1. #1
    HUY
    HUY's Avatar Become A Pro!
    Join Date: 04-29-09
    Posts: 253
    Betpoints: 3257

    Post descriptions of your losing models

    I understand that nobody wants to share their winning models because then they won't be winning any more. So I think there will be people who will have no problem to post their losing models, just to get some discussion going. A system qualifies as "losing" if it has been bet unsuccessfully or if its out-of-sample prediction is unsuccessful.

    I will make a start:

    Elo and Glicko systems on tennis. Tried to optimize the parameters using brute-force optimization and also Nelder-Mead optimization. Out-of-sample prediction very lacking.

  2. #2
    Miz
    Miz's Avatar Become A Pro!
    Join Date: 08-30-09
    Posts: 693
    Betpoints: 3132

    Anything that involves unsupervised neural network classifiers... for things like "win" in parimutuels. No matter what 56 graduate theses suggest.

  3. #3
    Miz
    Miz's Avatar Become A Pro!
    Join Date: 08-30-09
    Posts: 693
    Betpoints: 3132

    Ken Pomeroy's line projections. These don't work, but they are a gift that keeps giving, as lines tend to converge to them.

  4. #4
    HUY
    HUY's Avatar Become A Pro!
    Join Date: 04-29-09
    Posts: 253
    Betpoints: 3257

    Quote Originally Posted by Miz View Post
    Ken Pomeroy's line projections. These don't work, but they are a gift that keeps giving, as lines tend to converge to them.
    If the lines converge to them why don't they work?

  5. #5
    HUY
    HUY's Avatar Become A Pro!
    Join Date: 04-29-09
    Posts: 253
    Betpoints: 3257

    Quote Originally Posted by Miz View Post
    Anything that involves unsupervised neural network classifiers... for things like "win" in parimutuels. No matter what 56 graduate theses suggest.
    What input data did you give to the neural network?

  6. #6
    Miz
    Miz's Avatar Become A Pro!
    Join Date: 08-30-09
    Posts: 693
    Betpoints: 3132

    Quote Originally Posted by HUY View Post
    If the lines converge to them why don't they work?
    I should've said they don't work unilaterally. Sometimes they are incorrect and the lines seem to converge anyway... creating value... It's the Sagarin effect basically.
    Last edited by Miz; 03-02-13 at 11:38 AM.

  7. #7
    Miz
    Miz's Avatar Become A Pro!
    Join Date: 08-30-09
    Posts: 693
    Betpoints: 3132

    Quote Originally Posted by HUY View Post
    What input data did you give to the neural network?
    Tried various things. Supervised learning seems to work better. I also dislike the qualitative output, because it is harder to evaluate how accurate it is (harder to tell how big your adv is). There are better ways to include qualitative things like track conditions in a horse race, or "coming off a loss" in sports betting.

    It isn't that NN's don't work for prediction. Supervised is usually better than unsupervised.

    Many times for things like parimutuels where people like to try to use them, they are no better than the crowd (and usually worse than the crowd)... -EV
    Last edited by Miz; 03-02-13 at 06:08 AM.

Top