My Hockey Betting Algorithm Construction Paradox

My attempts to build an algorithm that determines NHL betting line value has encountered a paradox and I'd like to enlist the help Hockey Twitter to help my brain make sense of this. The premise of my algorithm project was to estimate the probability that a team wins, then subtract the implied probability of the betting moneyline and chart it against payouts to see if I'm getting value. If an event has a 75% chance of occurring, but the betting line is paying you like there's a 50% probability, then you are theoretically getting substantial positive value on the bet. You're getting a higher payout for something that's more likely than the betting odds imply. On a $100 bet, a 50% implied probability will double your money with $100 profit if it hits. A successful bet on 75% implied probability will pay only $33 profit. Increase the risk, increase the reward. Ergo; look for likely events that are paying you as if they are less likely.

 

I'm trying to find value under the theory that betting on it will lead to more money won in my gambling experiment. There are two charts below, one using 2021 data, the other using 2019/20 data, which is a much larger sample with a greater distribution of opponents. In these charts, “line value” is estimated by taking a team’s winning percentage, add it to their opponent’s losing percentage, and divide it by 2. Then you chart that line value on the x-axis with “betting payout” on the y-axis. This a crudely simple method, but produces a positive result using 2019/20 data. The larger sample produced exactly the result I was expecting, with the trendline passing through Y and X equals zero. If I made a $100 bet on every NHL game where I’m getting "value" in 2019/20, there is a significant profit.

 

The problem I've encountered in applying the same model to the smaller 2021 sample is that the inverse has been occurring. Betting on line value has produced negative returns, and betting against the recommendation would produce a net profit. The same phenomenon occurred when using goal differential to estimate win probability, positive results in 2019/20 and negative results in 2021. I’m at a loss to explain why this occurring. Obviously, the schedule is radically different in the pandemic world, with a much less random distribution of opponents. The same teams are playing each other repeatedly, which should theoretically make it easier to pick winners. Then again, it’s also more certainty for oddsmakers, who are possibly manipulating these unique scheduling circumstances in their favor with the lines being offered. The more value there is on the line, the less money you’re expected to win. It’s a contradiction.

 

One possible explanation is that betting underdogs was more profitable last season, with increasing likelihood as the schedule progressed. In 2021, favorites are paying out at a much higher rate of return. That doesn’t necessarily mean that upsets are less likely in the pandemic, but possibly that the bets on the underdogs are paying out less money than they should. The explanation may be a little bit of both. It’s also entirely possible that sample size is mostly to blame, and more games played will normalize results. As more games are played, the gap between good and bad could shrink. But through half of a season, the negative value is still paying off. I’ve been asking myself; should I start betting against the algorithm’s recommendation, or continue on this path assuming the data will normalize?

We may have reached a point in the season where the gap between good and bad teams shrinks, while the amount of profit being offered on favorites by oddsmakers stays the same and doesn't shift to reflect the shrinking skill gap. Upsets become more likely, a large number of bettors don't notice, and the sportsbooks bank the difference. That's what happened last season, but this current 56-game grind is its own animal. 

 

What say you?