Par-score / projected total.

Two ways in, pick yours.

The signal, in full.

Pass the match state; receive a complete, model-backed read.

1. 01

   Projected total

   The model's predicted final first-innings score from the current state, trained on the real scoring curve, so it does not blow up early the way run-rate extrapolation does.

   projected_final_total

2. 02

   Naive baseline

   A naive run-rate extrapolation is returned alongside the model projection, so the model's edge over the method everyone else uses is explicit.

   naive_extrapolation

3. 03

   Confidence band

   Every projection ships with calibrated P10-P90 bounds (85% coverage) that widen early and tighten through the innings, an explicit read on how much to trust the number.

   low_p10high_p90phaseband_width

Fit to the real scoring curve.

Fit on a large archive of completed professional matches, then validated on unseen games. The intelligence lives in the model.

• Real scoring curves

  Trained on the actual run-accumulation shape of thousands of T20 innings, which is non-linear.

  trained

• Gradient-boosted regressor

  XGBoost over state features, phase-routed quantile ensemble. Overall MAE 18.4. Calibrated 85% P10-P90 band.

  proprietary methodology

• Venue context

  Optional venue input adjusts for high-scoring vs bowler-friendly grounds.

  feature-engineered

• Out-of-sample MAE by phase

  Powerplay 26.4 runs (n=25,222). Middle 17.5 (n=37,219). Death 8.5 (n=19,783). Overall 18.4. Held-out test set, match-level split. Weakest early, when a projection is most valuable to a betting market, and tightest at the death. Published as a table because a single figure would hide where the model is strong.

  out-of-sample

What comes back.

{
  "projected_final_total": 179.5,
  "low_p10": 164.6,
  "high_p90": 201.2,
  "phase": "death",
  "band_width": 36.6,
  "naive_extrapolation": 170.6
}

Use cases.