Cricket's Most Dangerous Number Is the One Nobody Ever Wrote Down

17 March 2026. Two monitors burning in a Lahore flat, the air-conditioner dead. On screen, a file with nine fields and eight of them blank. Title — N/A. Source — N/A. Type — Unclassified. Summary — empty. The only populated cell read: Domain Label: cricket_world. The frame was alive; the cricket inside it was dead.

That same evening the Pakistan Super League's playoffs were four days away and the league was suspended. Two different events, one shape: the calendar present, the matches absent; the fields present, the numbers absent. That night I thought I was losing data. Eight months later, when a final was played in front of empty seats at Karachi's National Stadium, I understood the actual lesson — what was never recorded is not zero; it is unobserved. And unobserved data always looks like the most reliable data in the room.

Context: the season that stopped

PSL's fifth season was halted on 17 March 2026, four days before the semi-finals. It was completed in November 2026 in Karachi behind closed doors; on 17 November 2026, Imad Wasim's Karachi Kings beat Lahore Qalandars by five wickets in the final (source: PCB/PSL official records).

I had a three-season home-advantage model that weighed every match on the margin and run-rate of home and away sides. The number before 2026: 9.4%. The number for the closed-door phase: 2.1%.

The model did not break on a wrong number. It broke on a column that read Attendance: 0 and which nobody flagged. The journey from 9.4 to 2.1 is one story; how many matches sit behind the 2.1 is another — and the second matters far more than the first.

That week I wrote my first standing rule, and it still sits on page one of every draft: pre-2026 and post-2026 home records do not go in the same table, and the reason goes in print. That is not a rule about data hygiene. It is a rule about the fact that when a variable changes meaning, it has to change name.

Core: four varieties of silence

Administrative silence. When the league stopped, four matches simply "did not happen" — but in the spreadsheet they neither won nor lost. They vanished as rows. A model that assumed 34 matches in the season inflated every rate it produced, because every missing row makes the surviving ones heavier. An empty row does not shrink the sample; it tilts it.

Structural silence. Rain. A 20-over match becomes a 12-over match and the commentary says the sample got smaller. Wrong. The sample does not get smaller; it becomes a different sample. A 12-over chase has no death overs, so the phase-weighting on which my entire Expected Runs model rests simply does not apply. I have watched hundreds of these from the couch: the teams come back out, every column changes, the commentary sounds identical, and the mathematics has become a different discipline.

I built Expected Runs on 41,000 ball-by-ball PSL deliveries, weighted by batter, bowler, phase and field restriction. Of a twelve-part newsletter, eleven parts were read by roughly forty people. Part seven — on the death overs — was forwarded by an analyst in Multan and travelled from there. The lesson was obvious: phase weighting is the model's spine, and a phase only exists if the overs exist. Rain is the biggest enemy of the model that looks the most trustworthy.

Instrumentation silence. The same nine fields, eight of them N/A. The dangerous part is not that no data arrived; the dangerous part is that N/A reads like a fact. "Title: N/A" invites you to infer that the piece had no title. It did not — the piece did not exist. Or it existed and the parser failed to match the DOM, or it sat behind a paywall, or the JavaScript never rendered.

And this is where the damage starts: an empty file does not stop the next step, it poisons it. Had I declared it unanalysable and stopped, that would have been an honest sentence. Had I treated it as fit to draw conclusions from, every subsequent sentence would have stood on a lie — in flawless prose. A pipeline failure is not a story by itself; it becomes one when it returns in the next paragraph dressed as a fact.

My trade produces a daily version of this. When a player's fee is unknown in the transfer market, people quietly read it as free. The habit of treating every absent number as zero is the oldest disease in cricket analysis. Absence is not a measurement; it is a failure of measurement — and the two cannot share a column.

Selection silence. February 2026. Lahore Qalandars' three-person analytics desk. The franchise had finished bottom of PSL 2 with three wins from eight. I built a 214-player valuation sheet — every domestic cricketer with a T20 strike rate above 120 across the previous two National T20 Cups — graded into Platinum, Diamond, Gold, Silver and Emerging bands. A valuation sheet is a promise the league table tests.

The franchise ignored half the sheet. Of the half it used, three Emerging picks outscored its Platinum signing across the season. The most important part of that story is not what I got right. It is the 107 names nobody looked at — they were not proven bad, they were never observed. An unobserved player and a bad player do not belong in the same column, and a franchise that cannot tell them apart never gets to write its own error log.

In 2026 I coded all 64 World Cup matches by hand — shot maps, xG, PPDA, a dead air-conditioner. Then I committed the heresy of porting xG to cricket. Coding all 64 matches taught me where xG stops and cricket begins. The translation halts at three clean boundaries.

A wicket is an event, not a chance. In football, shot quality yields goal probability; in cricket, being out is frequently not a product of shot quality at all — LBW, run out, a catch, a bad call. Start counting wicket "chances" and you have converted an event into a probability. Next, 120 balls is an innings, not a match; football's 90 minutes is one unit, while a T20 innings is 120 discrete deliveries, and that is the real sample problem — a bowler sends down 24 balls in a four-over spell, and no rate should be built on that. Third, bowler-batter matchups are scarce: two players meet four or five times in a tournament, which is nowhere near enough for any meaningful matchup claim.

This is why I still staple a ball count to every claim. A number that cannot survive "how many balls is this built on?" does not make the paragraph. Editors used to call my drafts auditable, and then they stopped cutting them for length.

Contrarian: be suspicious of the beautiful number

Now the part that collides with everything above. Home advantage fell from 9.4% to 2.1% behind closed doors. It is a beautiful result, and beautiful results deserve the most suspicion.

Why? Because the closed-door sample was tiny and sat almost entirely at one ground — Karachi's National Stadium, which is also Karachi Kings' home venue. So "no crowd" and "the home side was strong" changed at the same time. I should have published the interval around 9.4, not the point. I did not. Where the sample is small, print the range, not the point — and state in advance the condition under which the model is wrong, not afterwards.

The "more data is better" reflex needs the same post-mortem. In cricket the marginal value of a fortieth metric is close to zero; the marginal value of a missingness flag is very high. A model that does not know its own empty spaces does not know its own reliability — and in knockout cricket that blindness is most expensive, because one innings is the entire sample and the model will still print a number with total confidence.

Forward

So where does the next start come from? From a season with crowds back and an audit that holds. My condition is already on the record: when the PSL completes its next uninterrupted, full-attendance season, if home advantage returns above 6%, then 2026's 2.1% was about venues, not crowds, and I withdraw the finding. If it stays below 4%, the crowd was doing real work. The gap between those two figures is my model's error band. The date is fixed in advance: within fourteen days of that season's last match.

When the stadium empties, the data gets louder. The question is not what is on the screen. It is what is not on the screen — and who decided, when, and with what excuse, not to write it down.