Build a College Basketball Elo Rating and Test It (Python)

If you built the football Elo, you already know the engine — and the beautiful thing about Elo is that it ports to any sport unchanged. Here we point the same update rule at a full college basketball season, then back-test it to see how well one number predicts the sport. Basketball, with its ~30-game schedules, gives Elo far more data to work with than football does. Full code: scripts/cbb-elo-rating-python.py.

The same engine, new data

Identical math: expected win probability from the rating gap, then nudge by K * (actual - expected). We tune two knobs for hoops — a slightly smaller K = 28 (more games means each one should move the needle less) and a +60 Elo home-court bump (neutral-site tournament games get none):

import sportsdataverse.mbb as mbb
df = mbb.load_mbb_schedule(seasons=[2025]).to_pandas().sort_values("date")
elo = {}
for g in df.itertuples():
    rh, ra = elo.get(g.home_id, 1500), elo.get(g.away_id, 1500)
    adj = 0 if g.neutral_site else 60
    eh  = 1 / (1 + 10 ** (((ra) - (rh + adj)) / 400))
    hw  = 1 if g.home_score > g.away_score else 0
    elo[g.home_id] = rh + 28 * (hw - eh)
    elo[g.away_id] = ra + 28 * ((1 - hw) - (1 - eh))

Back-test it

As we build, we tally whether the higher pre-game rating won — an honest, in-flight measure of predictiveness across thousands of games.

The result

6292 games. Back-test winner accuracy: 72.1%

Top 15 by Elo:
 1  Florida Gators        1829     6  Michigan State    1732
 2  Houston Cougars       1807     7  St. John's        1732
 3  Duke Blue Devils      1785     8  Drake Bulldogs    1728
 4  Auburn Tigers         1754     9  Chattanooga Mocs  1724
 5  UC San Diego Tritons  1733    10  Tennessee Vols    1722

Two things to notice. First, 72% back-test accuracy — even better than football, because more games per team make ratings more stable. Second, the top four (Florida, Houston, Duke, Auburn) were the actual Final Four — your homemade rating found them. But look closer and you'll spot Elo's blind spot: teams like UC San Diego, Drake, and Chattanooga rank surprisingly high. They piled up wins, but mostly against weaker mid-major schedules, and pure Elo doesn't fully discount that. The lesson: Elo rewards winning; it doesn't deeply weigh whom you beat. For that, you want the adjusted-efficiency approach, which explicitly corrects for opponent quality.

Elo vs. adjusted efficiency

• Elo is a win/loss model — fast, intuitive, great at prediction, but it overrates teams that beat up on weak schedules.
• Adjusted efficiency uses scoring margin per possession and explicitly adjusts for opponents — better at separating the genuinely elite from the merely undefeated-against-nobody.
• Use both. When they disagree (a high-Elo, modest-efficiency mid-major), you've found a team to watch in March — see what upsets have in common.

Tune and extend

• Back-test your knobs. Try K of 20, 28, 40 and keep what predicts best on held-out games.
• Margin multiplier (capped) so blowouts move Elo more than buzzer-beaters.
• Women's game: swap mbb for wbb — same code, same season convention.