Bengaluru's Death-Overs Ledger: The Lesson from a 34-Page Post-Mortem, and the Rule That No Single Number Survives Without a Confidence Interval

The first document I handed over at a Bengaluru franchise's data desk in June 2026 ran 34 pages and answered one question: why a side that finished eighth with three wins from fourteen matches conceded at 10.9 an over in the death overs against a league average of 9.3. The file did not raise its voice; it simply kept the receipts. The head coach read two pages and asked for a one-line summary. I have not written a document longer than four pages since, and every analytical claim now carries a three-line header — claim, sample size, confidence. This is the story of that document, and of a finding that still returns every preview: death-overs economy is a number that, without an innings count and a bowling-position context, is not data journalism but arithmetic advertising. First, context, because without it a number loses its witness. This was not a Test match and not an ODI series; it was a fourteen-match IPL league phase, where every ball sits inside the T20 regulation — powerplay six overs, middle overs seven to fifteen, death overs sixteen to twenty. These are three different sports. Comparing a T20 death-overs economy with a Test first-innings economy is like comparing a fifty-over strike rate with a twenty-over strike rate; both figures can be correct and the conclusion still nonsense. In professional analysis, format is the first filter, then venue, then toss and DLS, then players. My 34-page document followed the same order. When the IPL shifted to the UAE in 2026 and sixty matches were played in empty stands, I re-ran nine seasons of models. The home-advantage term that sat at 4.1 percentage points of win probability from 2026 to 2026 collapsed to 1.2. Producers wanted a 'momentum' metric; I declined and re-fitted every model with an environment variable, filing a two-page note nobody opened until the playoffs. The crowd was absent, but the data kept clearing its throat. Now the core finding, clear on page thirteen. That season the seventeenth over went to the side's fifth-best bowler in nine matches. This is not coincidence, it is a structural habit. The best bowler was being saved for the eighteenth and twentieth, while the seventeenth — often an over carrying both a set batter and a new one — went to less experienced hands. The effect on economy was direct: 10.9 was not a failure of bowler quality but the politics of over assignment. I built the claim on three columns: the bowler's death-overs economy in that sample; the batter's set-up in the same over; the captain's field placement. Without all three, 10.9 can be called a number, but it cannot be made to mean anything. I still keep a copy of that document, and its first page notes that every innings carries a different sample weight: if 61 percent of a bowler's wickets in that sample came against batters eight to eleven, then his death-overs economy is not a reliable measure but a measure of opposition weakness. From here comes the angle that cost me two television bookings and earned the quiet trust of three team analysts. At the 2026 Bengaluru mega auction, when Jaydev Unadkat sold for ₹11.5 crore, my valuation sheet had his floor at ₹4.2 crore. The gap was not quality but context. In fourteen innings of death-overs data, a large share of his wickets had come against lower-order batters, and that is why my file carried a ceiling below the market bid. My franchise walked away and bought two seamers at ₹2 crore each; between them they took 21 wickets, and I spent forty minutes after the auction explaining to the owner why the cheap deal was the expensive one. Since that day I have refused to publish a single-number valuation. Every preview now carries a floor, a ceiling and the innings count behind both. I open the auction ledger and find a queue of quiet apologies. Every bid was a sentence someone would later have to defend. There is a caution I repeat every time: correlation is not causation. The link between a 10.9 death-overs economy and eighth place looks clean, but it is probably not as direct as we would like. If that 2026 side had merely bowled better in the death overs, would it have reached the playoffs? Perhaps not. Its middle-overs run rate also sat below the league average, and the slow starts the top order produced in the powerplay pushed the later overs under pressure before they began. The death-overs problem was a symptom of a side whose structure was slipping. I would add that the fourteen-match sample is small — drawing any causal conclusion at that size is risky. One poor innings can drag an economy figure by 0.8. That is why I add a plain line to every analysis: what I am seeing is causation, or only a signal? Sometimes a signal is enough, provided it is called a signal. In 2026 a Bengaluru football analytics startup hired me to port my cricket phase model onto xG and PPDA for the Qatar World Cup. It took six weeks to admit the transfer did not work. Football's phases are not discrete; my fifteen-over blocks shredded ninety-minute matches into noise. I filed late, and my model underperformed the startup's existing one by a margin I still refuse to round down. But I wrote the failure up in full, with the numbers. Since then I keep a dated log of every model that failed and why, and I quote it in my columns. I trust a clean column more than a loud conclusion, especially in football. The transfer market is a rumour mill wired to a very patient spreadsheet. Now forward. If an IPL 2026 auction sees a team bid big for a death-overs specialist, the first question should be: in which overs, and against which batters, did those wickets come? The 2026 Bengaluru post-mortem shrank to two pages on a coach's desk, but its lesson has reached the scouting sheets. The question now is whether we are starting to mistake volatility for value. Perhaps. And the next time a commentator says some bowler's death-overs economy is 8.9, you can ask: across how many innings, against how many batters, and on which ground? All I would add is this — a number that cannot state its limits cannot state its confidence either.