Methodology

Elo is our prediction engine. It learns only from free, open Cricsheet ball-by-ball data, with a strict time cut so it is never graded on a match it trained on. Everything below is measured on seasons it never saw. For entertainment and analysis only; not betting, odds, or tips.

How a call is made: two tiers

The honest pre-match signal is very different between international and franchise cricket, so the confidence bar is set per tier.

Internationals

T20Is, ODIs and their women's equivalents carry genuine signal, so Elo makes a real call and only abstains inside a narrow 45-55% band (the alpha 0.25 conformal tier). On the held-out test it called about 73% of matches at about 66% accuracy.

Franchise T20

The IPL and its peers are close to a coin flip before the toss (IPL accuracy-on-calls was 50.7%). Elo keeps a wide 36-64% band and mostly says "too close to call", showing only a quiet lean. The real edge is the live in-match win probability (Brier 0.162), which the match page leads with once play starts.

The rule is a symmetric conformal set (abstain when max(p, 1-p) is inside the band); the band is the qhat learned per tier. Every surface leads with the probability; the confident "call" badge is the exception.

Models and held-out metrics

ModelVersionHeld-out headlineCard
Pre-match win (T20)v1.0-20260709-t20-k96Brier 0.220 vs 0.231 Elo-only baseline; accuracy 62%; ECE 0.031 (held-out test years).card
Pre-match win (ODI / WT20 / WODI)same family, per-format artifactsBrier 0.240 (ODI, barely beats Elo), 0.182 (WT20, the strongest), 0.226 (WODI, Elo-only).card
Live in-match win probabilitylive_meta family (per format, per gender)Brier 0.162 (T20), 0.197 (ODI), 0.164 (women's T20), 0.180 (women's ODI); Test win/lose/draw log-loss 0.692.card
Score simulator (Monte Carlo)matchup-sim v1Calibrated, not point-accurate: PIT mean ~0.50 and ~80% interval coverage on held-out seasons across T20/ODI, men's and women's.card
Fantasy projection (fantasy-v1)sim-derived expected pointsHeld-out IPL+T20I 2025: Spearman 0.230 vs 0.174 persistence (a modest, real ranking lift); top-11 overlap a tie; MAE marginally worse. A simulation-based projection, not a guarantee.card

Honesty

  • Time-based splits only: models are graded on later seasons than they trained on, never on shuffled data.
  • Elo's live scoreboard on the home page is scored by Brier on those held-out matches, split by tier, so the misses are visible too.
  • The archive has real gaps; see coverage for exactly what is included and the Afghanistan exclusion.
  • Live win-probability slope bands are set on evidence (a match-clustered bootstrap CI plus measured season drift), not hand-tuned. The current policy flags women's T20 live as marginally over-confident (calibration slope 0.886) and says so until a retrain corrects it, rather than hiding it.
  • Not betting, odds, tips, or investment advice.