Model card · pressure-v1

PENDING TRAINING

Pressure Score v1

A per-ball expected-value model that quantifies batting pressure in T20 cricket. For each delivery, the model estimates the probability distribution over (runs, wicket) outcomes given the ball context (phase, position, run-rate, wickets fallen). The difference between expected and actual outcomes accumulates into a per-innings Pressure Score — positive = the bowling side has exerted above-expected pressure.

Pre-disclosure scaffold. This model card is published in advance of training per §3.3-model (doctrine PR #312). The corpus, feature list, and architecture are specified. Eval results, feature importances, and the training date will be added when the model is trained. No claims referencing this model may ship to production until training is complete (enforced by build-time guard validate-model-card-provenance.mjs).

Purpose

Answers "How much pressure was a batter or bowler under/exerting, accounting for match context?" — a question raw economy or strike-rate cannot answer without context adjustment.

Metric emitted

Pressure Score

Profile pillar

P3 (Form & phase)

Output range

0–100 (higher = more pressure on the batting side)

Training corpus

The training corpus must trace to ball-by-ball deliveries. Any model that cannot produce this chain is not eligible under §3.3-model.

Source

CricketStudio ball-by-ball corpus — IPL historical (Cricsheet CC BY 3.0)

Seasons

2007/08–2025 (18 seasons)

Match count

pending training

Delivery count

pending training

Train/val/test split

Season-based split — most recent 2 seasons held out for validation and test; earlier seasons used for training. Prevents data leakage across season boundaries.

Notes

IPL 2026 ball-by-ball corpus excluded from training (it is the primary inference target). Cricsheet CC BY 3.0 license. No proprietary or undisclosed data sources.

Features (11)

All features are derived from the ball-by-ball corpus — no external or proprietary data.

Feature	Type	Description
`over`	numeric	Over number within the innings (1–20)
`ball_in_over`	numeric	Delivery within the over (1–6)
`phase`	categorical	Innings phase: powerplay (1–6), middle (7–15), death (16–20)
`balls_remaining`	numeric	Balls remaining in the innings at delivery time
`wickets_fallen`	numeric	Wickets lost by batting side at delivery time
`runs_scored`	numeric	Runs on board at delivery time
`run_rate_required`	numeric	Required run rate at delivery time (second-innings only; 0 for first)
`innings`	categorical	First (1) or second (2) innings
`venue_par`	numeric	Historical par score at the venue (from venue-signature data)
`batter_handedness`	categorical	Right-handed (0) or left-handed (1)
`bowler_type`	categorical	Pace / off-spin / leg-spin / left-arm-spin / left-arm-pace

Architecture & evaluation

Architecture

XGBoost (gradient-boosted trees) — separate models for runs regression (target: runs off the delivery) and wicket classification (target: binary wicket outcome). Two-headed design avoids conflating run-scoring and dismissal risk.

Targets

runs_off_delivery (regression), wicket_on_delivery (binary classification)

Evaluation metric

R² (runs regression), log-loss + AUC-ROC (wicket classification)

Eval results

pending training

Inference method

For each delivery in a captured innings: compute (actual_runs − expected_runs) + (actual_wicket − P(wicket)). Sum across all deliveries a bowler bowled (or a batter faced) in the innings. Normalise to 0–100 scale (100 = maximum pressure exerted per the training distribution). Positive = bowling-side advantage; negative = batting-side advantage.

How claims cite this model

Per §3.3-model, every claim derived from this model carries the model name, version, and a link back to this page in its provenance footer and in its ClaimReview JSON-LD block.

Atomic claim format

[Player] exerted a Pressure Score of [X] (Pressure v1, [N] deliveries, IPL 2026)

JSON-LD isBasedOn

"isBasedOn": { "@id": "https://players.cricketstudio.ai/methodology/model/pressure-v1#model" }

Governance

Training date

not yet trained

Retrain policy

Retrained each IPL season on the full historical corpus (season-based split updated). Model card version incremented on every retrain.

Deprecation policy

This version remains citable after a new version ships. Claims citing pressure-v1 continue to resolve. Deprecated status is set when a successor version covers the same inference scope.

Contact

hello@cricketstudio.ai

Version history

Version	Date	Notes
`vv1`	pending training	Initial release. Two-headed XGBoost on IPL historical corpus. Features: phase, wickets, RRR, venue par, bowler type.

Known limitations

T20-only. The expected-value baseline is derived from IPL data; applying it to other T20 formats (MLC, BBL, T20I) requires retraining or domain adaptation — not yet implemented.
Venue par feature uses historical par averages (≥3 captured fixtures floor). Venues with fewer fixtures receive a league-average default.
No batter-vs-bowler matchup feature in v1 — this creates systematic bias for matchups with large historical divergence from the population mean.
Expected values reflect historical distributions; unusual batting orders or pitch conditions outside the training distribution will skew scores.
Pressure Score is a relative metric within the training distribution — comparisons across eras (e.g. 2010 vs 2025) require era-adjustment, not yet implemented.
Pending training: eval results, feature importances, and calibration plots will be added when training is complete. This card is a pre-disclosure scaffold per §3.3-model.

Training corpus

Features (11)

Architecture & evaluation

How claims cite this model

Governance

Version history

Known limitations

References