okhp3-outcome-modeling-sports

okhp3-outcome-modeling-sports

OverKill Hill P³ · overkillhill.com · github.com/OKHP3

Apply the outcome-modeling core to sports without confusing raw wins with underlying performance. Sports evidence is organized as time-indexed team or player states, opponent-adjusted interactions, and structural events such as injuries, coaching changes, and roster moves.

Scope

Required workflow

1. Load okhp3-outcome-modeling-core and define the outcome, horizon, scoring rule, and as-of date.
2. Identify the sports entities, event grain, matchup relationship, and season structure.
3. Build time-indexed team or player states from only available prior events.
4. Adjust for opponent, venue, rest, schedule strength, role, and sample size where relevant.
5. Treat injuries, trades, coaching changes, and role changes as structural events or regime changes.
6. Separate descriptive performance from predictive features and avoid double-counting correlated metrics.
7. Compare simple baselines such as record, point or scoring differential, Elo-style strength, and replacement value against richer models.
8. Validate chronologically and report calibration, uncertainty, and failure modes.

Sports state model

Use a state vector appropriate to the sport. Preserve raw values and transformed values, with feature metadata. For a matchup, model the relevant difference or interaction between the two states rather than treating each team as an isolated row.

team_state(team, date)
matchup_delta(A, B, date)
outcome(matchup, date)

Do not assume that the best team wins every individual game. Aggregation can reveal persistent strength while realized games retain randomness.