Validity question.
During last year's basketball season, I thought of and developed a program for, of course, beating the pointspread.
It showed about a 57% win rate for about 1000 games.
This year, I threw out the previous year's sample, stopped tweaking, and tested it again (although against only one line, the final line as reported at Don Best; I would expect 1-2% improvement from line shopping).
Here are the results:
NBA sides: 121-88
NBA totals: 112-84
CBK sides: 260-211
CBK totals: 276-204
Combined: 769-587, 123.3 units won, 4.9 validity score.
By validity score, I mean (wins-losses)/sqrt(wins+losses). I don't know if this is the way to test for a 50/50 proposition. I know I didn't come up with a new formula here, but this is what I figured out as a good way, but that's just by instinct, because I have no math background. I just figure out how to do things as they come up. The program itself is written on Excel and I'm sure could be simplified because I end up doing a lot of brute force formulas only to later find simpler ways to write things.
Here are my questions, as I think about how I'll be betting this program next year (I didn't bet this year, because I was only interested in testing it).
In fact the first question is, should I bet it? Do I have the results validity to warrant investment?
Here are some things I like about it: I believe that the fact that the program works in NBA and college ball, sides and totals, is great evidence that it isn't a fluke. And it is the same program. It simply estimates a score. This means, I think, that it is fair to consider the total number of bets as the fair measure, and not have to break it up into the 4 subsets.
Here is where I'm unsure: the margins of advantage. IOW, how many points difference between my estimate and the Vegas line warrant a bet.
The above win/loss records come from 2 pt. NBA sides advantages, 2.5 pt. NBA totals advantages, 3 pt. CBK sides advantages, and 3.5 pt. CBK totals advantages. (I should say projected or estimated advantages, since the whole question here is whether I have an advantage at all).
I could cherry pick better win records.
NBA sides, 1.5 pt. advantage: 186-130
NBA totals 3 pt. advantage: 79-47
CBK sides 3.5 pt. advantage: 193-135
CBK totals 4 pt. advantage: 196-138
Combined: 654-450, +159 units, 6.1 validity score.
But I think cherry picking the betting margin is statistically dubious. My original margins were based on last years small sample development test, and was designed to make plays on about 1/4 of NBA games and 1/5 of CBK games (higher rate on NBA because the program becomes slightly more valid as the games played increases).
My compromise now, I think, re advantage margins, is to use some bet weights. I'm thinking of this: a top 10% margin (that is, a betting margin that would result in betting on 10% of the games) is set as a 1 unit bet, the 20% mark (25% in NBA) as a 1/2 unit bet, and the 5% mark is a 1.5 unit bet (none higher, though slightly more profit is available if I don't use this restriction).
Using this weighting method I end up with the weighted equivalent record of: 784.6-571.4, +154.5 units, 5.8 validity score.
So, next question: is it statistically fair that I use this weighting method? I do believe that a score estimator should win at a higher rate the further it gets from the Vegas line, and mine does (another reason I think it's valid), and therefore, some weighting is valid AND smoothes out the "cherry picking" problem.
Anyway, comments are much appreciated. As I said, I have no math background, and just gut my way through stuff, so I'd appreciate feedback.
I should add, because I think it's a validity factor, that the program was deduced, and not found through data-mining.
Anyway, enough babble. Any thoughts? Most particularly, would you bet this system next year if you were me?
Oh, and maybe I should add, no the system's not for sale, no I'm not a tout, no I don't want e-mails or PMs about it, and no I won't post picks next year because I don't want to be responsible in any way for anyone else's gambling.
|