1. Elo leaderboard
Each team starts at 1500. After every completed game the winner takes points from the
loser. The size of the swap depends on how surprising the result was: a heavy favorite
winning by 2 barely moves the needle; a 20-point underdog blowout reshapes the ratings.
The expected-win term folds in home court as a fixed additive advantage, so the model
doesn't treat a home win by a 1480 team over a 1520 team the same way it treats that
matchup on a neutral floor.
Margin of victory multiplies the base K-factor update, which is the whole
point of using Elo here instead of a win/loss counter: a 30-point win carries more
signal than a 2-point win, and the formula should reflect that.
P is expected home-win probability, r are the two ratings,
h is the home-court advantage in rating points, and s is the
scale factor (typically 400). The update Δ multiplies the base
K by the margin multiplier m and the actual-vs-expected gap
y − p.
- Ratings update game by game, in chronological order, replaying the whole season each refresh.
- Home court is baked into the expected-win formula, not applied as a post-hoc bump.
- Margin of victory scales the
K-factor, so a blowout moves ratings more than a one-point win regardless of who was favored.
- Ratings regress toward 1500 between seasons, so a single bad year doesn't permanently sink a team in the model.
2. Home-win predictor
The predictor trains on every completed game in the window before the target date, then
builds one matchup row for the home and away teams you pick. Two models train on the same
engineered feature set and run side by side: a logistic regression (linear, readable,
coefficients you can inspect) and a random forest (nonlinear, captures interactions the
linear model can't).
Keeping both visible isn't indecision. Linear and tree models weigh the same inputs
differently on a given matchup, and when they disagree the context table tells you why:
rest, injury load, and strength gap all show up as raw numbers beneath the winner call.
- Season win percentage tracks overall team quality over the full year so far.
- Team strength blends win rate with average scoring margin, which is a better quality signal than either alone.
- Recent form is a rolling window over the last few games, capturing whether a team is hot or cold right now.
- Rest days count the gap since each team's last game, since a team on a back-to-back plays differently than one on three days' rest.
- Injury features count out, doubtful, questionable, and probable players per team.
- Injury impact goes a step further, weighting status severity and estimated absence length so a star out for a week reads differently than a role player out for a day.
The benchmark for the current training window: home-court baseline accuracy
54.6%, logistic
64.8%, random forest
64.5%. The linear model edges
the tree model by about one point on held-out 2025 games.
