- Neural Net Bets
- Posts
- 2023/24 Alpha Test Summary
2023/24 Alpha Test Summary
How did the Neural Net Bet's beta model perform last season, and what did the testing betting phase return ?
Table of Contents
Model Summary
Key Stats
The key results of the 2023/24 Alpha test of IcePick (Neural Net Bets flagship model) are summarized in the table below. The rest of this document will go into more details about these results.
Metric | Value |
|---|---|
Games live predicted | 586 |
Win % / model accuracy | 64.33 % |
Record | 377 - 209 |
Total deposit | 13.51 units |
Final balance | 87.26 units |
Profit | 73.75 units |
Return on investment/deposit (ROI) | 545.9 % |
Performance
In the 2023/24 Season, the Neural Net Bets model live predicted 587 games.
The Win/Loss record was 378-209, meaning the model was 64.4% accurate.
During this same time, the sportsbooks were only 60% accurate (favorite won only 60% of the time). This difference is where our consistent returns are able to come from.
Out of these games, we predicted 73 underdogs to win. Out of these games, 45 won (9 of those in overtime), and lost 29 lost in regulation (5 of those in regulation). This means we do not have a significant drop off in accuracy, predicting underdogs with 61.64% accuracy.
When the favorite team was predicted the accuracy was 64.7% (332 - 181).
Not only does the model predict who is going to win, but it also gives a percent chance (and this compared to the implied odds from the bet results is how we decide how much to bet).
When predicted to win with a >60% chance, accuracy was 73% (153/211)
When predicted to win with a 55-60% chance, accuracy was 63% (111/176)
When predicted to win with a 52-55% chance, accuracy was 59% (70/118)
When predicted to win with a <52% chance, accuracy was 53% (41/81)
& keep in mind many of these wins would be predicted underdogs
The entire season (including beginning and end when the model is less performant) the accuracy was 61.6%.
Betting Summary
In the beta run / 2023-24 NHL season, 586 games were bet on (moneyline bets, meaning betting on who wins the game).
73 of these bets were on the underdog, which would pay out 2.17x on average
513 of these bets were on the favorite, which would pay out 1.59x on average
It was generally found that it was profitable to bet 1 unit on a parlay of all games with >60% chance to win, although this is not as statistically backed as the moneyline bets and are therefore not part of the official bets provided.
Over 10 different betting strategies were tested and compared (1 was actually used while the payouts of the others were calculated).
These strategies all revolve around setting a bet unit for yourself (which can remain constant for the whole season, or grow as your bankroll does). The models predicted win chance for a team and the implied win chance from the sportsbook line are then leveraged to alter this base unit.
Essentially, when the model is more confident when the sportsbook we bet more and when it is less confident we bet less.
Betting results
Below is an image of the bet returns (in bet units) over the course of the beta testing, as well as a summary.
The algorithm will output the team to bet on and how much to bet, for each game. Money must only be deposited when your balance cannot cover the total bet amount for the day ( & as you can see, there was only one point in the trial where we did not have enough bankroll to cover after our initial deposit).
The table below highlights the results of the betting season, and gives some example monetary values assuming a constant $50 unit is used.
Metric | Value | Value with $50 units |
|---|---|---|
Total Deposit | 13.51 units | $675.5 |
Total Balance | 87.26 units | $4363 |
Total Profit | 73.75 units | $3687.5 |
Return on Investment | 545.9 % |
Disclaimer
Keep in mind that this 545% return on investment is not guaranteed. Although it is almost guaranteed that the model & betting strategy will net a profit over the course of the season, the amount of profit is unknown. It is entirely possible that the return will simply be 100%, but is unlikely to be below this. In none of the 100’s of different models tested lasts season, was a negative return achieved.
The way to think about Neural Net Bets is not sports gambling, but a stock. Even a 100% return on an investment over the course of a few months is something to celebrate and would be regarded as a sweet return on your investment. This is how you need to think about the returns from the Neural Net Bets model.
Testimonials and Beta tester returns
Below contains some screenshots from some of the beta testers betting accounts, to put some real examples behind the results provided above. Note that the beta testers did not bet on all games (started at some point in February and ended in mid April). I myself am unable to provide any screenshots as I was testing if it was worth always betting in the sportsbook that has the best odds (& it is not worth all the extra effort)
