Anomalies in Repeat Winners

[traynor]

If someone has a large database available, it might be worthwhile to look at the percentages of horses that win (or win/place) in the next start and second start after a win based on the odds in the winning race.

I am seeing a VERY strong tendency for horses that went to post at (relatively) low odds in their winning races to come back with a strong performance in the next or second race after the win, and an equally strong tendency for horses that won at higher odds (3/1, 4/1, etc) to do poorly.

The probabilities are so symmetrical for both scenarios that it may be possible to toss a lot of favorites and chalky entries (and very few winners) based on little more than the odds in their (recent) winning race.

It might be worth looking at.

[Ray2000]

Seems right to me....

Using logistic regression to predict Win chance vs last race Odds and finish.
and considering no other factors.

Chart shows Win chance vs natural log Odds of last race Past performance

The odds were converted to log to keep in chart scale.... "0" = 1/1 Even money
The fitted lines are for different Finishes, the top curve is for wins as you pointed out.

Chart read Example
pplogodds of -0.5 ~ 3/5 on the tote
the predicted Win chance is 25% if comming off a win,.. 17% if finished 8th




Call:
glm(formula = CHARTWON ~ +PPLNODDS + PPFINISH_OFF, family = binomial,
data = z)

Deviance Residuals:
Min 1Q Median 3Q Max
-1.0798 -0.5723 -0.4775 -0.3917 2.5497

Coefficients:
Estimate Std. Error z value Pr(>|z|)
(Intercept) -1.024777 0.016809 -60.97 <2e-16 ***
PPLNODDS -0.288102 0.007343 -39.24 <2e-16 ***
PPFINISH_OFF -0.073504 0.004115 -17.86 <2e-16 ***
---
Signif. codes: 0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1

(Dispersion parameter for binomial family taken to be 1)

Null deviance: 103197 on 132291 degrees of freedom
Residual deviance: 99799 on 132289 degrees of freedom
AIC: 99805

[traynor]

Your results seem to indicate (much like my data and the data of a number of very serious bettors) that failure to account for the effect of confounding variables (in this case, the odds in the winning race) on factors considered predictive seriously diminishes the accuracy of those factors in predicting the outcome of future races. That diminished accuracy leads directly to decreased performance of decisions made on the basis of those factors.

Stated differently, the accuracy of factors considered predictive may be increased by calibrating the effect of confounding variables on those factors. Anyone intent on winning (rather than just "playing the game") would do well to exert the time and effort necessary to understand such concepts.

Thanks for the great graphics!

[traynor]

Note the tag on the lines below from the Saratoga Early Speed thread:

Stga Trot CloseToPace1CLastOddsAll 250 6.80 % ROI 1.16
Stga Trot CloseToPace2CLastOddsAll 250 8.80 % ROI 1.18
Stga Trot CloseToPace3CLastOddsAll 250 9.20 % ROI 0.98
Stga Trot CloseToPaceFinLastOddsAll 250 16.40 % ROI 0.90

What that means is that the data has been corrected for possible anomalies generated by different odds ranges. A very useful step if one is going to bet on the accuracy of one's information. It is also such a routine step--for me in calculating my own output and for others of my acquaintance in calculating their own output--that I sometimes forget it is not common knowledge or is not in common use.

[LottaKash]

On this subject, I believe that I have mostly figured out what classes repeat, and how often, and for how much....

That is what I would advise a "new" to handicapping-harness, to spend some time and effort with....Look to the classes for repeats....

[traynor]

On this subject, I believe that I have mostly figured out what classes repeat, and how often, and for how much....

That is what I would advise a "new" to handicapping-harness, to spend some time and effort with....Look to the classes for repeats....

I agree wholeheartedly that there are few areas more worthwhile to understand in harness racing than class, class levels, and performance related to those class levels.

[Ray2000]

LK traynor et al

The numbers for repeat Winners vs Trackmaster class ratings.
Includes multiple win streaks for 17888 starters who won last out.


TMClass WonLast Repeat Strike ROI
All 17888 3746 21% 0.79
50-60 189 24 13% 0.69
60-70 2171 411 19% 0.69
70-80 7979 1700 21% 0.77
80-90 6210 1309 21% 0.86 the juicy spot :)
90-100+ 1339 292 22% 0.76


Just one of the many ways to define class, but the trend is there.

[LottaKash]

LK traynor et al

The numbers for repeat Winners vs Trackmaster class ratings.
Includes multiple win streaks for 17888 starters who won last out.


TMClass WonLast Repeat Strike ROI
All 17888 3746 21% 0.79
50-60 189 24 13% 0.69
60-70 2171 411 19% 0.69
70-80 7979 1700 21% 0.77
80-90 6210 1309 21% 0.86 the juicy spot :)
90-100+ 1339 292 22% 0.76


Just one of the many ways to define class, but the trend is there.

Ray, thx for the numbers...

Still, despite these numbers there are subsets (not from a DB, I have none) to the "classes", that often have a very "high" repeatable rate....

In fact, there are a few that (paraphrasing from Mark Cramer) are just about "Automatics".....With certain qualifiers, that is....:cool: (secret tho)

[LottaKash]

I agree wholeheartedly that there are few areas more worthwhile to understand in harness racing than class, class levels, and performance related to those class levels.

Tray, coming from you, I am quite flattered....Being just a Pen & Paper guy n'all.............

[traynor]

Tray, coming from you, I am quite flattered....Being just a Pen & Paper guy n'all.............

Bluntly, if I did not know how to write apps to do what I want them to do, I would still be marking up a DRF with fine-point red Flair pens--my preferred approach for years. I don't like computers spitting out numbers and "ratings" that someone else developed, because I don't think whoever developed it knows any more than I do, and probably a lot less.

I got off to a running start in "profitable wagering" by realizing that there were about 10 "real class levels" between the $2500 and $4000 claiming levels at Turf Paradise (rather than just $2500, $3200, and $4000 listed in the PPs). Because of extended conditions, a $2500 claimer could face tougher opposition at that level than it would face at $3200--and the opposite was equally true. Horses that seemed (based on main conditions) to be "dropping in class" might actually be facing tougher competition than that in the "higher" class. And no one seemed to even consider such a thing. Horses "dropping" in class were overbet favorites. Horses "going up" in class were routinely discounted as "cheap speed that wouldn't be able to handle the competition at the higher class." It was a really good year.

The same situation still exists at many tracks, except it takes a bit more digging to uncover. Layering "class levels" into reality (rather than paper designations for the convenience of humans with limited attention spans) and then understanding which "moves" are potentials for improvement, decline, or similar performance puts a perspective on race analysis that all the number crunching with "speed" and "pace" analysis would be hard put to match. No computer required, either.