Three Numbers and a Gate: Where Data Integrity Stops Cricket's Transfer Window

At half past eleven on a Thursday night in a Motijheel office, a forty-two page file lands. Twenty-one pages of scanned scorecards, twenty-one pages of typed summary. The man who sent it wanted one number before Monday, and he had already chosen which number: tell me only how cheap this seamer is at the death.

I have a rule, in place since 2026. No model runs before provenance. Where did the number come from, who typed it, which filter sits on top of it, which innings were dropped. I learned that rule from an expensive mistake I will come to.

Forty-one claims in the file. Three failed the gate.

Claim seven: a seamer's death-overs economy of 6.8 across the last three seasons. The official scorecards give him eighty-four balls at the death in that period — fourteen point three overs. Claim nineteen: a top-order batter's National Cricket League average of 47.6. Reconstructing it, the not-out convention is applied wrongly, and the two innings of one rain-affected match have been counted twice. Claim thirty-one: an overseas batter's T20 strike rate of 148, taken from a league whose average first-innings score runs roughly forty higher than our domestic competition.

All three fail. And the recommendation on the last page rests on exactly those three.

Context: what the file is, and why now

Bangladesh's domestic game has no formally printed transfer window, yet it has one in practice. Before the Dhaka Premier Division season, clubs select players and agents work the phones. Before the Bangladesh Premier League, franchises submit retention lists. The board announces central contracts once a year. Money is fixed in those three places, and in all three places someone now asks for a number that nobody asked for ten years ago.

The reason is simple. Domestic crowds are thin and broadcast is limited, but budgets have grown. A club's wage bill, a franchise's retention cost, an agent's commission — all of it needs a slide in the boardroom. A slide with a number ends an argument. A slide without one lets the argument run, and a running argument shares out the blame.

This is where the politics of data begins, and where my job becomes difficult, because I am asked to supply numbers in the very room where the number's task is not to tell the truth but to legitimise a decision.

I read the supply chain in four layers. Official scorecards first: ball-by-ball, usually the most reliable, often incomplete at domestic level — no record of who bowled which over, of field settings, of dropped catches. Broadcast second: where there is a camera there is footage, where there is none there is memory. Commercial providers and private trackers third: fast, polished dashboards that do not publish their method. Agents and intermediaries fourth: here the number is a product, and what sells is not the product's quality but its appeal.

Four layers, four failure modes. Incompleteness, memory, opacity, intent.

In 2026 I built my first xG model, in football, from a small office in Motijheel, for the Bangladesh Premier League. I had fifteen years of observation in hand and still spent six extra weeks refining it, missing the mid-season deadline. Tracking Abahani Limited Dhaka's title run, I found their 2.4 xG per match was the highest in the league while their goals were 1.8. A gap of 0.6. The coaching staff dismissed it at first. After the Federation Cup semi-final, where they generated 2.7 xG and lost 0-2, the phone rang.

What those six weeks taught me transfers directly to cricket: process and outcome are separate objects, and a number's job is to explain process, not outcome. Football's xG is the language of process. Cricket's version of that language is subtler, because cricket's underlying record is abnormally complete.

That completeness is the trap.

Every delivery is logged. Every run, every wicket, every extra. As a result cricket's analysts rarely confront model uncertainty and constantly confront interpretive uncertainty. The distinction matters. Model uncertainty says my arithmetic may be wrong. Interpretive uncertainty says my arithmetic is right but the question it answers may be the wrong question. The second is harder, because you can no longer blame the scorecard.

I build models the way monks copy manuscripts: slowly, and with fear of error. Cricket needs that fear more than football does, because here the raw material looks so clean that verification feels unnecessary.

The core: three numbers, three kinds of death

Start with claim seven, the simplest and the most common.

Eighty-four balls. Fourteen point three overs across three seasons, four or five overs a season. A death bowler bowls sixty or seventy overs in a single T20 season; here fourteen overs have been asked to carry a three-year trend. The figure is not false. It really is 6.8. The problem is that it measures the fortune of those fourteen overs, not the bowler's skill.

Let us measure the uncertainty, because without it the rest of the conversation is decoration. Suppose his true rate is nine an over — unremarkable for a limited-overs seamer in domestic cricket. Expected runs across fourteen overs would be 126. Taking a normal over-to-over spread of roughly five runs, the standard deviation of the fourteen-over total is near nineteen. Ninety-five runs sits about one point six standard deviations below expectation, which under a normal approximation occurs with a probability near six per cent — roughly one time in eighteen.

A bowler whose real economy is nine will produce a fourteen-over stretch that looks like 6.8 about once in eighteen attempts. How many such bowlers are there in the league? Forty, perhaps. Then two of them are walking around with a 6.8 they did nothing to earn.

That is the secret of the small sample: it does not lie, it shouts. And the market buys the shout and calls it skill.

I owe a debt here. If I say the number is meaningless, I am overruling the scout's eye. The match that scout watched did contain a yorker in the nineteenth over that pinned a set batter. That happened. The eye did not lie. What the eye saw is what the number summarises — a summary of a summary. At the second remove the yorker disappears and only 6.8 remains.

Now the real question. Suppose the honest projection is somewhere between 8.5 and 9.5, with a wide band. The price you can defend at 6.8 is lower than the price you can defend at 9.2. That is where the number's true function surfaces: it was not built to predict a performance, it was built to move a price. 6.8 is not a forecast. It is a bargaining instrument.

I want to be precise, because this is easily misread. I am not saying anyone falsified the figure. Nobody did. Someone applied a filter — death overs — and the filter sounds reasonable. What nobody asked when applying it was how many balls sit inside it. Nobody asked, because the filter looked good. The spreadsheet was never the enemy; my blind trust in it was.

Claim nineteen is deeper.

An average of 47.6. The first problem is mechanical: not-outs. Divide runs by innings and the average falls; discard not-outs and it rises; divide runs by dismissals, which is correct, and the number changes. The second is that the two innings of one rain-affected match were counted twice, and that match contained his best score.

The third problem is structural. Which innings? Tier One and Tier Two of the National Cricket League are not the same standard of bowling. A batter averaging fifty in Tier Two meets the opposition's best two seamers when he is promoted. The innings that get dropped are usually the failures, and the failures carry the most information.

Here is the discovery, and I did not find the pattern; the pattern found me in the data: an average is not a measurement of a player. It is a measurement of the matches the player was permitted to play.

Consider how inverted that is. We read averages to decide who is good. Yet the average is produced after that decision — who plays which fixture, who is promoted, who gets an A-team call. The average is a consequence of the selection, and we are using it as the cause. Players who fail at the higher level are returned downward and their four Tier One innings vanish from the headline figure. Nobody does this deliberately. It happens because nobody asks which innings were excluded, and why.

I now write that question on the first page of every file, in large letters.

Claim thirty-one is the most dangerous kind, because the number is accurate and the world around it is wrong.

Strike rate is a ratio, and a ratio conceals a distribution. Two batters with 148 can be entirely different players. One makes a steady 20 from 13. The other makes 40 from 27 once and 4 from 8 five times. Same strike rate, different jobs. The first gives a side momentum. The second folds it.

Then the league effect. In a competition averaging 175 in the first innings, 148 is ordinary. In one averaging 135, 148 is excellent. One number, two meanings.

There is a temptation here that I have felt many times: build a conversion coefficient. This league's 148 equals that league's 136. Doing so is an error, because the difference is not only in run volume but in ball type, pitch character, boundary dimensions, powerplay regulations, and above all the standard of opposition bowling. A coefficient compresses five dimensions into one figure and erases the uncertainty of each.

The correct method is percentile mapping with stated uncertainty and a scouting overlay: he sits at the sixtieth percentile in that league, that league's bowling is stronger than ours, so the estimate skews upward and the band is wide. Less elegant, more true.

One lesson I carried from football has saved me repeatedly. PPDA is not a metric; it is a confession of how a team wants to suffer. Cricket's equivalent: a death-overs economy is not a metric; it is a confession of how a captain wants to take risk. But saying that triggers the trap. A proxy is never the thing itself, and the moment I believe I can read the confession, I must attach a falsification test. What happens to this number if the bowler is used in the powerplay? If the boundaries are shorter? Without those two answers, calling him a death specialist is poetry, not analysis.

Now the most important part of the file: the thirty-eight claims that passed.

One was flawless. A career first-class batting average, reconciled to source, dismissals correct, innings count correct. I ticked it.

Then I felt the problem. That single tick made the surrounding claims feel verified. Call it the halo of verification. One number passes and the reader assumes the file has been audited, and a file that appears audited is a file whose conclusions go unchallenged.

The number that passed the gate was the most dangerous number in the file. The three that failed confessed their own defects. The one that passed stood surety for the other thirty-eight.

So what should the model look like? Five rules, none of them theory.

A minimum sample floor, stated in balls and innings before any rate is quoted; for domestic death-overs economy I set it at two hundred balls, and below that the file says insufficient sample rather than a figure. A rival-hypothesis rule: before writing any conclusion, write the strongest alternative explanation. Percentile mapping instead of raw cross-league comparison. Publication of exclusions — which innings, which seasons, printed alongside the recommendation. And the least popular rule: the performance metric and the wage bill on the same page, because a number that sets a player's price must be seen next to the price.

The contrarian angle: data is not the disease

The conventional reading is that bad data is the disease and better data is the cure. Clean the data, fix the decision.

I do not accept it.

The disease is the demand that a single number settle a human argument.

Ask why the file arrived at all. A club was arguing about a seamer, but the argument was not about the seamer. It was between a former cricketer and an accountant. One had seen the yorker, the other had seen the budget. Both were right and neither would move. A third man brought data, and the data ended the argument, because a number sounds like truth and arguing with truth is rude.

The three failed figures were not arithmetic accidents. They were purpose-built instruments, each answering a question somebody in that room needed answered. 6.8 was needed to make the contract look small. 47.6 was needed to justify promoting a young player. 148 was needed to legitimise an overseas quota slot.

Second counter-intuitive point, closer to my own trade. Verification is a social act, not a mathematical one. When I ticked thirty-eight claims I was not doing mathematics. I was judging credibility. The sender is someone I have worked with for a decade and who has never lied to me. That relationship carried the ticks.

This matters because of Bangladesh cricket's material conditions. We have no public ball-by-ball archive for domestic competition. Not every delivery of a Dhaka Premier Division match sits in anyone's database. Broadcast is limited. So what an analyst calls data is often somebody's memory typed into a spreadsheet.

I do not dismiss memory. I have my own — an ICC Trophy final in Kuala Lumpur against Kenya, the pressure, the arithmetic. That is real data of a kind. But memory is a witness, not a record. Witnesses drift with time; records do not. We pass witnesses off as records because records cost labour and witnesses cost a phone call.

The data did not speak; I had to learn its silence first. And the loudest part of that silence is the information nobody collected, because what does not exist is never discussed.

Third, and least comfortable: the most honest number in the file was the wage bill. Nobody quoted it. A wage bill tells no story, only a limit. Stories sell; limits do not.

In 2026, when stadiums emptied, I examined 312 matches across the Bundesliga, the Premier League and our domestic league. Home advantage fell by 0.34 goals per match. The regression pointed at referee bias rather than crowd support. It was the first time data contradicted my own playing experience. I spent weeks rewatching my own tapes from the 1990s trying to reconcile them. I could not. That was painful and necessary, because it taught me that my eye is also a dataset, with its own bias.

If this piece ends with agents lie, it fails. The problem is not the agent. The problem includes me, the man who is happy to tick 6.8 because the number strengthens his side of the argument.

A fair objection: should