Weight Optimizer for Factors

[Sly7449]

Greetings,

Does anyone know where or if there is a tool or handicapping software that has the capability of optimizing the Weights appropriate to the Factors selected?

Does this feature even exist in any Handicapping Program or is there an add on program on the market?

Ideas welcomed.

Thanks

Sly

[MitchS]

Not sure if this is what your looking for. I use to mess with a program called Quick Horse. It has a "Supertune" feature and you can set-up your own criteria of factors to fine tune. To me though all you're really doing is backfitting the data. Couldn't really find a useful approach for the program myself...

[Tom Barrister]

If you do this, you're essentially backfitting the software to the results.

QuickHorse has what it calls "SuperTune" which adjusts the weights to produce an optimum ROI for past races. As with most backfitting, what worked in the test sample won't usually work elsewhere. I've seen some weird weights like:


FactorA 2
FactorB 1
FactorC 418
FactorD 3
FactorE 131


While it should be obvious that FactorC is probably not 418 times as powerfuol as FactorB, I've seen such weightings produce an ROI of 1.30 or more when backfitted to ,000 races . When the test is run on a control group (a different set of1,000 races) with the "SuperTuned" weights, the ROI is in the 0.80-0.85 neighborhood, where it belongs.

You can find Quickhorse at http://www.quickreckoning.com/quickhorse.htm

Bob Pitlak also might have something along the lines of what you're looking for on his website. It's http://www.sports-bet-advantage.com

[Wizard of Odds]

Rather, fit to maximize [Actual Wins - Expected Wins]:)

[Murph]

Thorostats (http://www.thorostats.com/) new interactive HOOF report does exactly what you are asking.

Set the weight for up to 45 different factors and have the weighted rankings alone AND applied to the POWER number for a better odds line.

Give it a try - backfitting can work well in several situations. I am finding the
HOOF ranking very helpful in evaluating favorites.

Murph

[podonne]

Backfitting can be largely avoided if you use a more robust sampling methodology when calculating the weights. I like to use two sets of 33%.

[Murph]

Backfitting can be largely avoided if you use a more robust sampling methodology when calculating the weights. I like to use two sets of 33%.Please elaborate, podonne. Define "more robust sampling" and explain two sets of 33%. I just play on what has worked in the past.

[PaceAdvantage]

Give it a try - backfitting can work well in several situations. I am finding the
HOOF ranking very helpful in evaluating favorites.Are you still associated with the Thorostats company? If so, you shouldn't be making these kinds of postings, especially without telling everyone of your affiliation.

[Murph]

Are you still associated with the Thorostats company? If so, you shouldn't be making these kinds of postings, especially without telling everyone of your affiliation.Seriously? You have an issue with that post?

If you do just delete it, Mike.

[PaceAdvantage]

Seriously, I didn't have an issue until this line:

"I am finding the HOOF ranking very helpful in evaluating favorites."

[Murph]

I don't understand why that line is an issue, it is simply my opinion.
It is in reply to a posted question. Thorostats (http://www.thorostats.com/) new interactive HOOF (http://docs.google.com/View?docid=dhpw4d2k_44fncmdsg2)
report fits what Sly7499 is asking about perfectly. You should have a look at it yourself.

I only responded in kind and on topic with my opinion. Was that the wrong thing to do?

[Sly7449]

Greetings,

I would like to express my Thanks to All that provided guidance on my question.

Did not mean to get anyone's toes stepped on but I may have provoked a line crossing in the process.

Gentlemen, your replies did encourage my thought process. I have tried doing the Weight assignment by the Binary Search Method which is an agonizing process. I made an attempt in trying the David B. Fogel formulae but that was not too successfull.

From some of your replies, I gather that there are mixed opinions on the validity of the Objective. One response indicated that it was great for identifying Favorites, Ouch!

Another suggested that there is no need for Weighting/tuning if a good foundation is laid. Hmmm. Interesting!

Looks like my initial query should have been, Is there any Value Added in trying to apply Adjusted Weights to factors and if so, how best could this be accomplished.

Thanks All

Sly

[socantra]

I don't understand why that line is an issue, it is simply my opinion.............

I only responded in kind and on topic with my opinion. Was that the wrong thing to do?

To tout someone onto a course of action which would profit you without revealing that fact to them would at the very least be considered ethically challenged. Your opinion is somewhat tainted by your affiliation with the product. The person you are advising should be made aware of that fact.

[Jeff P]

Looks like my initial query should have been, Is there any Value Added in trying to apply Adjusted Weights to factors and if so, how best could this be accomplished.Some factors are more important than others. So of course there's value in applying adjusted weights to factors.

Take the following simplistic model:Number = [( EarlySpeed x Wt1 ) + ( LateSpeed x Wt2 )] / ( Wt1 + Wt2 )If you think early and late speed are of equal importance then Wt1 and Wt2 should be equal in the above model.

Using models similar to the one shown for "Number" above, I've found that the importance of early speed vs the importance of late speed changes with race distance, race surface, and class of race.

For example, in a 5f race, in the above model, Wt1 would be higher than Wt2. Also, on the AQU inner dirt surface, in the above model, Wt1 would be higher than Wt2.

However, using the same model, in an 8.5f race, optimal results for "Number" might well be achieved when Wt2 is slightly higher than Wt1.

On most polytrack surfaces, using the above model, optimal results are achieved when Wt2 is made higher than optimal Wt2 for natural dirt surfaces.

Q. So how do you find optimal Wt1 and Wt2 for races of different distances run on different surfaces?

A. Lots of work.

Here's how I do it:

1. Create a factor similar to "Number" using arbitrary weights for Wt1 and Wt2. Perform number crunching and record the results for "Number" for all starters in a large database. Query that database, breaking the results out by distance and surface.

2. Create a second factor similar to "Number" using slightly different weights for Wt1 and Wt2. Let's call this factor "Number2." Perform number crunching and record the results for "Number2" for all starters in a large database. Query that database, breaking the results out by distance and surface.

3. Compare results for "Number" to results for "Number2" in specific distance/surface categories. One set of weights will usually provide stronger results than the other. Differences noted can indicate which set of weights is closer to correct.

4. Repeat, repeat, repeat...

The process can be a long one. But it CAN get you the correct set of weights for the factors in your model.


Greetings,

Does anyone know where or if there is a tool or handicapping software that has the capability of optimizing the Weights appropriate to the Factors selected?

Does this feature even exist in any Handicapping Program or is there an add on program on the market?

Ideas welcomed.

Thanks

Sly
JCapper has an interface called the IVTable Wizard that allows users to create their own factors. The process I recommend to my users is nearly identical to what I just described in this post. But be forewarned: Getting to the correct set of weights is WORK.

And yes, in case anyone is wondering, I'm the evil genius behind JCapper. <G>

-jp

.

[socantra]

And yes, in case anyone is wondering, I'm the evil genius behind JCapper. <G>

-jp

.

A disclosure made redundant by your identification as an Authorized Advertiser for jCapper and the jCapper link at the end of your post. :)

[Sly7449]

JP,

Thanks for your guidance.

I guess that I have been stubborn in my assigning equal weights to Factors that I have found to be Significant. (Short cut).

I have eventually stripped my Race segmentation as you described and what shows a 4.50 ROI on a Backtest resulted poorly when forward tested.

My sample size was small, 150 races.

What I discovered this past weekend was that the Larger I made my Sample Size, the Worst both Backtest and Forward Test became.

I have reached a Backtest sample size of 1755 races for a Track, Distance, Surface, Race Type (Clm NW x). Forward sample was 315.

End result on both test were discouraging, but then again, that was using Equal Weights.

I'll try using a Fibonacci and see if I get improvement.

Thanks for your insight.

Sly

[podonne]

Please elaborate, podonne. Define "more robust sampling" and explain two sets of 33%. I just play on what has worked in the past.

Hi Murph, sorry i didn't reply right away.

Sampling is a tool to use a large population of races to test a hypothesis. Most people will sample by using all the races in their particular database. A more robust method would be to divide the entire race database into several randomly selected sets and test your factors on each set independently. If the result you get is the same (or close) in both sets, you have a much better answer.

Because my database is rather large, I use two sets, each containing a randomly selected set of 33% of the entries.

Hope that helped :)

[podonne]

Here's how I do it:

1. Create a factor similar to "Number" using arbitrary weights for Wt1 and Wt2. Perform number crunching and record the results for "Number" for all starters in a large database. Query that database, breaking the results out by distance and surface.

2. Create a second factor similar to "Number" using slightly different weights for Wt1 and Wt2. Let's call this factor "Number2." Perform number crunching and record the results for "Number2" for all starters in a large database. Query that database, breaking the results out by distance and surface.

3. Compare results for "Number" to results for "Number2" in specific distance/surface categories. One set of weights will usually provide stronger results than the other. Differences noted can indicate which set of weights is closer to correct.

4. Repeat, repeat, repeat...

The process can be a long one. But it CAN get you the correct set of weights for the factors in your model.



This is similar to a genetic algorithm and you should get better results by applying a slightly different approach. Instead of doing it one at a time, create a large number of variations and test them all. Then select the two that scored the highest, and combine them together 100 times to create 100 "children", applying a small amount of "mutation" to each one. Test the children, pick the best two, etc...

[thoroughbred]

My companies software program, CompuTrak, takes weight changes into account.

If you go to our website: www.revelationprofits.com and go to our documentation to find "Engineering Analysis of Thoroughbred Racing", the equations for weight change are shown.

[Sly7449]

Greetings,


Once again I venture into this arena having stopped for a while. I did surrender due to the fact that what I thought was Practical Handicapping, started to appear to be Impractical.

Doing this manually for a few Factors and for a few races is a task by itself.

Very time consuming whereas if it could be automated, then one could set it and go to sleep and upon wakening, its done.

Hmmm.

Sly