I see a lot of postings about the Kelly Criterion posted here at SBR. If you don’t know what the Kelly Criterion is, read
http://en.wikipedia.org/wiki/Kelly_criterion . And while the good Dr. Kelly “proved” that his Kelly Criterion was the optimal amount to wager in a Bell Telephone Tech Journal paper published in 1956, you as a sports bettor should NOT use it.
The chief reason is not that Kelly it is a bad formula, but rather that the Kelly bet fraction you calculate [basically, your losing percentage plus the casino advantage subtracted from your winning percentage, i.e. Kelly = Win% – (Lose% + vig) ] is for all intents incalculable by almost all sports bettors, probably even the Ganchrows of the world.
Why is this? Because you must calculate that win % as a Bayesian Prior. If you don’t know much about Bayes and his theorem, read
http://en.wikipedia.org/wiki/Bayes'_theorem.
Now, Kelly itself works pretty well in Blackjack, and came to fame in Ed Thorp’s Beat the Dealer, but the Bayesian Prior’s you calculate in BJ are 1) smaller as a whole than most sports win fractions thus less likely to affect your bankroll if you make a mistake 2) much easier because you have a well known (but not easy) statistical method to calculate Kelly Criterion in 21.
I contend that in sports situations you have four problems to deal with:
1) understanding Bayesian Priors in the first place – this is actually the easiest part, but if your head is spinning after reading some of Ganchrow’s posts you may have a problem.
2) calcing your win % given that your particular method may be suspect - I contend that angle players have a real problem here.
3) the particular sport league you are betting against evolves due to changing rules, changing players, and changing team management i.e. coaching decisions.
4) the line changes due to market perceptions of these league changes.
Now, I can write a lot about each of these four points, and I doubt that any of you are convinced of all four, but if you believe any one of the four is in place, you cannot accurately calc your Prior. And without an accurate Bayesian Prior, you cannot do Kelly…period.
OK, that’s enough to start the fight. Tell me why I’m wrong.