you want to use Techform ratings or whatever then merely compile your own table from past results.

The first column is rating difference between A and B.

The second is the proportion of times B beats A.

This gives you relative chances which you can apply to any number of runners.

Hi thevig,
the reason i mentioned standardising the data is that you can place an accurate score against every start a horse has, then using the Mean & Standard deviation of each horses score you can then develop an Oddsline.
so to find the z score of a two horse race, you would do the following. zscore=HorseA (mean score)-(minus) HorseB(mean score)/(divided by Horse B(standard deviation) you then look up the zscore on any zscore table on the internet this is the probability that horse B would beat horse A , if horse A ran an average race.
you then reverse the Mean scores and use Horse A standard deviation to find the probability of Horse A beating horse B if horse B ran an average race.

[ This Message was edited by: woof43 on 2003-03-05 17:56 ]

Woof,

Having now spent some time looking at z scores and their alleged relevance to team ratings I still have more work to do. However I'm convinced you're barking up the wrong tree.

The idea of taking the mean and stddev of a horse's past performance ratings is nonsense. For starters, typically performances peak after a number of runs into a campaign so your method would overrate 1st ups and those short of their peak distance range.

As for Techform (published freely on UNiTAB) which gives their favourite 100 in all races there is no sane way you can average a horse's Techform past performances. Like it or not you have little choice but to collate Techform differences with relative win chances, as I outlined.

yes you raise a valid point regarding Fitness, The Mean gives you a starting point, you make slight adjustments to the Mean based on past analysis of your population(database)or if you have enough specific data on each horse you would use its own data.
While your looking at z scores the other relevant area for horses, is to describe its distribution pattern by that you look at the Kurtosis of its data set.

The only way to assign ratings and the value thereof under varying scenarios, is to have a significant database with reliable statistics.

I have just got myself a database and you would be surprised the relative importance of some factors on profit, not strike rate have.
Similarly, you'd be very surprised how relatively insignificant some data is to filter losing bets.

E.G. Barrier position has very little impact on a large sample of data. Now whether you want to filter it to different scenarios, like front runners and distance is another thing, but overall I'm speaking, barriers have a minute impact on success or failure.

Database is the only way to go if you want to take this seriously.

Good Luck.

[ This Message was edited by: Chrome Prince on 2003-03-06 23:13 ]

Chrome: I am assuming that the rating I have includes a consideration of all of the important factors influencing the result which have been used to create a numerical rating. I want to convert ratings to odds. A mean rating and Z score does not convert the rating to a probability of the horse winning, rather it gives the probability that a numerical rating is exceeded. It is not clear how the difference rating scores connected to the frequency of the number of times A beats B gives the probability that a wins. Maybe A and B have never raced.

Vig,

although you've managed to figure out that the other methods are unsuitable, why do you insist on ignoring my method which clearly works and I believe is the only practical approach?

Now, much as I despise Z-scores they do contain some theoretical beauty.

Woof appears to be the only one besides me who realises that Z-scores CAN be used to calculate the probability that A beats B. If A and B are Normal, then so is "A-B", the distribution of every possible matchup of A vs B.

Trouble is I believe it's highly unlikely in typically chaotic races that ratings are Normal.

Hi there, I gave an example of A vs B, but this isn't the method I employ, an idea is to use your database to find the avg winner score per Class of race, then find the avg Stdevep of each winner in each Class. This will provide you with the avg score each winner needs to acheive in each Class. Then instead of using B when calculating A z-score your substitute it for your claculated avg winner score.