sport.dev is a free platform that uses machine learning models to predict sports.
It uses historical data, team lists, injuries and other model outputs to predict win% and player stats.
The predictions are intended strictly for informational use and not betting advice.
Note:
No sportsbook data or market price data is used in any models. Market price data can get predictions reasonably close, however the aim of this platform is to compose enough signals purely from data to train improving models for each sport.
Known issues:
- Right now every model is designed to work optimally for regular season play roughly after the first 10% of matches have been played.
- Start of season lacks enough recent data for player form and roles/combinations.
- End of season/finals for each sport is also played differently and may be hit or miss. Model weight adjustments for end of season/finals should be ready soon.
- Football Draw prediction weights needs a lot more work.
- Every league in a sport relies on global models for that sport rather than being league specific. Some leagues like French Ligue 1 or Brazilian Serie A do not fit my current euro models very well due to a few things: financial and squad disparities, travel volatility and heavy squad rotations. I'll take a deeper look at this soon when building out league specific steps for football.
- Rugby League (NRL) had model degradation from round 13 to round 23 from an upstream model stall that forward filled stale data into several primary feature inputs. This was difficult to detect because it started during State of Origin (a mid season tournament using the same players) so I had assumed the model drift was from player fatigue or injuries. This has been fixed but it did lower the overall accuracy % for this season.
Credits:
- League and Team logos are only used for reference purposes, no products or services are being sold using those protected images
- Live score data from ESPN
- Weather data from Open-Meteo.com