Don’t Tell Me The Odds is a hockey handicapping site that uses sophisticated algorithms to predict money lines, spreads and totals for the NHL and virtually every other major league:
On the site you will see multiple NHL models. The top performing model is called RUS. Three other models (1,2,3) are provided that are optimally used as inputs in your own models.
European predictions will be posted overnight (eastern time). NHL predictions will be posted by around noon (eastern time).
If you have any questions or requests for predictions not currently being offered please contact [email protected]
A 1-month subscription costs $1099 (CAD) for about 900 games of predictions.
Our statistics tables are free to use.
RUS is a proprietary model that provides the score and win probability for every NHL game. This model is influenced by roster choices so it will be periodically updated throughout the day as team’s finalize their lineup decisions.
The model was backtested using the same methodology here.
The evaluation exercise I will undertake here involves taking 7 seasons of past results (2009-2016) and repeatedly go through the following steps:
This process allows assessment of the model behavior with data it has not seen during its building phase and by having it run through all the games in a given season we get a feel for potential losing and winning streaks we might have to face.
As is common in the field, the rationale behind my staking plan is the Kelly formula, which accounts for both the difference between bookies and modeled odds (the advantage we think we have on the bookies) and the probability of winning. The outcome from the Kelly formula is an estimate (f) of the fraction of our bank we should stake on a given bet to guarantee an optimum return on the long term. This in essence is what drives the stake when following the Kelly strategy: your bet size increases with both your chance of winning and the advantage you think you have on the bookies.
An example of the Kelly Criterion:
The Kelly Criterion formula is: (bp-q)/b
B = the Decimal odds -1
P = the probability of success
Q = the probability of failure (i.e. 1-p)
Using a coin as an example of Kelly Criterion staking, consider you are betting on a coin to land on heads at 2.00. However, the coin is biased and has a 52% chance of ending up on heads. In this case:
Q = 1-0.52 = 0.48
B = 2-1 = 1.
This works out at: (0.52x1 – 0.48) / 1 = 0.04. Therefore the Kelly Criterion would recommend you bet 4% of your bankroll.
The Kelly Criterion works very well when you know the exact probability of your bet (a coin landing on heads) but runs into issues when less precise odds (a hockey team winning a game). Overconfidence in one’s ability to know their precise odds can lead to over betting (betting too much) which will lead to long term loss. That is why we recommend people use a fractional Kelly Criterion, in our examples we use Half Kelly Criterion meaning we take our recommend bet size and divide it in half. Building off our coin flipping example, if we were using a Half Kelly Criterion we would bet 2% of our bankroll.
Kelly Criterion also works based on compounding our bankroll. Example:
Day 1 Bankroll: $1000
Day 1 Winnings: $20
Day 2 Bankroll: $1020
An issue with the Kelly Criterion is that it is not explicitly built to handle simultaneous bets (betting multiple games being played at the same time). There is research exploring how to properly allocate one’s bankroll for simultaneous bets see here, here and here. In our backtesting below we used the following Kelly Criterion based staking strategy:
We divide our current day’s bankroll by the number of games on that day:
Here are the model results if you invested $1,000 in 09/10 and then continued to reinvest your winnings each season. Profit each season is noted on the left of the graph at the corresponding dotted line.
Here are the model results if you started each season with a $1,000 bankroll and did not reinvest your winnings. Profit each season is noted on the right of the graph.
As you can see in the charts above, the returns are not linear, there will be ups and downs throughout any given season. To further demonstrate this point we looked at the data that produced the 2nd chart above (start each season with a $1,000 bankroll and did not reinvest your winnings). We broke the data into consecutive 5 day segments:
We implore people to use this information smartly. Make sure you find a staking plan that suits your risk appetite and your overall bank size. Do not bet beyond your own personal means.
All seven years previously have shown to provide value, but the future is not guaranteed. We do not sell this service with the promise of profitability. We do sell this service with the promise of providing the same analysis the generated results detail above.