Sample Size

[TGJB]

To address Michael\'s question on the other board about sample size (over 600 races) for the study:

If the study was measuring winners it would be a different story. Let\'s say both sides top ranked horse would win 30% of the time (around the rate for winning favorites). That would mean only around 200 races would be scored for each side, since if one side or the other didn\'t have the winner it wouldn\'t score. It also would mean that ordinary randomness-- and photo finishes-- could easily turn that 30% into 20 or 40% over what would be a fairly short period. And if you tie in betting (ROI) it gets even worse, because the results are leveraged-- if you go 15/25 on photos and the other guy goes 25/15, and the average payoff is $8, it can have a dramatic effect on the results. (If you happen to win the photos with the favorites and the other guy wins them with longshots-- inevitably how handicapping contests are decided-- that too can slant a study a lot). For a study based on betting you need a very big sampling for it to be meaningful.

The study we have been discussing, which is intended to measure figure accuracy and does not have a betting component, doesn\'t have any of those problems. a) You don\'t have to win the race, just beat the other guy, so every race scores, making the sample size effectively 3 times as big, and b) it\'s only one point per race, no matter what. A result-- or a photo- isn\'t leveraged because of the public\'s opinion of a horse. Even coming out +10 in photos will not have a huge effect on the outcome.

What we might get is something like a score of 350 to 300, which would mean someone will have been right 50 times more than the other. That would be meaningful. If it comes out 330 to 320, not so meaningful.

[BitPlayer]

TGJB -

Do TG and Ragozin really disagree that often?  I would think that your sample size of over 600 would be substantially reduced by races in which the horses who run 1-2 are ranked the same by both services, producing a tie and no point for either side.

When it\'s over, I would think the statisticians who read the board should be able to give us a p value indicating how probable it is that the winning margin is the result of relative merit versus random variation.

[TGJB]

Bit-- we\'ll find out, and I do hope the stats guys chime in later.

[miff]

JB,

Is the comparison starting tomorrow? Where will the comparison figs be posted.


Mike

p.s. Reviewing tomorrow\'s TG Spa sheets, I note the July 4th Belmont insanity where the track went insane changing speeds. It would be interesting to see how Rags handled July 4th given their averaging.

[NoCarolinaTony]

I was planning to give it an F- Bar Analysis myself.

NC Tony

[TGJB]

Yeah, I got your f-Bar right here...

It looks like the guys will be up and running, although everything is being done on the fly for the first day. Should be posts with the ratings and rankings up tomorrow morning.

Leaving right now.

[bobphilo]

Right Bit,

a 2 sample t test would tell if the difference between the 2 is statistacally significant and not do to normal variabilty.
Ties would not reduce the size of the sample, though if there were enough of them, it could lead to the conclusion that there is no significant difference in the 2 methods. That is also a possible outcome.

Bob

[gatodelsol]

If you want to increase the sample size, then you should also take each service\'s lowest rated horse and see which one finishes lower.  It would be just as meaningful (more meaningful to those that use the numbers more to eliminate horses from contention) if the only goal is to see which numbers are more accurate.