1. #1
    Carl-Haakon
    Carl-Haakon's Avatar Become A Pro!
    Join Date: 02-08-13
    Posts: 35

    Staking problems

    I've been trying to model ice hockey totals for a few european leagues and my model seemed to be doing fairly well before I got hold of any odds.

    I evaluated it by checking, for each game, how great of a probability the model had assigned to the correct outcome of the game (which in this case was over or under 5.5 goals scored). The average assigned probability for the correct outcomes was 53.3%, which seemed OK. I then scraped the market average odds for O/U 5.5 for these games and checked to see what the average bookmaker's average probability for a correct outcome was. This turned out to be a 51.7%; I didn't adjust their implied probabilities for overround either, so this should have been an indication that my model would be profitable.

    Here's the weird part though: it's not. I've tried simulated betting for out-of-sample seasons with every staking strategy I know of (kelly, percentage, flat), and most of those simulations have ended up bankrupt. I thought that it might just be that the model itself is unprofitable and that my four test seasons where just lucky, but then I simulated betting over those seasons and ended up bankrupt there aswell! I've tried setting a higher threshold for the minimum edge the simulator should be willing to bet, but that doesn't work either.

    Does anyone have any idea what I might be doing wrong? The model itself is calibrated: e.g. fourteen percent of all predictions that were deemed to be fourteen percent-likely came true. Is this enough to accurately gauge ones edge?

    I'm going to hunt for bugs tonight again, but I'm pretty exasperated at this point. If anyone has any ideas, please share

  2. #2
    matthew919
    Update your status
    matthew919's Avatar Become A Pro!
    Join Date: 11-21-12
    Posts: 421
    Betpoints: 5869

    Quote Originally Posted by Carl-Haakon View Post
    I've been trying to model ice hockey totals for a few european leagues and my model seemed to be doing fairly well before I got hold of any odds.

    I evaluated it by checking, for each game, how great of a probability the model had assigned to the correct outcome of the game (which in this case was over or under 5.5 goals scored). The average assigned probability for the correct outcomes was 53.3%, which seemed OK. I then scraped the market average odds for O/U 5.5 for these games and checked to see what the average bookmaker's average probability for a correct outcome was. This turned out to be a 51.7%; I didn't adjust their implied probabilities for overround either, so this should have been an indication that my model would be profitable.

    Here's the weird part though: it's not. I've tried simulated betting for out-of-sample seasons with every staking strategy I know of (kelly, percentage, flat), and most of those simulations have ended up bankrupt. I thought that it might just be that the model itself is unprofitable and that my four test seasons where just lucky, but then I simulated betting over those seasons and ended up bankrupt there aswell! I've tried setting a higher threshold for the minimum edge the simulator should be willing to bet, but that doesn't work either.

    Does anyone have any idea what I might be doing wrong? The model itself is calibrated: e.g. fourteen percent of all predictions that were deemed to be fourteen percent-likely came true. Is this enough to accurately gauge ones edge?

    I'm going to hunt for bugs tonight again, but I'm pretty exasperated at this point. If anyone has any ideas, please share
    I think I'm missing something here- you're using the same O/U of 5.5 for every game, and modeling the binary outcome with a logistic regression (or something similar)? Why not use something like an ordered probit model and model the exact number of goals scored, for all integers? PM me if you want to talk more.

  3. #3
    Carl-Haakon
    Carl-Haakon's Avatar Become A Pro!
    Join Date: 02-08-13
    Posts: 35

    Exactly; the 5.5 OU is the one moste widely available, so I take the average market odds for those outcomes and compare it to my probabilistic predictions for those outcomes.

    Actually, my model gives exact probabilities for each score and they all sum up to 1 (by construction).

  4. #4
    Carl-Haakon
    Carl-Haakon's Avatar Become A Pro!
    Join Date: 02-08-13
    Posts: 35

    I'll PM once I've enough posts (one more after this)

  5. #5
    statnerds
    Put me in coach
    statnerds's Avatar Become A Pro!
    Join Date: 09-23-09
    Posts: 4,047
    Betpoints: 103

    not profitable over what sample size?

Top