Why do we still use Impact Values?

[podonne]

Originally Posted by Overlay

Quirin presented the chi-squared test as a means of determining whether results (which he expressed as impact values) that seemed to be positive or negative were a fluctuation without any particular statistical significance that might not continue into the future, or an independent trend or factor that was not interacting with other variables to produce the outcomes that it was showing, and that could more reliably be counted on to maintain its positive or negative character in subsequent samples. (Of course, Quirin used comparatively small sample sizes.)

Quirin is absolutley right to do so. But why even list the impact value, then? Why not just chi-squared and ROI? or chisq and Win rate?

[podonne]

Originally Posted by TrifectaMike

Podonne,

You might be interested in this thread

http://www.paceadvantage.com/forum/s...=square+method

It came to an premature close, but you might find some of the answers you are looking for at the end of the thread.

Mike (Dr Beav)

Hi Mike,

That is quite a thread! I hope at some point you will continue your instructions. I might suggest writing an article, though, to cut down on the mid-lesson distractions.

But, since it seems like the same people commenting there are reading \ commenting here, I will not belabor the point.

Instead I will just say this to anyone who might read this thread in the future. There are several strong, well-researched, statistically valid ways of calculating the relevance of a binary factor from a 2x2 contingency table. ChiSquared is one, also Fischer's exact test and the binomial test. Impact Values are not.

I see them as a crude and unnecessary attempt to combine relevance (i.e. statistical significance) and magnitude in the same statistic. Relevance is easily measured per the above, and magnitude is as easy as ROI or win rate depending on what you're looking for. An accurate notion of relevance and magnitude is incredibly important, and an Impact Value alone provides neither.

Reading the comments so far, I think people use them because they're easy to read, and because they're different. Its kind of like jargon, or the thought that you've discovered something special, something unique to horse racing. And perhaps there is research proving that an Impact Value provides a more accurate picture of a factor's value than a ChSq test and a win rate\roi.

If there is, lets get it and not keep it as a horse racing secret! Social scientists, stat-heads, math wizards have been doing it wrong this whole time. But, there's a reason statistics books don't mention them, and you didn't learn about them in school, and my boss (Me: MBA, Boss: PhD in Stats) was confused when I described how they're calculated.

If we're going to do something different, we need to prove that its better first.

Thanks to everyone for listening. This forum continues to provide an incredibly enriching experience.

[traynor]

A nice genetleman attempted to use his handicapping methodology to start a business a few years ago--including a handicapping community forum and such--and flopped miserably because he failed to distinguish between impact values for entries and for races.

If impact values are applied to races, (unless something really weird is going on) it is possible to have a perfect correlation with winners by restricting wagers to horses with four legs. That is, every race will be won by a horse with four legs. Unfortunately (and a distinction the nice gentleman failed to grasp) is that more than one entry in each race has that attribute. His statistical analysis of his application ignored that fact (according to the nice gentleman, at the insistence of an Eastern European programmer who stated that his (the programmer's) algorithms used the best statistical procedures available) and declared huge winning percentages for each of his key factors. Ignoring the fact that there were multiple entries in virtually every race with those key factors.

Impact values are simple, but provide valuable information. They answer the question, "How many entries with this specific attribute actually win their races?" It ain't rocket science, and is not intended to be. It is intended to make a simple calculation of the existence or non-existence of a particular factor in entries that win their races. Precisely as Quirin (the PhD and tenured full professor of mathematics and computer science at Adelphi University) explained in his book.

Impact values are dead simple to use. Establish how many horses have a specific attribute. Establish how many horses with that specific attribute won their races regardless of the number of entries with that attribute that may have been entered in an individual race. Divide the latter by the former, and you have an "impact value." Pretty simple stuff. Nothing complex intended or implied.

It may be that the cause of confusion is how those impact values are used. The only thing the impact value expresses is the ratio of winners with the attribute in question in comparison to the total number of entries with that attribute. Nothing more.

[traynor]

A chi square is nice for expressing magnitude. Impact values are only concerned with the existence or non-existence of a particular attribute. The magnitude of the attribute is irrelevant (in that calculation).

[Maximillion]

Originally Posted by traynor

A nice genetleman attempted to use his handicapping methodology to start a business a few years ago--including a handicapping community forum and such--and flopped miserably because he failed to distinguish between impact values for entries and for races.

If impact values are applied to races, (unless something really weird is going on) it is possible to have a perfect correlation with winners by restricting wagers to horses with four legs. That is, every race will be won by a horse with four legs. Unfortunately (and a distinction the nice gentleman failed to grasp) is that more than one entry in each race has that attribute. His statistical analysis of his application ignored that fact (according to the nice gentleman, at the insistence of an Eastern European programmer who stated that his (the programmer's) algorithms used the best statistical procedures available) and declared huge winning percentages for each of his key factors. Ignoring the fact that there were multiple entries in virtually every race with those key factors.

Impact values are simple, but provide valuable information. They answer the question, "How many entries with this specific attribute actually win their races?" It ain't rocket science, and is not intended to be. It is intended to make a simple calculation of the existence or non-existence of a particular factor in entries that win their races. Precisely as Quirin (the PhD and tenured full professor of mathematics and computer science at Adelphi University) explained in his book.

Impact values are dead simple to use. Establish how many horses have a specific attribute. Establish how many horses with that specific attribute won their races regardless of the number of entries with that attribute that may have been entered in an individual race. Divide the latter by the former, and you have an "impact value." Pretty simple stuff. Nothing complex intended or implied.

It may be that the cause of confusion is how those impact values are used. The only thing the impact value expresses is the ratio of winners with the attribute in question in comparison to the total number of entries with that attribute. Nothing more.

Nice post....And imo taking the above a step further into a "consensus"
type of selection process just isnt for me...I dont want "consensus" in my selections,I would rather strike out swinging for the homerun...just mho

[InControlX]

The challange of profitable computer-based handicapping seems to be that the horses have properties, like herd instincts for one, which do not fit well in classical statistical evaluations. The horses' performances are not simply catagorized such as the statistical survival of bacteria vs. experimental drugs. Nor is there much in performance gaps to work with since a typical winning margin represents only a fractional percent in energy terms, and often negative when route paths are observed.

I'm not implying the computer-based predictions are not worth the try. At a bare minimum the programs can scan hundreds of entries for one's favorable patterns almost instantly and relieve hours of equivalent manual drudgery without processing errors. However, a successful technique must go further than classical regressions, introducing new methods unique to thoroughbred evaluations including probability estimates of chaos factors.

Those who find and implement these techniques are more likely to have some horse dung on their shoes as well as keyboards at their fingertips.

ICX

[TrifectaMike]

Originally Posted by InControlX

The challange of profitable computer-based handicapping seems to be that the horses have properties, like herd instincts for one, which do not fit well in classical statistical evaluations. The horses' performances are not simply catagorized such as the statistical survival of bacteria vs. experimental drugs. Nor is there much in performance gaps to work with since a typical winning margin represents only a fractional percent in energy terms, and often negative when route paths are observed.

I'm not implying the computer-based predictions are not worth the try. At a bare minimum the programs can scan hundreds of entries for one's favorable patterns almost instantly and relieve hours of equivalent manual drudgery without processing errors. However, a successful technique must go further than classical regressions, introducing new methods unique to thoroughbred evaluations including probability estimates of chaos factors. I like this statement.

Those who find and implement these techniques are more likely to have some horse dung on their shoes as well as keyboards at their fingertips.

ICX

Mike (Dr Beav)

[pondman]

Originally Posted by InControlX

Those who find and implement these techniques are more likely to have some horse dung on their shoes as well as keyboards at their fingertips.

ICX

By the time you've come up with a solution the meets over or it's old news. What if there are not enough occurences? For example, I've got my eye on Sacramento. I'm waiting for horses which have just broken their maidens. I'm looking for repeat secondary jockeys (I stay away from super jocks.) If they show up I've got $300 on them. The bets would go against convention and the speed head, because they are going to be slow in PP. It's the kind of bet I walk away with + 10k. And if nothing shows up, I lose nothing. My point: how long would it take you to do this with math?

[traynor]

Originally Posted by pondman

By the time you've come up with a solution the meets over or it's old news. What if there are not enough occurences? For example, I've got my eye on Sacramento. I'm waiting for horses which have just broken their maidens. I'm looking for repeat secondary jockeys (I stay away from super jocks.) If they show up I've got $300 on them. The bets would go against convention and the speed head, because they are going to be slow in PP. It's the kind of bet I walk away with + 10k. And if nothing shows up, I lose nothing. My point: how long would it take you to do this with math?

That is the purpose of data mining and reverse engineering. Data mining looks for correlations (usually in large masse of data) that might escape notice in visually scanning records. Reverse engineering works from winners, going backwards to find indicators in the PPs that would have indicated that winner. Neither is perfect, and both need a LOT of common sense to make them work as they should.