MAKING A PAR FILE?

[HIGH ROLLER]

hi, about how many races must one do in order to come up with an average par for a particular class and distance at a track?

[Charlie Judge]

I use a minimum of 20 examples.

There is a great way to prove this:

Look at a series of times (or speed ratings) created by taking the means and standard deviations of a series of race results. Plot the standard deviation as a function of the sample size, and you will see the curve of the standard deviations has an inflection point at about 20 samples. After 20 samples, the standard deviation curve gets much flatter. I will calculate some values and post the graph in a future post. I did this work a few years ago, and I do not have the graph handy right now.

CAJ

[OTM Al]

20 is ok, but rule of thumb in the statistics world is actually about 40. More is always better of course and 40 is an awful lot when taking into account all the factors to come up with pars. I certainly wouldn't trust anything less than 20 though.

[Dave Schwartz]

Sorry to play devel's advocate here, but...

Where do you get 20 races for each class at a particular distance?


Last year at Belmont Park - How many races do you suppose they ran at 7f at Belmont Park that matched the following conditions?

fast track
claiming (any price)
3up or older


8! They ran EIGHT races! And that was spread over all class levels! How do you get 20 or 40 in each level?

Even at 6fur they only ran 13!

I contend that making pars cannot be done by simply "averaging." If you are doing averaging, I'd love to see where you are getting the sample size.


Regards,
Dave Schwartz

[GameTheory]

Originally posted by Dave Schwartz
Sorry to play devel's advocate here, but...

Where do you get 20 races for each class at a particular distance?


Last year at Belmont Park - How many races do you suppose they ran at 7f at Belmont Park that matched the following conditions?

fast track
claiming (any price)
3up or older


8! They ran EIGHT races! And that was spread over all class levels! How do you get 20 or 40 in each level?

Even at 6fur they only ran 13!

I contend that making pars cannot be done by simply "averaging." If you are doing averaging, I'd love to see where you are getting the sample size.

Well, you *can*, but they are not necessarily any good. I make class-based pars via non-linear regression (the only thing I use pars for is to make variants). First I throw out maidens, 2yr olds, and some other "unreliable" race types, and then the rest of them all get thrown together in the same batch (claiming, alw, different prices & purses, ages, sexes all together) to build a race time predictor. Then to get a par for a race, I just input all those variables (age, sex, purse, race type, non-winners restrictions etc) and it spits out a time (it can also spit out a standard error for the prediction, but I don't actually use that in my current system). This way I can get a par time for unique race conditions that are actually not in the sample (except for maidens & the other types that I throw out). So it really is based on some (unknown) grand average with various adjustments for each variable, and it is all automated, updated weekly.

Now your pain-stakingly hand-made pars are undoubtably more accurate, but the variants created by using mine have shown themselves to be good enough to be useful. (i.e. using the variants that I make with them is better than not using them)

[OTM Al]

Also, you can set your par not conditioned on the track, ie all 3 and up NW1X allowance at 1 mile on dirt regardless of track. Then the variant captures track differences and daily track conditions rather than just a single track's variation and will give you numbers that can be compared across tracks rather than just to a single track ala Beyer figs.

[Dave Schwartz]

GT,

You know, I would expect no less from an intelligent guy like you. Great idea!


Dave

[Charlie Judge]

Dave:

I dont make pars per se. I create par ranges using Jim Cramer's speed numbers for the race winner. I look at all races for the condition in the database going back to 1994, and calculate the median and the 25% and 75% quartile range. I look at the par ranges for speed and pace and create a 2-dimensional box bounded by the 25% - 75% interquartile range for pace and speed for the condition.

I agree an average is not the way to go. The median is much more helpful, and a range is even better, as is shows the distribution of winning efforts. A single number par doesn't really describe the data. It is interesting to look at the breadth of the range to determine how consistent the winning numbers are for each condition. Grade 1 races have the smallest range.. down to 2yo MCL Fillies which have the greatest variability.

My data is based on

track distance surface age sex class claiming price statebred

variables.


I will post some representative data later.

CAJ

[Dave Schwartz]

Charlie,

That would, of course, be a valid statistical approach. After having seen your work in the past, I would repeat what I said to Game Theory: From you I would expect no less than such quality.


Regards,
Dave Schwartz

[Tom]

I make Beyer pars for NYRA several other tracks. For the NYRA database, I combine Aqu-Bel-Sar for most races becase I am using a Beyer number and not a time. AndI use races going back to 1998, so I have some hefty sample sizes for some classes. I mix distances when it looks like they can be mixed. To see that, I do everything in Excel, so I print a line chart of the fig sin the sub sample and then put in a trend line and look to see where figs depart from the trend line and eliminate them. Sometimes, the lowest and highest figs are out of wack, sometimes only the highest. I split meets for 2 and 3 year old races, and am very careful about mixing one and two turn races, such as the one turn mile at Aqu and the two turn mile on the inner. Unless the graphs show a tight fit to the trend line. I usually use the median as the par fo rthe class, and then look for a high and low range of normal figs. Number above and below the range are races of interest, either slow or fast for the class. Agian, using the graphs to visually check the data points.
I recaluclate every so often, so my pars are changing. I find the pars for the 2 and 3 years olds can be very differenet from crop to crop.

[hurrikane]

Interesting thread.

I don't think pars per se are attainable any more with the way conditions are written and the way they move horses around to slot tracks. Just MHO.

I do like GT idea and do something like that now.

[OTM Al]

Game Theory-

As an advanced student in Economics, I'm curious what technique you used to come up with your non-linear model. Did you use maximum likelihood estimation or take a best fit approach over several different representative models or something else entirely?

[chickenhead]

I have a question for GT and anyone else. Does anyone include in their mechanical pars any projection type data, or have you found that either not worth it or just impossible to do?

[JackS]

IMO any type of projection without proof, is misleading and probably useless. Pars created this week will probably be a little different than last week or next week. The changing conditions of horses, weather and the track itself lead us to use "reasonable" pars created from horses in races long since run. My idea is to use last years complete season to create a set of pars, then use a minor adjustment (fast or slow) in an attempt to equalize the races were playing today. I'm with OTM AL, once you have created a set of pars, create one number that can be transported from track to track. Using this method the neccesity of maintaining volumns of pars can be elminated. Maintainance of this simple set can be weekly, monthly or yearly. I prefer yearly as I doubt slight daily or weekly variations to be proof of anything other than a normal circumstance. Anyone making their own pars will have a set of par tables that only resemble someone elses set. Trust and live with your own ideas and tables. They are probably as good as and maybe even better than someone elses.

[chickenhead]

what do you mean by proof? do you use pars mainly for track to track comparison, or do you use them for daily variants as well?

[JackS]

I guess proof would come with results. If you can project a track to be two points fast and the results continously substantiate that two points is correct, I'd take that as proof. When would you begin to accept the two point difference? If you accepted it after one week, the next week could return to normal or below. Anyone using projections sucessfully should continue to use them. The problem as I see it, there are two many variables that can cause a track to appear "off".