The ₹11.5 Crore Lesson: IPL Auctions Need Phase Splits, Not Strike Rate

January 2026, Bengaluru, the mega auction. Jaydev Unadkat's name came up at the bidding table. My value-over-replacement sheet had priced him at ₹4.2 crore. Rajasthan Royals bid ₹11.5 crore. That night, from a Delhi newsroom, I published the gap — Unadkat was the clearest overpay of that auction. Five months later, running my first xG and PPDA model in Kazan, I realised the mistake was not only in the bid. The mistake was in the instrument I had used to measure him: economy rate, with no information about the phase in which it was conceded. After that night I stopped reading the auction as a money game and started reading it as India's actual scouting system. This is the story of that error, and of the correction my spreadsheet had not priced.

The IPL mega auction is the most transparent valuation platform in Indian cricket, because every selection call comes with a public price. Zonal selectors never told you why they preferred one player over another; a franchise spreadsheet shows its reasoning with every bid. Over the past decade that spreadsheet has converted Ranji numbers into premiums — sometimes correctly, sometimes at ruinous cost. Where the old zonal selector leaned on experience and regional bias, today a franchise model sits in the auction room and decides which bowler's ball is the most expensive at the death.

That shift is not neutral. It made selection more democratic, but it handed selection to a new bias — one that reads numbers without their context. The figures on the auction table are usually a whole season's average, in which cheap powerplay overs and expensive death overs are mashed into a single cell.

Economy rate is the metric that does not tell you what question it is answering. A bowler's economy of 9 can be the average of three different bowlers: 8 in the powerplay, 7 in the middle, and 12 at the death. The average becomes 9, and that 9 hides where the bowler actually earns his keep. For an auction this is a serious problem, because a death over costs several times what a powerplay over costs — the match is decided there.

And this is where India's real problem hides. India's second-choice XI is the hardest team to make in world sport, and that surplus of talent is precisely what disguises selection failures — the replacement is always good enough that the mistake never gets caught. Auction money prices that surplus, and often in the wrong place.

In 2026 I built my first detailed model on exactly this flaw — ball-by-ball win probability from scraped 2026 IPL scorecards. May 2026, the final in Hyderabad. Mumbai Indians made 129/8. At the innings break my model gave Rising Pune Supergiant a 68% chance of winning. Mitchell Johnson bowled the last over; RPS needed 11. RPS finished 128/6. Mumbai won by one run.

For the next week I did not defend the model. I rebuilt the death-overs prior and published the calibration failure myself, with the graph. I stopped asking why the 129 was wrong and started asking what it was measuring. The answer was clear: my model was not weighting death-over wickets the way they actually matter. Defending 129, a bowler does not merely need to stop runs; he needs to take risk — and that risk is not recorded in any column of the scorecard.

That lesson became my phase-split model. Three buckets: powerplay (1-6), middle (7-15), death (16-20). A separate baseline for each. In the IPL, a good powerplay economy is roughly 7.5-8.5, middle overs 7-8, and death 9.5-10.5. A bowler who concedes 9.5 at the death is far more valuable than one who concedes 8.5 in the powerplay — but raw economy puts them in the same row.

At the death, stopping runs is not enough. The real currency there is the wicket, and then the dot ball. A bowler who concedes 10 at the death but takes a wicket every fourth ball is more valuable than one who concedes 8.5 and takes none. So my phase-split model combines three things: economy, dot-ball percentage, and wickets per over. Each is weighted separately, then converted into a composite score — and that score, not the tournament-wide economy, should be the basis of an auction bid.

This is where the Unadkat case gets interesting. The overpay was real; the correction was the part my spreadsheet had not priced. Rajasthan bought him as a death-over specialist, but his death-over sample was small — a handful of overs in one season. A death-over average built on a small sample cannot separate noise from signal. My model treated the raw death-over number as seriously as a 60-ball powerplay sample. That is where it slipped.

The right method is to pass every phase number through shrinkage — pull a small sample toward the phase average, more so the smaller the sample. Then compare what remains against replacement level. That is the true meaning of value over replacement: not how good a player is, but how much better he is than the available alternative, and in which phase.

In the Ranji circuit, domestic coaches have repeatedly told me that their best death bowler is often the cheapest buy, because his overall economy average hides his work. For me this was not an anecdote — it was human evidence of the flaw in the model I was fixing. The number and the dressing-room voice were saying the same thing, only in different languages.

I read every match as market versus model. Before an auction I write down both the projected closing bid and my own number, so readers can see which side I am on. In 2026 the market put ₹11.5 crore on Unadkat; my model stalled at ₹4.2 crore. The truth was somewhere in between — and that truth opened slowly over the next two seasons.

Here is an uncomfortable truth I repeat every time: the model did not break; it met a score it had never been taught to fear. Unadkat's overpay was the market's error, but my ₹4.2 crore number was not the whole story either. The market can be inefficient and still be smarter than my first explanation — a lesson that cost me.

This is where correlation and causation part ways. That expensive bowlers often underperform does not mean the price is the cause. The real cause is usually that auctions bid on small-sample death-over numbers, and the next season regresses those numbers toward the mean. That is regression to the mean, not a lack of talent.

So I publish the error bars alongside every claim. The things the model cannot see — temperament, fitness, dressing-room fit — I name first, so the reader catches them before I do. Before I argue with the line, I make the line show me its prior. Any claim that arrives without a baseline, a sample, or a caveat is incomplete in my eyes — even if it is mine.

I repeated the same lesson in 2026, when, after the pandemic, home win rates in matches played behind closed doors fell from 43% to 34%. There too the metric had not changed; what changed was the environment the metric was not measuring. Auctions work the same way — if a changing environment (pitch, ball, format) has no place in the model, the same old average becomes expensive again.

In the next auction, the franchise that bids on phase-adjusted, shrunk death-over economy rather than raw economy will stay a step ahead of the market. The question is not which bowler has the best average. The question is in which phase that best average was built, over how many overs, and against whom. The team that asks this will avoid the ₹11.5 crore mistake. The team that does not will look at the same graph next season — and say the model was wrong.