IPL Crunch '26 — Ball-by-Ball Data Analytics · 289,673 Deliveries

Everyone has IPL opinions. Very few back them with data. This project answers three high-stakes questions using real ball-by-ball IPL data (2008–2026): Do toss winners actually win more? Which phase — powerplay, middle, or death overs — most predicts victory? Who are the top performers when it actually matters? The dataset: 289,673 deliveries across 1,218 matches and 19 seasons from Cricsheet.org. The goal: produce findings rigorous enough to inform real franchise decisions, not just repeat common cricket wisdom. Process Started with raw Cricsheet ball-by-ball CSV. Cleaned methodically: excluded wides from balls-faced (inflates SR otherwise), excluded run-outs from wicket counts (not bowler skill), restricted player rankings to last 5 seasons only (retired players skew all-time data), and used innings 1 only for phase correlation to avoid chasing selection bias. Built phase aggregations (Powerplay 1–6, Middle 7–15, Death 16–20) per match, computed Pearson correlations against match outcome, and calculated win-contribution % per player. Then trained 3 independent ML models — Random Forest, Gradient Boosting, Logistic Regression — on pure in-match features with zero toss metadata. All three independently ranked Middle Run Rate as the #1 predictor. Results Toss winners win just 50.5% — statistically a coin flip. The decision (field vs bat) matters 9.4pp more than winning the toss itself. Middle overs (7–15) have the highest Pearson correlation with winning (r=0.333) AND the largest run differential (+11.7 runs). All 3 ML models independently confirmed this — Random Forest at 69.3% test accuracy. Surprise finding: field-first win rate collapsed from 65% (2016) to 43% (2023) as adoption hit 80%+. Winning chasers score fewer death-over runs than losing chasers — proving the game is won in the middle, not at the death. Reflection I'd incorporate venue-level toss analysis — the aggregate field-first advantage masks venue-specific patterns (some grounds strongly favour batting first). I'd also build a second-innings phase model separately, since chasing dynamics create fundamentally different run-rate incentives that a single model can't fully capture. Finally, I'd add player consistency metrics (how often a batter performs in wins vs losses across seasons) rather than just cumulative win-contribution %, which can be biased by team quality.