IPL Crunch '26 — 17 Seasons of Ball-by-Ball Data Decoded

Every IPL fan has opinions — but nobody backs them with numbers. The gap I identified: most cricket analysis is anecdotal, built on commentary narratives rather than ball-by-ball evidence. Does winning the toss actually matter? Which phase of play separates winners from losers? Who are the true best performers across the full IPL era? I set out to answer all three with 17 seasons of ball-by-ball data — 950+ matches, 200,000+ deliveries — and let the numbers challenge the assumptions everyone holds as fact. Process Started by parsing raw JSON from cricsheet.org — 1,235 match files — into a flat ball-by-ball CSV using Python. This taught me how nested sports data is structured before I ever touched analysis. Cleaned season labels (fixing slash-format strings like "2020/21"), derived missing boolean columns like is_wicket from wicket_kind, and deduped to a match-level table for toss analysis. First attempt at phase analysis pooled both innings — a methodological flaw I caught and corrected by splitting 1st vs 2nd innings separately. Also caught a narrative error where my written finding contradicted my own code output on toss decision win rates. For the Impact Player era split, I discovered the gap between pre and post-2023 death over scoring — a structural finding the brief didn't ask for but the data Results Three data-backed findings: (1) Toss winners win ~50% of matches — statistically random — but captains who choose to field win 53.7% of the time. (2) Death overs produce a run-rate gap of 0.295 between winning and losing teams — nearly double the powerplay or middle overs gap. (3) Chasers win 55.6% of IPL matches, yet 65.9% of toss winners choose to field — suggesting the instinct is correct but the coin is irrelevant. Dataset spans 17 seasons. All analysis reproducible via public Kaggle dataset with a single URL load. Reflection I'd build the innings split and Impact Player era analysis from the start rather than adding them later — it taught me to think about confounding variables before writing findings, not after. I'd also validate every written narrative cell against the code output before treating the notebook as done. Next time I'd add a venue-level analysis — toss advantage varies significantly by ground, which is a richer story than the aggregate number alone.