Impact Factor versus true probability
 

What is the definitions of the impact factor calculation introduced by Quirin?

Are there any other alternative definitions that are better?

How do these IFs compare to "true" win probabilities of a horse racing factor or can they be adjusted to do so and how?

What i am getting at, say, for example of last time out winners (LTOW) is that in some races there will be several LTOWs but in some races none at all. That needs correcting for, to find a "true" probability of the LTOW factor.

Here is a link that may address what you are getting at. It is by Gordon Pine.

http://www.netcapper.com/TrackTracts...e/TT010223.htm

Well, there is no such thing as a "true" probability, but putting that aside the only way to adjust for the different scenarios precisely is to consider each scenario separately, i.e. create an IV/probability distribution independently for each different scenario you're interested in. So separate races where there is exactly one LTOW, compute the frequency, then do two, etc. Of course you need huge initial samples because your some of your subsets are going to be so small as to be meaningless otherwise...

Aside from the odds-related aspect that Gordon Pine mentioned, Mike Nunamaker also adjusted the impact values in his studies for the field sizes of the races in his sample. However, I don't know how you could take a set of impact values calculated according to Quirin's definition, and recalculate them after-the-fact to fit either Pine's or Nunamaker's method, without knowing odds data and field sizes for each of the individual horses or races in the sample (which are often not available, short of starting from scratch with your own calculations from original past performances). To me, Quirin's method suits the purpose well enough, as long as a sufficient sample size is used.

The true strength of the impact value for a particular factor or set of factors (and the effect on individual winning probability) can only be calculated or considered in the context of how all the other horses in the race stack up with regard to the same factor(s).

098poi,

Thanks for the Gordon Pine link.
A/E is definitely a useful index for seeing if a factor is over or underbet but it does not tell you anything about win probability for that factor. Using the price of a horse brings in all sorts of other handicapping factors which together sets its price. So presuming it is the single factor being investigated that does all that is very misleading.

Game Theory,
As you might suspect I am trying to avoid the incremental approach. I would guess it is a compound IV that needs to be found, as it is a combination of factors that may make the race winner, rather than just one single factor. The closer you look the more confused it gets.

Overlay,
Thanks for info.
What you are saying is that a large sample used for an IV calculation converges onto a "truish" probability. That knowledge, +/- 5%, is what I am after. Does the price really effect the chance of winning, or does the chance of winning effect the price?

Has anyone got any relevant information on Fred Davis's views, who possibly introduced the IV method in his 1974 book "Percentages and Probabilities"

I, too, have difficulty with the concept (or with my understanding of it) of the odds themselves affecting a horse's winning chances, although Gordon Pine used it as the basis for his A/E gauge. To me, the odds are a dependent variable driven by the public's interpretation of the whole body of information that is available about a given field. And the public, although remarkably accurate in the aggregate, can be wrong on a regular basis about the odds that it assigns to individual horses in individual races. Of course, horses at lower odds win at a greater rate than horses at higher odds, but (again, in my opinion) it's their superior performance attributes that both produce that winning result, and that also account for the fact that they go off at low odds.

Overlay, how do you handle factors that simply don't apply to all horses in a race? 2nd race off layoff, 1st off claim, trainer angles that only apply to one horse, etc...

The only factors I would use as stand-alone angles like that (that would not apply to all horses in a field) are those that would be sufficiently strong that they would statistically qualify as independent variables (such as the type Quirin discussed in Winning at the Races), that by themselves account for the winning percentages they produce, apart from any other variables or factors that might be present in the horse's record. Even then, though, I think there has to be consideration of the angle's winning percentage in light of the horse's odds, rather than just betting any particular angle blindly.

If your research indicates that horses that fit a particular condition run 3 points better than if they did not you can make that adjustment to thier expected speed rating.

Alternatively if you determine that a situation improves a horse's winning chances by 5% you can adjust the win probability and renormalize.

I do some of each.

I didn't say anything about betting them blindly, or even betting them at all. They could be negative angles. I was just wondering how you might adjust your oddsline for inherently "non-relative" factors where you can't compare horses within a race on that factor other than to say this horse has it and this horse doesn't...

===========================
From the article:
What if you knew that horses that won their last start went off at odds lower than average? In fact, they do. That would mean that the horses in this group might be winning a lot because of their low odds. Their winning the last race might not have anything to do with it.
==============================


He didn't really mean to say this ... did he? (The remainder of the article doesn't reveal the answer.)

(Sorry, I wasn't meaning to imply anything negative about anybody's particular betting style.) The factors (and there are only three of them) that I apply to individual horses (rather than to the full field) are strong positive ones (the type of independent variables that I previously described) that fall outside the scope of the factors that I'm using to rank the whole field on. When I've arrived at a composite value ratio for all the horses in the field using my "full-field" factors, if any horses in the field then qualify for the individual factors I employ, I multiply only the composite value of those horses by the value associated with the individual positive factor(s) for which they qualify, and use the new composite value for those horses as the basis for calculating their fair odds.

sjk,

To me that seems the crux of the problem.

If you take last time out winners they have, in UK, a "positive" IV but a "negative" A/E. The public can all see that the horse won last time out and its price falls (possibly too low) but its apparent chance of winning, compared to the field, has apparently risen, based purely on long term averaging study of winners last time out.

Any odds line adjustments could cancel out or make matters worse if the factor used was not the actual cause of a better chance. If the factor is not an independent one, it might be double counted or worse if you allow for it. So a winner last time out in a poor race, one that has been rested too long after win, one that got injured in winning that race, one racing at a different distance or ground condition is not so good as other last time out winners. You are then back to studying the actual horse in detail and possibly the many interacting IVs, A/Es which now seem not so helpful as at first sight they could be.

So, that is really what I am asking here, is there anything new that has made them more useful in practice?

I would think that winner last time out is far from independent of the speed figure data and would be a poor candidate for the type of adjustment mentioned in my post.

On the other hand a horse that was claimed by a 20% winning trainer away from a 5% winning trainer might well be a candidate for an adjustment and the fact that he was claimed is clearly independent of the PP information.

If a horse has the speed to make a clear lead on the other runners that is also information that is at least partly independent of his PPs viewed in a vacuum and might also inspire an adjustment.

I guess I am agreeing with you that a careful understanding the interrelations of the various factors (both from a conceptual and from an analytical perspective) is pretty important to getting it right.

Nice thread so far. :ThmbUp:

postet by robert99- In practice what I tend to do is this: Run a large sample for a given factor and take note of both the impact value and roi. After you've run enough samples for many different factors (I've done tens of thousands over the years) you'll gain a little knowledge about what to expect from impact values and roi. When you run a sample on a good factor it will stick out like a sore thumb.

Realistically it will have two attributes:

1. The factor is predictive. It will have a nice clean curve for impact values where horses that are strong in the factor being looked at relative to their fields will win more races than horses that are weak in that factor. If you run enough samples you'll notice that a LOT of factors meet this criteria. What they fail to meet is the next attribute which is:

2. The factor is ignored by the betting public. This is what I refer to as a hidden positive. The roi for horses that are strong in the factor being looked at will not only be much higher than horses that are weak in that factor - but - the roi for horses that are strong in the factor being looked at will be noticeably higher than the roi for top horses in other factors that you have looked at.

This last part is key. When a hidden positive is present over a significant sample, it's a very strong indicator that you are looking at a factor that is crying out to somehow be used in your overall mix.

I find that if you use hidden positives your assessments will be superior to assessments made by the vast majority of your competition. This should be a no brainer. But amazingly enough, it took me decades of trial and error to finally reach this conclusion.

If you fail to use hidden positives in your mix then you are just spinning your wheels and doing the same thing everybody else is doing. The result will be that you will face a very tough uphill climb in trying to gain a true edge over your competition.

That's what I have found to be true in practice.


-jp

.

My experience with Impact factors, or databases in general, is that they dilute the specific to the general. While a general trend is fine to learn the game, once you get to that plateau you have to have the specific to differentiate you from the crowd.

I tend to follow tracks that fall outside of these generalities just for the fact that, in studying the general, the specifics are lost. It has worked well for years as race track "foot prints" are not only specific, but they respond to their own unique variations.

NICE post JP and much food for thought!

I'm gonna enjoy chewing on this bit of Filet!

That just means your factors, or your database approach is too general. There is nothing inherently more or less general about using a database, it is just faster. There is no reason you can't program a database approach to be as specific as you like...

I always look for subsets - say you have a general stat about last out winners. Within, that is likely a subset of those who are profitable and another that is not. (There are).

Jeff P,

That sounds most encouraging and makes a lot of sense.
Thanks for posting.

Jeff, I agree with your assessment of the value of hidden positives, but how long does any such factor stay hidden before it starts getting noticed and overbet, requiring you to reinvent the wheel? It seems to me that using more fundamental factors provides greater durability over time, while still providing ample opportunities for finding value (especially employing a weighted blend of several such factors, and looking at the winning chances of all horses in a field, rather than concentrating only on finding a single "angle" horse, or the one most likely winner).

Warning. This post could get long. And it may ruffle some feathers because it goes against what a lot of others out there have advocated. Can't help that. The subject matter deserves a complete answer. Those of you willing to wade through it with an open mind I promise you can find some good info buried in here. What I'm posting about is stuff that is relevant to me whenever I think about the essence of winning. The process I go through isn't the only way to win. But the process I go through involves horses with hidden positives. And because the game is parimutuel in nature and because there is enough competition to make the pools pretty efficient - I believe at a core level betting horses with hidden positives is one of the requirements to being a winning player.

You'd be surprised. Factors stay hidden from the public for a reason. Mostly because either the factor itself is under the radar or the public refuses to believe that factor to have any real relevance. We just had a couple of lengthy threads arguing the validity of speed in workouts. In that thread I showed how to create a simple model using workout speed as the basis for profitable play. Speed in workouts isn't stand alone profitable so it was something I was willing to post about in order to illustrate how hidden positives can be used to good effect. I know of a couple of other factors (which I'm not willing to post about here) that have been stand alone flat bet profitable since I first came across them about 15 years ago.

The really interesting thing to me is that most people (and please understand I'm not picking on you) make assumptions about what I do. I get a sense from phone calls and emails a lot of people think that I use stand alone factors in a vacuum... that as soon as I discover a promising new factor that meets the criteria I posted that I'll start betting that factor as a stand alone. Nothing could be further from the truth. A lot of what I do is to take interesting and overlooked (but related) stand alone factors and use them as building blocks to create compound factors. The compound factors I'll produce will fall into one or more areas of handicapping endeavor. For example, to my way of thinking (and I need to stress that phrase so I'll say it again) to my way of thinking the five general areas of handicapping endeavor are:

1. Pace
2. Class
3. Ability from Speed Figures
4. Current Form or Condition
5. Human Connections

Every one of the compound factors that I produce is designed to represent in a unique way a horse's ability in one of the above five areas.

The interesting thing to me is that some of these compound factors outperform (in terms of win rate and roi) the sum of their parts. Then I'll turn around and use these compound factors and weight them in such a way as to create a decent power rating. So in essence the bulk of what I do is to do my own research and use what I discover to create my own compound factors and roll them together to make a power rating that is fundamentally better (IMHO) than whatever the vast majority of those I am betting against happens to be doing.

In essence I agree with you. I really do employ a weighted blend of several fundamental factors.

Continuing on with what you wrote:
From there I'll combine the power rating with the way I perceive people will bet the race to create a reasonably accurate odds line. The more work I do (the more I re-invent the wheel) the more I find I keep making better and better odds lines.

As an aside the most accurate odds line I currently make predicts race winners with more accuracy than the public does when they select the top four horses ranked by post time favoritism. But... in practice I find that accuracy (in terms of probability) in an odds line does not necessarily translate to being able to use that odds line as a pure play or pass decision making mechanism for profitable play.

I've invested a ton of my own man hours into the never ending quest to develop better and better odds lines. No matter how accurate an odds line I make in terms of predicting race winners... and as mentioned I am able to do that with a very high degree of accuracy.... one fact remains: Whenever I simulate flat win betting EVERY horse that goes to post at odds higher than that indicated as a theoretical break even point by the odds lines I create I lose significant money.

That tells me one of two things:

1. I suck at making odds lines.
2. The pools are pretty efficient by nature.

HOWEVER, using an odds line to provide a theoretical break even point for flat betting EVERY horse above that theoretical break even point is NOT how I use the odds lines that I make.

In practice I find that when I create spot plays keyed off of combinations of hidden positive angles and run those against the odds line theoretical break even point in play or pass decision making... THAT approach consistently wins money.

And the one thing I keep coming back to whenever I think about why that approach works for me is that is has me betting almost exclusively on ONLY THOSE HORSES with enough hidden positives in their records to make this game wonderfully worthwhile.


-jp

.

Couldn't finish editing my last post before the time limit expired... so I'll pick up where I left off...

Overlay, you posted:It looks like our approaches differ in the way we use odds lines.

Please understand that I am in no way saying that my way is the ONLY way. If someone else is able to do it by looking at the winning chances of all horses in a field, rather than concentrating only on finding a single "angle" horse, or the one most likely winner then even though their approach is different from what I do I'll be the first to step up and congratulate that person on a job well done.

Ultimately every player has to find an approach (whatever it might be) that works for them.


-jp

.

Even if they don't realize it, anyone making a profit is by definition betting on hidden positives. That's where overlays come from. (Just as anybody who makes a profit is therefore by definition betting on overlays, even if they don't believe in overlays.)


Personally, I love using single factors (or combos of esoteric factors) in a vacuum, although I use oddsline-based approaches as well. You talked about workouts. I posted a long while back about how I had made a flat-bet profit model using just info from workouts with no other factors. No one believed me because workouts aren't "accurate" and even if they were you've got to consider other factors. You've just GOT to. (Those who insist that "inaccurate" = "useless for making money" really don't understand the game they are playing, IMO.) Certain trainer and pedigree plays have huge ROIs and to look at other factors just seems to dilute them because they make the most when they don't seem to have anything else going for them. I find it actually isn't that hard to come up with flat-bet profitable angles that are based on weird things as long as you can live with low win percentages (10-15%).


I got to the point where on could make an oddsline better than the public (slightly) on any measure you'd care to use to determine that. But they didn't provide many betting opportunities, so I gave that up and now make them differently. They do provide a profit at the theorhetical break-even point and above, but only if I stick to the "contenders". I've never been able to make a full oddsline and have the low probability "overlays" make money. (Although when using spot plays I often bet outrageous long shots and win.) When betting from an oddsline, I restrict play to those where I assign a probability of above average for the field, which is always (1/field size). i.e. in a 5 horse race I need at least a horse assigned 20%, but in a 10 horse race 10% is ok.



I have found that making oddslines is one of the great mysteries of life. Jeff, you might want to try the following. If you've got a model, that model is made up of factors that you use as inputs, and then the model spits out a probability value, or something you convert to one. Now, some factors are presumably "more important" to the model than others -- it gives them more weight. And also presumably, you test your models on out-of-sample data and see how they do. Now, take one your "important" factors, and when running a test replace the true values for that factor with random values (in the correct range for that factor), or just scramble the values for that factor in each race (have horses swap values). Compare your randomized test with the true test. As long as you don't use the public odds as one of the random factors, and as long as your model isn't giving a hugely disproportionate weight to that factor (I'm presuming your model doesn't have any one factor that totally dominates), there is a great chance you'll find the randomized tests done in this way perform BETTER than using the actual factor values MORE than half the time. How important is your factor now? Welcome to data mining bias...



I would really like for you to elaborate on this -- "keyed off" of how? And the break-even point is determined how? Are you saying you are using your oddsline as-is, but only considering horses for betting that pass some additional factor filter?

GT, you wrote:It looks like your experiences with odds lines are almost exactly the same as my own. I've also found that when I stick to contenders (which by my own definition are horses with hidden positives) I can make money simply by applying a theoretical break even point provided by an odds line for play or pass decision making.

You also posted:Yes. That's exactly what I'm saying.

I'm going to use a really simple model to illustrate this. In this case, "keyed off of how" is a simple UDM or spot play defined by the following 2 rules:

1. The horse must be ranked 1st in its field for Power Rating

2. The horse must qualify according to the rules in Boxcar's "B Angle" as set forth in his post from a year or two ago.

OMG I can already hear the flood of groans out there in cyberland. :eek: :rolleyes:

But let's use THAT as an example of a working UDM and let it define whether or not a horse qualifies as a potential play.


Now, what I've done next is run that against all horses in my calendar year 2006 database broken out by OR3. Before we go any further I need to define OR3. In JCapper, OR3 stands for Odds Ratio Three or Post Time Odds divided by the program's JPRTote Odds Line. In a perfect world OR3 = 1.00 represents a theoretical break even point. Any horse with OR3 >= 1.00 represents an overlay and any horse with OR3 < 1.00 represents an underlay.

Next question: Wha does the data show?


Now, the UDMs I use in real life work a lot better than the overly simple model I've used here. But the way I use them - and my process in getting there - is almost identical to what I've shown here. Each UDM or model will have its own break even point for OR3 (and other odds ratios within the program.) In the end my play or pass decision making comes down to just 3 things:

1. Does the horse qualify according to the rules of an active UDM or model?

2. Does the horse look ok on the track (my subjective view.)

3. Does the horse have odds high enough to be above the model's break even point for a significant odds line ratio? Which BTW can even be part of the model itself.

If the answer is yes to all three, chances are I'm betting.

GT, in a nutshell... without going into any great detail about the actual factors in my models... that's what I do.


-jp

.

Jeff:

Thanks so much for a marvelously thorough and informative response.

Help me to understand impact values, something I had never heard of until joining this forum. Evidently I have been using IV without knowing it for years or maybe I just play an angle. Tell me an approximate IV for my process.

I use the Bris custom card feature and scan an average of 50 or so races per day. I look for a particular factor which takes about 10 seconds per race. Out of 50 races which would be an average of 400 or so horses I may find 2 horses with said factor. Many days there are none and some days there are 4 or 5 so I don't know the actual average but I'm guessing at 2 per day. At the end of a month there may be 40-60 plays and these average 25% or more winners at an average mutual of 10/1. So what would be the IV of my angle?:confused:

From "The Handicappers Condition Book" by James Quinn pg 18, "The probability of winning equals simply the percentage of winners having a characteristic divided by the percentage of starters having the characteristic."

So in your case (assuming each race you are talking about only has 1 horse with your factor) with 50 races with 8 horses in the field per race (my guess) that gives you 50 horses with your factor and 400 horses total. So 50 divided by 400 equals 12.5% of starters have that factor. If your horse wins 25% of the time then 25 divided by 12.5 equals 2 or an Impact Value of 2 which is 100% more than would be expected. ( A value of 1 would mean that the factor is not significant either way. If you had Aunt Millie randomly guess the winner by picking just names she liked she would probably have been correct between 12 and 13% of the time. A factor can produce a negative expectation also for example (from the book) 1st time starters over all lose more than would be expected from their percentage of starters.
You sound like you are on to something good but check the link to Gordon Pine's article, it goes more in depth and confuses me a bit!!!

Handicapping within angles
 

I use seven angles as screens to find live longshots. Then I handicap the strength of the field and other factors, workouts, trainer stats, distance, etc. Jockey, weight, and beyer speed figure are not involved in the process. After this it's go or no go.
For example, freshened fillies are one of my screens and I passed on a 75-1 winner at Calder yesterday because the other factors did not confirm a play. I did catch one at the same odds a few years ago because both the class of the race and workouts confirmed it. Despite missing an odd win here or there, the extra step has produced positive results.

I'd be interested in hearing more about the "freshened fillies" filter. I think that when a f or m goes sour, they just don't seem to get back without a break and i noticed this from the owners side first but then picked up on it from the handicapping side as well. do you see any difference with respect to age ?

john

Not quite right, I'm saying that out of the 50 races I only find an average of 2 plays.

Frsh fillies/cold colts
 

The only age consideration involves very young two year olds. Horses raced in April or May of their 2yo year may have been breezed before their knees closed. So if one wins at first asking and then is given time off I pass.
The others are most productive in dirt races up to seven furlongs. I like at least four weeks off, some display of talent earlier in their career, and decent workouts depending on the individual trainer. Blinkers on or off is a positive.
Restricted races are the most productive but it works all the way up to stakes races. The longshot cited had three months off following a couple of races where she was assigned zero beyers. The race she won following the layoff was nw2 claiming $10k at Fairgrounds. Her maiden win was in a $20k race in New York so she passed the talent question. She was working well in the morning (note to discount if the works are limited to three furlongs) and a low profile jockey broke her on top and she wired the field.
One problem that arises is that there can be more than one in a race that has been freshened.
For confirmation of this angle review the pp's for fillies in sprint races in the form and look for that line that shows time off. You'll see some nice prices or otherwise solid efforts.
Colts usually need a race. The exception are the male offspring of Meadowlake and Phone Trick who will run fresh. Might be a conformational thing as these are frequently the boxy typed who pound away at their knees.

Well then,if 098poi has the formula correct,you have an Impact Value of 5000.
TTRAC

Impact value is the pecentage of winners with the factor -- the percentage of ALL winners in a sample with the factor divided by the percentage of all the starters with the factor.

So, let's say you have a sample of 1000 races and you find 50 plays. Let's say you have 13 winners out of the 50. And let's say the total number of starters in the 1000 races is 8000 horses.

IV = ( %winners with factor ) / ( % starters with factor )
= (13/1000) / (50/8000)
= 0.013 / 0.00625
= 2.08

Odds don't come into it.

So, out of 400 horses per day, 2 have the factor you're referring to, which would be 1/2 of 1% (0.5%) of all the horses. 25% of the horses with the factor win, which means one winner every two days (on average) out of the 100 races screened over those two days. So, 1% of the winners are coming from 0.5% of the horses, which would be an impact value of 1.0/.5, or 2.00, meaning that horses with the factor are winning two times their "fair share" of the races.

Your initial post said at the end of the month you have 40-60 plays. That is where I got the 50 from. You only look at your plays. What percentage would win as their "fair share" relative to those that actually won. Forget about how many races aren't plays. It doesnt matter if only 1 in 500 races has a play, after 50 plays you should get the results I said.

I see said the blind man. Thanks:confused:

GameTheory,

IMHO, a better formula is:

Wins / ExpectedWins

where:
Wins=actual wins

ExpectedWins= the sum of 1/field for each horse in the sample.


Dave

Going for a record here, bumping a thread over ten years since last post. ;)

This was an interesting read. I wonder if people’s thinking has changed at all in the intervening time period.

A guy called Roger Biggs (not the train robber!) here in Australia was/is big into the IV’s, wrote a few self published books about them. Personally, I think they have limitations. I think for your standard form factors like last start finish pos etc the market already has it all accounted for.

As you say, the market may very well have those factors accounted for from the standpoint of not overlooking them or omitting them, but (to me) the usefulness of the IV's is that they allow you to determine not just if the factors have been taken into account, but also whether, in the process of being taken into account, they have been undervalued or overvalued (especially in combination) by the public.