Contender Selection

[tanda]

I am developing my software for my own personal use (not commercial). I have previously posted some results of testing for comments.

I would like any comments from software users and developers on the following results.

I have revised the contender selection methodology. The test results are limited in size, so that is the one obvious explanation. Nonetheless, is anybody familiar with these types of results even in a limited size test?

All races are played in which all horses have at least one past performance line. That results in about 75% of races.

On Saturday, the software identified 236 contenders (69.4%) and 104 non-contenders (30.6%) in 46 races. The non-contenders won 2. The I.V.s were 1.38 v. 0.14. A flat bet on all contenders would have returned a 1.10 R.O.I.

The test included Gulfstream's card where the fav won 8 races, the second fav won 1 and a 14.40:1 shot won the other. The software played 8 of the 10 (7 where the favorite won and the 14.40 shot race). In other words, the highest odds on a winning contender that day (other than the 14.40 horse who was a contender) was 1.80:1.

There were 10 contenders who were between 9.40:1 and 17.70:1, but none won higher than 17.70:1. No huge prices, no 40:1 shots, etc.

The average contender won at 4.92:1. This seems sustainable as most tracks have average winning odds near or higher than that figure (although non-contender bombs help drive that number up).

Even for a 46 race sample, the results appear quite nice and surprising, especially in light of all the low-priced contenders who won at Gulfstream and the lack of any winner higher than 17.70:1.

Any thoughts?

[ranchwest]

Your ROI computation doesn't match your other figures.

32 winners at 4.92:1 yields $378.88. The bets would total $472 on 236 contenders. The division yields an ROI of .802.

46 races is not nearly enough data for meaningful conclusions. IMHO, 500 would be a minimum and the more the better.

[rrbauer]

Not sure where ranchwest's 32 winners came from. Looks like 44 winners to me for a return of $520.96 which yields the 1.10 ROI compared to $472 bet.

I agree that you need more data and once you get it you should slice and dice into categories to see where you might improve the ROI by excluding certain categories from your model. Also, I like to toss out the highest payoff from the mix when I deal with data of this type to mitigate the effect that a "bomb" might have on my averages.

[ranchwest]

Looks like I misread the message. I thought 69.4% was the winning percentage, but that was the percentage of total runners.

[tanda]

As I stated, 46 races is a paltry sample. Nonetheless, I was wondering how likely it is that any contender selection process could show a flat bet profit on all contenders (making up about 70%) of the field for a 46 race sample especially when none were more than 17:1.

As to throwing out the highest, I would agree if the highest winner was 60:1 or something. But the highest was 17.70 and there were several in the range 12 to 17. Thus, 17.70 does not seem an abnormally high number. When throwing out the top winner in studies, I first consider how typical the number is. If it is a number that should typically arise, then it should stay. If you bet 70% of the horses in a field, you should expect to catch a 20:1 plus shot every now and then, so 17:1 seems reasonable.

I will add results as I accumulate them. Maybe in 40-50 race samples.

[ranchwest]

I'm curious. Does your methodology guarantee at least one non-contender per race?

Since horses win approximately in proportion to their odds, if you are to continue hitting 95+%, then you're extremely likely to also hit horses over 20-1.

Your sample averages only 7.4 horses per race, which could partially account for the low number of big payouts. In 5 horse fields, the money gets spread around and there aren't a lot of 20-1 horses. Get some 12 horse fields in there and there may be more high odds horses coming in.

[GameTheory]

Those are good results, and probably indicate you are on the right track, but a small sample doesn't mean much. I've tested some methods that on one day will win race after race after race at 5-1 and up for huge ROI's, and then lose 30 straight the next day.

It is easy to get tricked because good results you naturally pay attention to. If you're getting mediocre results, you tend to keep tweaking the method until you hit have a good patch of results, but you can do the same thing betting randomly. You really need 100's of races to tell much of anything. Not only that, you need 100's of UNTESTED races, because if you keep adjusting things until you do well on a certain sample (even a big one), most times you'll have simply learned to predict THOSE races, not unseen ones.

[GR1@HTR]

Although it is a very small sample, I'd say you are off on the right foot. March forward and best of luck.

[tanda]

No, it does not guarantee a non-contender.

It evaluates horse on four form factors. The factors are liberal. A horse is a non-contender with two or more form defects.

Then, it uses three ability ratings. Two are related. If a horse is not within a certain amount of the top horse in the one rating or both of the related ratings, then it is a non-contender.

So, it is a combination of form and ability. The standards are liberal so that a failure is fairly significant. Most races have 2-3, some many, a few none.

[andicap]

Just wondering Tanda, what's the average number of contenders you have in a race?
(I know it changes with the size of the fields, etc.)

I believe a saw a similar method over at netcapper.com by a guy named "Valupix" who also bet a sizable number of horses per race and claimed a long-term 4.3% profit. Not a lot, but he was churning a lot of money through the machine.

[tanda]

So far, average field size is 7.42, Cs are 5.50 and NCs are 1.92.

So, it does not eliminate many horses ... but only 2 out of 198 have won in the new sample. So even if it only eliminates a few, if they win at such a low rate, it is still significant. Most are fairly obvious throwouts, but many are mid-priced horses and a few have been favorites.

Some other interesting results with a new sample of exactly 100 races.

Top horse when favored won 30 of 52 with 1.42 R.O.I.

Top horse when not favored won 4 of 48 with 0.30 R.O.I.

When top horse favored, R.O.I for remainder of field was 0.39.

When top horse not favored, R.O.I. for remainder of field was 1.02.

Contenders in all races, favored and not favored, 0.95.

Longshot contenders, all races, 1.15 R.O.I.

Longshot contenders, top horse not favored, 1.60 R.O.I.

[rrbauer]

When the top horse is favored: Bet it.
When the top horse is not favored and it's a "longshot": Bet it.
When the top horse is not favored; and, it's not a "longshot", bet the longshot(s).

[andicap]

TANDA,

What do you consider a longshot? 6-1? 8-1?

[andicap]

Originally posted by rrbauer
When the top horse is favored: Bet it.
When the top horse is not favored and it's a "longshot": Bet it.
When the top horse is not favored; and, it's not a "longshot", bet the longshot(s).

Richard,
I trust you are summing up Tanda's results. If so, I didn't see the 2nd one, i.e. betting the top horse when not favored and it's a longshot in his summary doing well.

[tanda]

Andicap,

For this thread, I do not know if I had a specific definition in mind.

I was posting some preliminary results for discussion.

One standard I do use is Field Size:1. So, in an eight horse field, 8:1 or more is a longshot. Or 10:1 of field size:1, whichever is smaller.

I agree with your comment on Richard's comments, although I have some additional results that are not quite as bad for non-favorites who are top ranked.

I am still trying to get a handle on the optimal betting strategy (assuming results hold up with additonal testing). But the idea of betting longshot contenders in races where the favorite is vulnerable (not top ranked) is promising.