ratings-strength-models

Ratings & Strength Models (Sports)

Overview

Estimate latent team (or player) strength from past results, then use those ratings only as they existed before game T.

Ratings are often the strongest simple sports baselines and strong features for logistic/ML models. The non-negotiable rule: as-of only.

Bundled standalone path:

standalone sequential rating construction
held-out evaluation against simple baselines

When to Use This Skill

Use when:

Team power ratings / rankings over a season
Matchup priors for win or margin models
Opponent adjustment
Strong baselines before feature-heavy ML
User says “Elo,” “power ratings,” or “team strength”
NFL/NBA/MLB Elo baselines

Do not use as a substitute for:

Need	Go to
Raw form windows only	`time-series-sports` / `feature-rules`
Full ML horse race	`predictive-modeling`
Validation design	`validation-design`
Season win distributions from ratings	`simulation-sports`

Installation

The standalone Elo builder and evaluator require pandas:

python -m pip install pandas

Parquet input also needs pyarrow or fastparquet.

Workflow

Choose the result signal (win, margin, goal diff, points ratio, EPA, runs).
Choose update style (sequential Elo-like vs batch season-to-date).
Fit/update using only past games.
Export an as-of rating for each team before each game.
Predict matchups from rating differentials (+ home).
Evaluate with season walk-forward (validation-design).
Compare to constant / home / form baselines (baseline-models).
Report params, as-of rule, metrics, limits.
Optional: feed into simulation-sports.

Model Families

Family	Best for	Notes
Elo / Glicko-style	sequential win strength	easy as-of; margin optional
Offense/defense split	score models	attack vs defense ratings
Least-squares power ratings	season-to-date margins	solve system each week
Hierarchical strength	players with uneven n	partial pooling
Possession models	pace sports	sport-specific later

Start with Elo-like or margin power ratings before exotic variants.

The bundled scripts implement only the sequential Elo row builder and its score table. Glicko, offense/defense, least-squares, hierarchical, margin, and possession models are methodology in this skill and require separate code.

Read rating_families.md when choosing Elo, power ratings, or an offense/defense model. Read parameters.md before setting update, home-advantage, carryover, or margin rules. Read evaluation.md before designing forward scoring and baseline comparisons.

Standalone Elo Builder and Evaluator

The builder reads a user-owned CSV, Parquet, JSON, JSONL, or NDJSON table with one row per completed game. Defaults require season, game_date, unique game_id, home_team, away_team, home_score, and away_score; all names are configurable.

python /path/to/ratings-strength-models/scripts/elo_asof.py \
  --input games.csv --out elo_asof.csv --k 20 --home-adv 65

python /path/to/ratings-strength-models/scripts/eval_elo_baseline.py \
  --input elo_asof.csv

The builder saves both teams' pre-event ratings before updating either one. Its two-row-per-game output includes team, opponent, is_home, elo_pre, raw rating_diff, home-adjusted elo_diff, win_probability, and actual. Ties receive 0.5.

The order field must strictly order each team's events. If calendar dates can tie, use kickoff timestamps or a verified event sequence. The evaluator scores home rows once per game and reports season-level Elo log loss, Brier score, accuracy, a fixed 0.5 reference, and—after the first season—a home-win-rate reference estimated only from prior seasons. Pass a matched legal baseline with --baseline-prob-col to score it on the same rows. This is a score table for already-built as-of probabilities, not a full walk-forward parameter-tuning pipeline.

Design Choices

K factor

Higher K: reacts faster, noisier
Lower K: stabler, slower to adapt
NFL/NBA starting point: K ≈ 20 on 1500-scale Elo
Long-season competitions often justify a lower starting K

Home advantage

Add home_adv to home team pre-game Elo when computing expected score
Start fixed; estimate later from data if needed

Margin

Optional multiplier on K using point differential
Diminishing returns (log/sqrt) to avoid blowout domination

Mean reversion / season regression

Regress ratings toward mean between seasons for multi-year series
Important for long panels and roster turnover

Initialization

Common: all teams 1500
Report init explicitly

Other Families (patterns)

Least-squares power ratings (season-to-date)

Each week, solve team strengths from margins-to-date; store as-of ratings before each game. Recompute weekly — never use end-of-season solution for early weeks.

Offense/defense split

Each team has attack and defense parameters; predict score lines, not only winners. Still as-of only.

Hierarchical player strength

Partial pooling when player n is uneven (statistical-modeling mixed models / Bayesian refs).

Evaluation

Primary

walk-forward log-loss / Brier for win probability from rating diff
MAE for predicted margin if modeling spreads

Always compare to in the full methodology

constant baseline
home-only model
form logistic baseline when available

The bundled evaluator implements the 0.5 and prior-season home-rate references plus an optional supplied probability baseline. Constructing home-only logistic, form, or alternative-rating predictions requires baseline-models or another fold-local workflow before calling the evaluator.

Success

beats constant on most folds
preferably competitive with form logistic

Hard Constraints

As-of only — never use post-game rating to predict that game.
Update order must follow real chronology.
Do not fit K on the final test season repeatedly without nested validation.
Report initialization and scale (e.g. mean 1500).
Do not treat ratings as causal player quality without more structure.
Every rating claim must name sport, seasons, event ordering, and parameters.

Anti-Patterns

Final rating applied backward to early games
Huge K chasing noise
Ignoring home advantage
Mixing playoffs and regular season without a flag
Claiming ranking quality without forward prediction metrics
Using form and Elo interchangeably without a bakeoff

Reporting Template

Rating model: Elo
Sport:
Signal: win (margin multiplier: …)
Params: K=…, home_adv=…, init=1500
As-of rule: pre-game rating before update
Validation: season walk-forward
Metrics vs constant / home / form logistic: …
Calibration:
Limits: no roster continuity, no injuries, …
Reproduce: python /path/to/ratings-strength-models/scripts/elo_asof.py --input …

Output Contract

Done means:

As-of rule stated
Params stated
Walk-forward metrics vs baselines present
Limits stated
Repro command present

Worked Examples

python /path/to/ratings-strength-models/scripts/elo_asof.py \
  --input games.csv --out elo_asof.csv --k 20 --home-adv 65

python /path/to/ratings-strength-models/scripts/eval_elo_baseline.py \
  --input elo_asof.csv

For season simulation, pass a documented pre-event probability artifact to simulation-sports; do not treat final ratings as if they existed before earlier games.

Bundled Resources

references/

File	Contents
rating_families.md	family overview
parameters.md	K, home_adv, margin, regression
evaluation.md	how to score ratings

scripts/

File	Contents
`elo_asof.py`	build pre-game Elo table
`eval_elo_baseline.py`	score pre-event Elo probabilities against 0.5, prior-season home rate, and an optional supplied baseline

Related Skills

Need	Skill
Baselines	`baseline-models`
Features	`feature-rules`
Form windows	`time-series-sports`
Predictive ML	`predictive-modeling`
Validation	`validation-design`
Simulation from ratings	`simulation-sports`

Quick Command Card

python /path/to/ratings-strength-models/scripts/elo_asof.py \
  --input games.csv --out elo_asof.csv --k 20 --home-adv 65

python /path/to/ratings-strength-models/scripts/eval_elo_baseline.py \
  --input elo_asof.csv