Nine Columns, Nine ‘Not Applicable’: In Football Data, the Hardest Skill Is Knowing When to Stop

2:40 a.m. Nine columns on the screen, and under every one the same sentence: ‘insufficient information, assessment not possible.’ The title field reads ‘not applicable,’ the source field reads ‘not applicable,’ and the information-point list is entirely empty. Twenty minutes earlier the file had arrived labelled ‘analysis report.’ What I am holding is not analysis. It is a pipeline’s death certificate.

The easy route was to fill the blanks. One club, one formation, two xG numbers — no reader would have noticed. The first lesson of a data monk is this: an empty cell is also information. And the information it carries is that we know nothing about this subject.

How the pipeline works needs stating plainly. The first stage breaks an article into ‘information points’ — each point a verifiable claim, tagged with source and date. The second stage sorts those points across nine dimensions: tactics and technique, club finance and transfers, results and public opinion, league context, governance, management and dressing room, risk, media narrative, and industry transmission. One rule governs everything — every conclusion must be traceable back to a numbered information point.

Tournament cycles put that rule under maximum strain. During a major tournament, a dozen transfer tickers, two-line social posts and clips arrive every hour. For transfer rumours the first filter is source tier — who is writing it, who is leaking it, where the agent’s interest sits. If the source field itself is blank, every downstream layer is worthless.

Technically, three distinct diseases are hiding here, and each needs a different cure. Retrieval can fail — paywall, 404, anti-scraping block, or a non-text source such as video or podcast. The article can arrive but the deconstruction step returns an empty parse — model error or timeout. Or the article is genuinely too thin to yield three separate claims — a two-sentence transfer ticker. The first needs a substitute source, the second a re-extraction, the third a short-form note.

The core argument sits right here. The most valuable output of an analysis pipeline is not an accurate forecast — it is a correctly declared zero. A wrong xG figure gets caught; a coach, an analyst, even a fan watching the match can say ‘that was never a chance.’ A fabricated tactical narrative does not get caught. It gets quoted, printed, and still circulating a year later. A wrong number lives for hours. A wrong narrative lives for a season.

This is where the audit-trail idea earns its keep. Each information point can be hashed together with its source, date and quoted fragment, and written to an immutable log. Then the principle ‘no point, no claim’ stops depending on anyone’s honesty — it becomes machine-checkable. No public blockchain is required; a tamper-evident record is enough, one that stops the next stage from quietly filling a cell.

The xG and PPDA model I built in Chattogram in 2026 for the Abahani–Sheikh Russel fixture rested on the same principle. Fourteen shots, 2.3 xG for Abahani, 1.7 for Sheikh Russel, PPDA 8.7 against 11.2. The model called a 1-1 draw; the match finished 1-1. The numbers matched because every number sat behind a timestamped entry, and because every reporter was compelled to file a post-match data sheet. I did not introduce the sheet because the model was right; the model was verifiable because the sheet existed.

The second lesson came while running a live dashboard during the Croatia–England semi-final at the Russia World Cup. Croatia 1.4 xG, England 0.8; Luka Modrić covered 12.8 kilometres and completed 67 passes, and his late pressing dragged England’s PPDA down to 12.9, the match ending 2-1. That night the most important cell on the dashboard was latency — how many seconds late the number arrived. Without the delay known, a number is not information; it is only light. The dashboard is not the match, it is the match’s map; and a blank patch on a map cannot be filled by guesswork.

One line in the report was, to my mind, the most valuable in the whole document: ‘the absence of identifiable risk is not evidence of the absence of risk.’ Six rows in the risk matrix, every one marked ‘not identifiable.’ The same holds in football — a club without data does not lack risk; lacking data is itself the largest risk.

So where does the threshold sit? My proposal: a minimum of three information points, plus at least one named entity (club, player, coach or competition) and a dated source. Below three, this is not analysis but a short note. That is a threshold, not a law, and both costs must be shown. Lower the gate to three and thin tickers slip through, flooding the chain with weak reports; raise it to five and legitimate short items get rejected. I pick three, because the errors are not symmetrical — a missed item can be recovered later, a fabricated analysis cannot be withdrawn.

One structural flaw also surfaced. The instruction for entity extraction says ‘identify from the information points above’ — with an empty list, that is a circular trap. The fix is simple: let entity extraction read the raw article text, so entities survive even when point extraction fails.

Now the other side. Null-handling can itself become a hiding place. A pipeline that says ‘insufficient information’ every time never faces accountability, is never proven wrong, and slowly becomes useless. We give awards to the model that called 1-1 correctly; nobody remembers the model that said ‘not yet, the input is incomplete.’ The second demands more courage, because it has nothing to show.

The real failure is not the empty report. The real failure is that at the second-stage boundary, three different faults cannot be told apart. That is an observability failure, not a data failure. Three different diseases, three different cures, and no diagnostic label anywhere.

What to watch in the next round. Whether a FAILED_EXTRACTION flag ever appears in the article log. Whether source, author, date and URL become mandatory fields. And whether a re-extraction returns five information points from the same source. If a report cannot show its own data’s birth certificate, the question is simple — is it a report, or a guess? Start with the xG, but end with the cold Tuesday, the one when leaving the cell empty turned out to be the best analysis on the desk.