Optimizing Z-Scores.

[Grind-It-Out]

I have a database of handicappers' results.

As was recommended in Ganchrow's SWEET post,http://forum.sbrforum.com/players-ta...r-success.html, I am using Z-scores to determine which handicapper to follow.

My questions:

Should I give more weight to bets that have occurred recently? I was thinking about using the capper's past 2,000 picks to form the z-score. The most recent pick would have a multiplier of 1, and each pick thereafter would have a multiplier of .0001 less than the previous.

Should I separate Z-Scores by sport? If a capper has a z-score of 1.5 overall but 2.8 on baseball, which value should I use when betting on baseball? Perhaps halfway between the two?

Should I Discount Very Small Sample Sizes? One of the guys I'm looking at is 12-5 with a z-score of 1.33. Pretending that 1.33 was enough confidence to bet, my gut tells me that 17 games isn't enough of a sample size. Should I ignore my gut, or should I ignore this capper?

How Should I Discount Related Picks? If a capper bets on both the spread (or runline) and moneyline of a game, should I just ignore one of the two? What if one wins and the other loses, or one wins and the other pushes?

Thanks for any help!

[Dark Horse]

You should probably go by sport, allowing for fluctuations from one season to the next (don't expect a 60% ATS NFL season to be followed by another 60% ATS NFL season; regardless of Z-score for the first season). There is no reason to give greater weight to the last bet than the previous.

If you trace the long term records of excellent players, over several years in different sports, you will see a long term ceiling in the 57-58% ATS range. Anything higher than that, in the short term, and in my opinion, should be considered a temporary fluctuation. Of course, the larger the sample size, the higher the Z-score with that type of success rate.

[djiddish98]

I'll dust off some stat knowledge and put in some two cents.

First, I wouldn't weight bets, unless you know of someone becoming better at capping a specific sport, or they are fine tuning their model. Unless they state otherwise, presume that they are using the same criteria for each bet they place, so there's no weighting. What might seem like a "hot streak" is most likely luck, so don't over-factor it.

I would probably do Z-score by sport, simply because I can imagine certain handicappers have better knowledge / models for a specific sport. Handicapper A's results in Sport X are probably unrelated to the results in Sport Y.

For 3, it's been a whiles since I dabbled in stat, but wouldn't a T-test here be better with these smaller sample sizes?

My answer to your last question depends on personal preference. If you'd like to have a set of unique picks with little / no correlation, then you'd exclude. If you're looking for gauging a handicapper's success, then include it all.

I would think it would be easier to just go through and include everything, since presumably you're calculating the units won, not just the records. It seems easier than manually filtering through your data to exclude items.

[Grind-It-Out]

First, I wouldn't weight bets, unless you know of someone becoming better at capping a specific sport, or they are fine tuning their model. Unless they state otherwise, presume that they are using the same criteria for each bet they place, so there's no weighting. What might seem like a "hot streak" is most likely luck, so don't over-factor it.

Great point. My thinking was that I should weight the bets to account for changes in handicapping models, but that would hurt me in every case where a capper didn't change his model.

I would probably do Z-score by sport, simply because I can imagine certain handicappers have better knowledge / models for a specific sport. Handicapper A's results in Sport X are probably unrelated to the results in Sport Y.

Since these aren't actual bets, just picks at a free monitoring service, I can certainly see where an expert in one sport may try his luck at another sport that he wouldn't actually bet in real life. So, I agree with you again.

For 3, it's been a whiles since I dabbled in stat, but wouldn't a T-test here be better with these smaller sample sizes?

I never took statistics, unfortunately. I just try to test out what I read about. I'll look it up.

My answer to your last question depends on personal preference. If you'd like to have a set of unique picks with little / no correlation, then you'd exclude. If you're looking for gauging a handicapper's success, then include it all.

I would think it would be easier to just go through and include everything, since presumably you're calculating the units won, not just the records. It seems easier than manually filtering through your data to exclude items.

If I include it all, I can get all the info I need from one (pretty frickin awesome) SQL query. Running the query and processing the info takes about 30 seconds. If I separate, then we're talking about some loops of queries. In that case, I would guess that the whole process would take 15 minutes or so. So, including everything is much easier on my mind and my computer!

Thanks!

[Flying Dutchman]

Good luck. Handicappers are such a variable lot that I never got anything working off any kind of metric, z-scores included.