ipl crunch3

Standard IPL analysis relies on batting averages, economy rates, and strike r built for Test cricket that measure volume, not context. They cannot distinguish a match-winning 40 under extreme pressure from a comfortable 40 in a dead rubber. Franchise auction rooms misprice players because no tool quantifies clutch performance, phase-specific impact, or momentum collapse. Toss analysis is reported as a single aggregate, hiding dramatic venue-specific variation. The result: analysts describe what happened, never why and franchises make million-dollar decisions on the wrong number Process Ingested and merged 548K+ ball-by-ball deliveries across 16 IPL seasons from Cricsheet Cleaned and standardised franchise name changes, removed Super Overs, flagged rain-affected matches Engineered 7 custom cricket metrics from scratch unavailable in standard cricket databases Built over-level aggregations for phase-wise (PP / Middle / Death) win-factor analysis Trained and compared three ML models with out-of-time validation on held-out seasons Applied SHAP explainability to quantify which features drive win probability predictions Delivered an interactive HTML dashboard, complete Python pipeline, and plain-English report Results Death overs economy is the #1 match-outcome predictor — stronger than powerplay or wickets Toss advantage is 52.4% overall but climbs to 64% at dew venues and drops to 50% at three grounds 34% of PoM awards misidentify the true match-winner by Clutch Performance Index High boundary dependency teams lose more close matches than singles-focused teams Gradient Boosting model achieved AUC 0.79, accuracy 72% on out-of-time validation Go beyond standard cricket stats and build metrics with real franchise decision-making value Reflection 7 custom metrics built from first principles — Pressure Index, CPI, CSI, DOE, PAR, MSS, EUP — none exist in standard cricket data products Out-of-time model validation (train 2008–2021, test 2022–2024) instead of random splits — far more credible for sports data Toss analysis disaggregated by venue, not reported as a single number — revealed the myth hiding inside the average PoM award accuracy tested against CPI — exposed a structural bias in traditional cricket recognition Boundary dependency framed as a vulnerability, not an asset — counterintuitive finding backed by close-match win rate da