Skip to main content

📰 Summary (use your own words)

Explaining the methodology behind how 538 simulates their game prediction models.

✍ïļ Notes

  • First iteration relied on [[Elo ratings]]
    • This worked well to keep track of the general strengths of teams over the years
    • But it does not capture any player movements, injuries and resting quickly
  • In 2015, 538 introduced a system called CARM-elo which extends the Elo framework to use their CARMELO player projections
    • This improves the offseason affects to a team's projected strengths
    • It still had trouble with super teams that went easy during the regular season and turns it on during the playoffs - it couldn't properly measure these teams
  • In 2018, fully moved away from Elo and uses player rating system
    • This system is based on RAPTOR metric
    • Using a blend of basic box score stats, player tracking metrics, plus/minus stats to estimate a player's effect (per 100 possessions) on his team's offensive and defensive efficiency
    • This metric forms a prior for each player and it gets updated as the season goes on
    • Then team's strengths can be calculated based on the line-ups they have available during game day
  • Make game predictions by combining the team's talent level weighted by expected minutes played, and multiplied by a regular season or playoff scalar adjustment and converted into expected points scored
    • The expected winning percentage is calculated with [[Pythagorean expectation]] for regular season this has the form
    • winning%=ProjectedPtsScored14.3ProjectedPtsScored14.3+ProjectedPtsAllowed14.3winning \% = \frac{Projected Pts Scored^{14.3}}{Projected Pts Scored^{14.3}+Projected Pts Allowed^{14.3}}
  • The season is then simulated with Monte Carlo and then the predictions are made