Bayesian Analysis of NHL Betting Strategies

Background and Context

While browsing sportsbooks, I noticed that the point spread for NHL games was almost always 1.5 goals, with significantly more favorable odds than those offered on the moneyline. Since teams trailing by one goal often pull their goalie in the final minutes of a game, creating opportunities for their opponents to score into an empty net, I suspected that games would frequently end with a goal differential of two or more.

Having recently learned about Bayesian statistics, I thought this would be an ideal opportunity to apply what I had learned, particularly because I already had a prior belief about the outcome. I wanted to determine whether the odds offered on the puck line were genuinely as favorable as they appeared, or whether they were simply too good to be true.

Data Collection and Setup

To complete this project I needed data on both empty nets as well as betting odds. For the empty net data I used MoreHockeyStats and for odds data I had codex build me a webscraper which pulled from OddsPortal. One important note is that the odds posted on the website were the best available odds across several sportsbooks. If redone using only one sportsbook, it is possible that results may differ. Upon looking at the empty net data it became clear that teams did not have uniform results, meaning that I needed to use a hierarchal model (via JAGS). 

Methodology

In order to test how things actually worked out, I created 4 different strategies. The first would only bet moneyline, the second only bet on the spread (-1.5, not +1.5), the third bet on moneyline 50% of the time while betting on the spread the rest, and the fourth was the most complex. As mentioned previously, I used a hierarchal model to separate based on their empty net success percentage. In this fourth strategy, if the winning team was in the upper group it would bet the spread, and if they were in the lower group  the strategy trended conservative and bet the moneyline. This fourth strategy was titled "experiment".  

Within each strategy type, I wanted to account for bad, neutral, and good scenarios. In our bad scenario, it is assumed the correct winner is picked 40% of the time. In the neutral case we assume the correct winner is picked 60% percent of the time, and in the good case we assume 80% accuracy when picking the winner. For each case of each strategy I created 1000 simulated portfolios which placed a $10 bet on every single game from the start of the 23-24 season until the end of the 25-26 season. I tracked what the final balance of each portfolio was. 

Results

These results were very interesting and contradicted my earlier belief. The worst performing strategy was the one where games were strategically picked to take the spread. From there it was a 1-1 correspondence of the more frequently the spread was taken, the worse the portfolio performed. I found that the strategy in which exclusively the moneyline was taken in fact performed the best across all cases! I found this very interesting and it certainly changed the way I thought about NHL betting strategy.