Zero Is Not 'No Data': Silent Pipeline Failure in Football Analytics and the Threshold of On-Chain Proof

August 2026. The press box in Nizhny Novgorod. The 67th minute of Croatia versus England. On my laptop screen Croatia's live xG read 1.4, England's 0.8. Luka Modric had covered 12.8 kilometres, completed 67 passes, and his late pressing had dragged England's PPDA down to 12.9. The match ended 2-1 to Croatia.

I have written about that evening many times, because it is the foundation stone of my professional life. Today I want to write about a different frame from the same evening, one I never published.

For one moment the screen went entirely blank. Every cell, every bar, every label. Then it came back. I assumed the feed had dropped. It had not. My browser was forty seconds behind. The data was arriving. I was not watching it.

Latency and failure are two different things, and telling them apart is the most neglected skill in this trade. Start with the xG, but end with the cold Tuesday.

Eight years later, in September 2026, the same question came back to me through an analytical pipeline. Nine analytical dimensions of an article — tactics and technique, club finance and the transfer market, results and the public-opinion cycle, league landscape and team positioning, rules and governance, management and the dressing room, risk profile, media narrative, and industry transmission — every one of them returned the same phrase: insufficient information.

None of them was empty. All of them were filled. Filled with a zero.

This is where blockchain enters. The problem blockchain promises to solve best is not payments. It is data provenance. And the problem it can least solve is also right here.

Context: information points, Stage 1, Stage 2

Our work runs in two stages. Stage 1 extracts information points from a raw article. Each point is an atomic, verifiable fact: this team took fourteen shots in this match; this deal was worth this much; this coach's contract runs to this year. Stage 2 is the nine-dimension deep analysis built on those points.

The rule is simple. Stage 2 never goes beyond Stage 1. No team, player, match or figure is invented. That is our constitution, and it has not been broken once.

Now back to 2026. From Chattogram we built a standard xG and PPDA model for Abahani Limited Dhaka versus Sheikh Russel KC. We tracked fourteen shots. Abahani's xG was 2.3, Sheikh Russel's 1.7. PPDA was 8.7 against 11.2. The model predicted a 1-1 draw. The match ended 1-1.

That success gave us a habit, and the habit gave us a blind spot. The habit: every report opens with a data table, and no match report goes to press without xG and PPDA. The blind spot: we began assuming that when data exists, decisions get better. We forgot that the more urgent question is what happens when data does not exist.

The blockchain industry has spent years chasing the answer to that question in a different vocabulary. On-chain oracles, data attestation, provenance layers. All of them mean one thing: where did this information come from, who said it, when did they say it, and has the claim changed since.

In our pipeline, none of those four was visible. The nine-dimension analysis finished, and nobody could say whether the raw article had ever entered the system.

The core: four failure modes, one symptom

Nine dimensions returning the same result does not mean nine problems. It means one problem, seen nine times. Working through it, I separated four distinct failure modes, and each has a direct analogue in the blockchain world.

First, ingestion failure. The raw article never arrived. The fetch layer timed out, or the source page was moved, or the crawler was blocked. On-chain, this is oracle downtime. No feed, so no number.

Second, parsing failure. The text arrived, but the extractor could not read it. Encoding broke, structure changed, or the language was one the tokeniser does not recognise. On-chain, this is a malformed payload: the data arrived, but the contract cannot interpret it.

Third, silent pass-through. The text arrived and parsed, but the information-point count was zero. The pipeline did not stop. It treated the zero as valid input and passed it downstream. This is oracle-staleness tolerance set far too wide. The heartbeat has stopped, and the contract still treats the old price as true.

Fourth, and the most dangerous: fabricated output. Had the pipeline proceeded on empty input, teams, players, matches and scorelines would all have been invented. The dashboard would have looked beautiful. No cell would have been blank. And precisely for that reason nobody would have suspected it.

This is where blockchain's most important lesson hides, and it is not about tokenomics. It is about Solidity. A require() statement in a smart contract is not politeness. It means that when conditions are not met, the contract stops itself and reverts. Failure there is a recognised, valuable output.

In data pipelines, failure is usually an exception rather than an output. And an exception means it goes to a log, nobody reads the log, and the pipeline runs on.

In our case the pipeline behaved decently. It did not invent. But it did not stop either. It ran the analysis across all nine dimensions, wrote insufficient information into each, and produced a report whose value is zero.

The effort spent producing that empty report could have caught the empty input at a tenth of the cost. Which brings us to thresholds.

The minimum-information gate: where to place it, how low to set it

A gate needs four conditions. Title and source must be populated. There must be at least one information point. There must be at least one named entity — club, player, coach or competition. And the author's stance and the article's purpose must be identified.

Now the hard question: what is the information-point threshold? One? Three? Ten?

At one, the error risk is high. Building nine dimensions on an article with a single verifiable fact is nine buildings on one brick. At three, short reports get rejected — yet short reports often carry the real signal. At ten, nothing but long features survives, and most of our work is not features.

I settle on three, but I test the sensitivity. Moving the gate between one and five shows what is lost and what is saved. Lower it and recall rises while precision falls. Raise it and the reverse. Which matters more depends on what the downstream task is.

A live match dashboard needs a low gate, because speed matters more there — a wrong xG value gets corrected five minutes later. A contract analysis or a disciplinary report needs a high gate, because a wrong number there circulates for years.

On-chain, that difference becomes sharper. Write something to a chain and it cannot be deleted. So the gate there should be higher, and that is our trade's central trade-off: raising verifiability costs speed, and raising speed costs memory.

The dashboard is not the match; the dashboard is the match's shadow. A shadow can fall in the wrong place, but a shadow has one virtue: it never moves on its own. When a number moves on its own, someone is behind it. The question is who.

The contrarian angle: blockchain does not make bad data true

A comfortable misconception circulates in both football analytics and blockchain. The misconception is that writing data on-chain makes the data true.

It does not. Writing on-chain means only that the data has not changed, and that who wrote it and when can be proven. Provenance and truth are different things. A bad parse committed on-chain becomes permanently, immutably wrong. Garbage in, gospel out.

There is a second problem: cost and latency. In our live xG template we update every fifteen minutes. Anchoring every update on-chain would mean dozens of transactions per match, dozens of fees, and a guaranteed delay each time. In a quiet draw that may be acceptable. In the last three hours of a transfer deadline it is not.

So the threshold should be decision-based, not event-based. Not every number needs anchoring. Only the numbers that change someone's decision — a coach's substitution, a scout's recommendation, a regulator's sanction — deserve an anchor. The rest should circulate and stay correctable.

One more point, less discussed. The empty input that came back may not be a bug at all. The article may genuinely have been empty of information: a headline with nothing underneath. In that case the pipeline's response was exactly right. But one error remains even then — catching the emptiness took the full nine-dimension analysis.

In the right architecture the gate would sit at dimension zero, before analysis begins. Cost: nothing. Saving: the entire analytical budget.

The arithmetic of a near-miss

I am not calling this an accident. I am calling it a near-miss, because the pipeline ultimately did not fabricate. Nobody typed a number. Nobody inserted a club name. The constitution held.

But the arithmetic of the near-miss is interesting. Nine dimensions, at least three analytical conclusions each, each with its own evidence trail and its own confidence level. The result is a long, well-structured, entirely honest report whose informational value is zero and whose architectural value is considerable.

I would rank it among our best documents, because it shows where the system is strong and where it is weak. The strength: it does not invent. The weakness: it stops late.

Football knows this delay well. A team plays badly for fifteen minutes, concedes nothing, so nobody changes anything. Then it concedes in the sixtieth minute, and the change arrives. Data pipeline failures work the same way: until there is visible damage, the system stays live.

And because this is written mid-tournament, one addition. Tournament pressure breaks data discipline first. Pressure makes people abandon process, and the first thing abandoned is verification of information points, followed by the information points themselves. A large share of the numbers printed on the night of a group stage's third round are never checked by anyone.

Signals for the next round

I will track three things next.

First signal: whether the same empty output appears in other records. Once is an incident. Repeatedly is an architectural defect, and architectural defects need engineers, not analyst patience.

Second signal: a hash of the raw text. If a hash of the raw text were stored at ingestion, we could today prove whether the failure was upstream — the article never arrived — or downstream — it arrived and was lost. That single hash would redirect the whole investigation. This is blockchain's real contribution: not price, but memory.

Third signal: the gate's own ledger. How many articles stop at the gate each month, and how many deserved to stop. If most blocked articles were actually acceptable, the gate is set too high. If the zero passes through, there is no gate.

Five minutes of waiting, a fifteen-minute template. Our entire trade sits between those two numbers.

The final question is simple and hard to answer: if we cannot confirm that a sentence ever reached us, what exactly are we verifying on-chain?

A Tuesday

That 2026 match ended 1-1 and our model was proven right. I was proud that day. Now I understand that a correct number and a reliable pipeline are two separate achievements. When both land on the same evening, that is luck, not routine.

A system that can recognise its own emptiness, and announces that emptiness rather than hiding it, is the one that deserves trust. The empty input did that, late but honestly.

Start with the xG. But end with the cold Tuesday, when the dashboard is blank and you have to decide: the number is missing, or the number has not arrived yet.