When the Models Went Silent: On the Limit of Cricket's Numbers, Where Every Forecast Loses Its Bet

A few years ago, on a May night in a Mumbai studio, a number flickered at 34% on the screen in front of me. It was the moment my ball-by-ball model, trained on 812 IPL innings, gave Mumbai Indians only a 34% win probability at the innings break of the final in Hyderabad. The scoreboard read 129/8, and Rising Pune Supergiant's bowlers were in charge. That night Krunal Pandya made 47 off 38, and the match was decided by one run. My model was wrong — no, my model was incomplete. Its death-overs wicket-value term had a flaw I would spend three weeks fixing. But from that night, a habit began: I stopped filing match reports as verdicts, and began every piece with a stated prior, the sample behind it, and the conditions under which it fails.

This article is not the story of one match. It is the story of a process I developed at my desk over eight years — and of the limit where cricket's numbers do not deceive us, but go silent.

Context: When Data Speaks, Does It Speak the Truth?

Cricket analysis today is caught in a strange paradox. On one hand, we have more data than ever — ball tracking for every delivery, coordinates for every shot, fatigue logs for every player. On the other, we have more confusion than ever about what that data means. When I left a Kolkata daily's sports desk in 2026 and joined a twelve-person pricing desk in Salt Lake, my first task was to build a 'true run value' model. At the time I thought the numbers would show me the truth of the match. Today, in 2026, I know that numbers only show what we ask of them — and often we ask the wrong questions.

IPL ball-by-ball data is rich and free. Ranji Trophy data is thin. Regional-language coverage and pre-2026 women's cricket data are nearly absent. This is a bias I call 'data-availability bias' at my desk — where the data that is easily available becomes our subject. This means our best analysis often happens in areas where data is already rich, and our most important questions — like how India's second XI is selected, or how women's cricket is truly valued — remain unanswered.

Core Analysis: The Three Moments When the Models Went Silent

First moment: The 2026 World Cup and the death of 'possession'. In June 2026 I was running the football book. My pre-tournament model, built on two seasons of xG and PPDA from the Bundesliga and Premier League, priced Germany at 14% to fail the group stage. On 27 June Germany lost 0-2 to South Korea and finished bottom of Group F. Three days later Spain completed 1,029 passes against Russia and went out on penalties. I wrote a 2,400-word piece arguing that possession is a lagging indicator and that PPDA measures intent far better. I retired possession percentage from my vocabulary for good and replaced it with field tilt, PPDA and progressive passes per shot. It was a decision that made my tournament previews unfashionably dull and unusually accurate.

Second moment: The 2026 empty stadiums and the 36 all out. In May 2026, when the Bundesliga returned to silent stadiums, I immediately re-ran my home-advantage regression. Across six IPL seasons and 1,200-plus matches it was a 1.07 multiplier. Behind closed doors it collapsed to 1.00. I published the finding before the IPL's September restart, and it held across 60 matches. Empty stands taught me that atmosphere was a variable I had priced at zero. But in December of that same year, in Adelaide, India were bowled out for 36 despite holding a 53-run first-innings lead. My in-play model had given them a 74% chance of not losing. Thirty-six all out in forty-five minutes, and every model I trusted went silent. I spent that winter logging ball-tracking for every wicket. Since then I have added a mandatory 'context coefficient' paragraph to every piece — crowds, travel, calendar, pitch age — and refused to compare pre-2026 and post-2026 numbers in the same sentence.

Third moment: The 2026 Qatar and the Saudi shock. On 22 November 2026 Argentina led Saudi Arabia 1-0 at Lusail and my in-play model had them at 94%. They lost 1-2. I published the post-mortem within forty hours — not an excuse, but a variance ledger measuring the model's calibration error across 5,000 historical matches. The same week, my pre-tournament note on FIFA's added-time directive (27 minutes in England-Iran) had already justified the trip. Every number I trust carries a scar from a season I got wrong.

Contrarian Angle: The Gap Between Correlation and Causation

Here is an uncomfortable truth that every serious student of cricket data knows but few write about: most of our models mistake correlation for causation, and then sell it as 'insight'. For example, there is a strong correlation between high strike rates in the death overs and wins. But is a high strike rate the cause of winning, or does the likelihood of winning allow players to play more aggressively? My 2026 model ignored this distinction, and that is why Krunal Pandya's innings beat it.

Similarly, 'distance covered' and 'high-intensity sprints' are packaged as effort metrics. But pointless running also produces pretty numbers. A player who runs 12 kilometres in the wrong place looks more 'hardworking' than one who runs 9 kilometres in the right place. In 2026 I retired possession percentage, and nobody at the desk noticed. But the truth is that the same flaw exists in many cricket metrics today — like 'economy rate' that ignores wicket-taking ability, or 'average' that ignores match situation.

My next concern is the new women's cricket dataset. Since 2026, the WPL and the 2026 home World Cup win have drawn the commentary establishment's attention quickly, but modelling still lags. India's real problem is not talent but the machinery that ranks it — and the women's game is still an unassimilated dataset in that machinery. Until we develop separate metrics and separate sample windows for women's cricket, all our comparative claims will remain incomplete.

Takeaway: The Signal for the Next Round

So the next time you see a cricket statistic — whether it is a strike rate, an economy rate, or a win probability — ask three questions: What is its sample? What is its window? And under what conditions does it fail? Every number you trust should carry the mark of a moment when it was wrong. If it doesn't, it is not a number — it is decoration. And there is no place for decoration in cricket — especially when the next ball can confront you with a truth no model can see.