The Empty Payload: Why 'I Don't Know' Is Cricket Analytics' Most Honest Answer

The file that landed on my desk last week had 'N/A' in the title field. 'N/A' in the source field. The one-sentence summary — blank. The list of core viewpoints — blank. And the information-points field, the spine of the whole structure, was an empty pair of square brackets. Of fourteen fields, exactly one carried a value: the domain label, reading cricket_world.

The skeleton was fully intact. Headings sat where headings belong, table borders were drawn, and the cells were silent. I looked at that empty form for a long time. Fifteen years on a news desk and eight in a model room taught me one thing — an empty cell never stays empty on its own. Somebody fills it with an assumption, and then sends it out as data.

It is time to test that assumption.

Context

Stage-1 extraction has a simple job. Take an article, pull the title, record the source, classify the type, compress it into one sentence, mark the author's stance, note the purpose, extract the information points, separate the viewpoints, list the entities — which team, which player, which board, which tournament — assess time sensitivity, grade source quality. Then stamp a domain label on it.

Of those fourteen fields, one is what I call the load-bearing field: the list of information points. The other thirteen depend on it. A summary of what, if there are no points? An entity of what, if there are no points? Time sensitivity of when, if there is not a single date?

In cricket terms, this is a scorecard where the runs column printed correctly and every batsman's name cell is blank. You can tell from the runs that it was a big match. You can never reconstruct who batted, in which over, against whom. And if you cannot reconstruct it, what you write is not a report. It is a guess.

The lesson first became physical for me in the model room in Indiranagar in 2026. Building a PPDA-plus-xG model across all 380 matches of the 2026-17 Premier League, I found that the machinery of the model was not the point — the completeness of the input rows was. In the matches where over-by-over data was whole, a repeatable pattern emerged: sides whose PPDA climbed above 11.0 after the 60th minute conceded 0.42 more xG in the final fifteen. My first hundred live positions under that filter closed 68-32. The filter worked because the rows underneath it were not incomplete. One empty row would have wrecked the arithmetic.

In cricket that distinction is vivid. Shakib Al Hasan's 606 runs and 11 wickets at the 2026 World Cup is a number with a tournament attached, a date range, a match count, and a checkable home in the ICC's official tournament statistics. Whatever claim I build on it, the claim has a floor. Against that, 'a source close to the team has indicated' has no tournament, no date range, no verification. Neither sentence looks like a number. One is a number. The other is a rumour.

The real problem is that an empty report has an economy. A piece is wanted every day. A headline is wanted every day. On the days when no data arrives, the space still has to be filled. Some people fill it by promoting the label into evidence — it says cricket_world, therefore it was about cricket, therefore I can write whatever I already know about cricket. That is the most dangerous jump in the trade.

Core

So what caused the empty payload? Three possibilities, ranked by probability.

First, pipeline failure. Fetch, parse, or extraction — one of the three stopped working. This is the most likely, because the bones of the schema survived while the muscle did not. Even a genuinely empty article normally leaves a title or a URL behind. Here there is neither.

Second, a paywall, a JavaScript-rendered page, or a bot-block. The server handed someone a page, but the page held no readable text.

Third, the article really was empty. The least likely of the three.

Diagnosis, though, is not the real issue. The real issue is that a half-filled report is far more dangerous than an empty one. An empty report announces that something broke. A half-filled report does not. It enters your feed, settles into your database, feeds your training set, and six months later someone makes a decision on that row.

A number without a sample size is just a rumour with a decimal point.

Take the domain label. It reads cricket_world where the schema wants Cricket. Many would call that a naming question. It is not a naming question; it is a join question. I have watched this error repeat in cricket. One feed calls the match 'BAN v IND', another calls it 'Bangladesh v India'. The system reads two separate rows and counts the same match twice. The match count inflates, the win-probability model learns from a phantom fixture, and that model's output ends up as a graphic on television.

I watched that March 2026 India-Bangladesh T20 World Cup match at the Chinnaswamy from the stands. India won by one run, and Bangladesh needed two off the last three balls. MS Dhoni's run-out of Mustafizur Rahman from behind the stumps in that final over turned the result. One ball, one decision, one front-foot call. A dataset that tags that single ball wrongly has zero decision value, however large it is.

So bring the label question back. Wrong label, no join. No join, duplicates. Duplicates, an inflated sample. An inflated sample raises confidence without raising accuracy. And when confidence rises faster than accuracy, analysis turns into prophecy, and prophecy turns into a loss.

The model is not a prophecy. It is a lamp, and lamps cast shadows.

Duckworth-Lewis-Stern is a good example of that shadow. It is among the most powerful tools in the sport, because it does not merely describe a match — it decides one. Yet if the rows feeding it are incomplete, if an over's output is tagged wrongly, if a broken innings is not flagged separately, the resulting error belongs to the input, not the model. The user cannot see the difference. The user sees a number, and the number makes a decision.

One more thing belongs here, and it surfaces daily in this transfer window. The numbers we have argued about in recent IPL auctions — Mitchell Starc at 24.75 crore in 2026, Pat Cummins at 20.50 crore, Sam Curran at 18.50 crore in 2026 — are traceable. They can be checked against the official auction record by date. The report saying 'three franchises are circling this player' has no record behind it, only a source, and the source has no name.

Agents are the largest invisible cost in this market. The noise they manufacture distorts the price and the squad requirement at the same time. In cricket the machine runs like football's, only faster, because the auction calendar is short and the window shuts on a hard deadline.

Every transfer is a bet on a system, not just a player.

And the heatmap. Over recent years we have started treating bowling heatmaps and pitch maps as proof of analysis. The trouble is that a heatmap can tell you where the ball landed and cannot tell you what role the bowler was playing. The new-ball enforcer in the powerplay and the death-over yorker specialist can produce near-identical heatmaps while doing entirely different jobs. Without the role, a heatmap is decoration — reading tea leaves with better software.

Contrarian

Now the part where I argue against myself.

The plain fact is that an empty report is not a shortage of news. The empty report is the news. A pipeline that cannot produce even a headline has told me something positive about itself. If I quietly file it as 'source unavailable' and move on, I miss the actual event.

That is where the danger sits. Silence is easy; not staying silent is tempting. The reader is not waiting for a filled-in space. The reader is waiting for a story. A story needs data. And when there is no data, people fill the cells with inference. That is how an 'analysis' gets built with not one verifiable point inside it and a complete-looking exterior.

The second danger is subtler. Seeing cricket_world in the label, plenty of writers will assume the piece was about cricket, and will build their analysis on that assumption. That is precisely the error I have made, and it is the one that has cost me most.

In 2026 I published a full 64-match model before the World Cup in Russia. It gave Croatia a 3.2 percent chance of reaching the final, because it overweighted their qualifying xG of 1.31 per game and gave almost no weight to shootout and extra-time resilience. Croatia reached the final. I lost 41 units on outright positions. For eleven days after the final I rebuilt: keeper save data specific to shootouts, substitution patterns in extra time, and a full published retraction with the error log attached.

That episode gave me a line I still use, and I know its limits: in 2026, Croatia taught me that heart is an unlisted variable. I also know the analogy does not transfer cleanly to cricket. Ninety minutes of football and three formats of cricket are not the same object. In cricket, 'heart' is often measurable — death-over economy, fourth-innings averages, super-over strike rates instead of shootouts. What can be measured does not need to be called miraculous.

So I come back. In a report with zero information points, my job is not to guess. My job is to stop, and to write down why I stopped.

I keep a ledger of every wrong number. It is my most honest teacher.

I will admit this much: stopping is not easy. At 55, with eight substantial experiences behind me, the trade keeps teaching you one sentence — we want your opinion. Nobody sends you an empty cell. They want a full one. Under that pressure, plenty of people fill the cell. I filled it once myself, and I paid for it.

One more thing. During the empty-stadium matches of 2026 and 2026, we all assumed home advantage would shrink. It did not. Empty stadiums did not remove home advantage. They exposed how much of it was noise. The empty payload did the same thing to the pipeline — strip out the noise and the structure is visible. Here the structure is clear: fourteen fields, one label, zero information.

Takeaway

I am setting a rule that takes effect at my own desk today. Before anything is published, two things get counted: the number of information points and the date range. If either is zero, the piece does not go. It is a strict rule, and it cuts against me, because my own model once printed 3.2 percent off a sample close to nothing.

The question goes to the reader: of all the analysis published this week, how many would survive that count? How many are complete-looking structures with not one information point inside, their cells stuffed with the author's confidence?

The label was wrong. But because the label was wrong, we now know where the pipeline cracked. The next report will have that crack repaired. That is the actual story.