Three Numbers That Failed Verification: An Audit Ledger for Cricket Analytics

I spent a domestic match last month sitting in a near-empty stand in Bangalore. In the fifth over the left-arm spinner bowled four straight dots, and on the fourth the batter’s bat barely turned. The young man next to me said, “He’s in form today.” I said nothing, because I did not yet have a single control percentage from that bowler’s last ten matches in front of me.

After play I opened the scorecard and found the same bowler’s economy published as two different numbers by two popular data feeds — 7.42 and 7.83. The gap is 0.41 and it changes no result. The cause of the gap changes rather more: one feed counted byes and leg-byes against the bowler, the other did not. Same name, same match, two truths. The stadium was empty; the numbers were not.

Last week three entries in my own Stage-1 data packet failed cross-check against the public record. The rule is simple: flag the discrepancy before the dimensional analysis, never after. This piece is about those three flags, and why three small mismatches ask a very large question about how cricket analytics is built.

Method note: a number without a stated rule belongs to nobody

Sample size first. Sixteen innings from a domestic tournament and forty from a franchise league do not carry the same weight; the confidence interval on the first is wide enough to support two contradictory conclusions, both of which will look reasonable.

Then definitions. By control percentage I mean the share of a batter’s shots that went where the bowler or fielder controlled them — miss, edge, block, or straight to hand. Dot-ball percentage is a different object: legal deliveries on which no run was scored. Collapse the two and you lose the distinction between building pressure and merely stopping runs.

False-shot percentage measures how often a batter reacted to the wrong line or length. Catch efficiency measures catches taken against chances offered. Role-adjusted strike rate measures the same batter in the powerplay, the middle and the death — because the same 130 means excellent at the death and ordinary up top.

I write these six definitions down for a reason. The three flags below were born inside them.

There is also an absence, which no flag caught because it is not an entry. The scorecard counts overs, not the balls inside them. Which delivery a batter left on purpose and which one he swung through and missed does not survive into the summary sheet. I count the silence between the deliveries, because the scorecard is a lossy compression of the match.

Flag one: where did the denominator go

The first entry was a death bowler’s economy. One feed showed 8.90, the other 8.24. Chasing the source, I found the two feeds count legal deliveries differently — wides and no-balls are absorbed into the over in one place and excluded in the other.

It sounds trivial. Economy is a fraction: runs over overs. Change the denominator and the fraction moves, and with it the auction price, the selection committee’s impression, the broadcaster’s line. Every published rating without a stated definition is an unfinished sentence.

The second problem is arithmetic. This bowler’s death sample is eleven innings, 28.4 overs. At that size one bad over moves the economy by up to 0.3. The variance the two feeds are arguing about is smaller than the variance luck produces on its own.

Flag two: who chose the window

The second entry was an opener’s strike rate — “146 in his last twenty innings.” The number is true. The window is suspect. Nine of those twenty were small chases, where the cost of risk is low and the bowling is less aggressive.

Change the window and the picture changes. Over the last thirty innings the figure falls to 138; in the powerplay alone it is 119. All three numbers are true. The window that knows the outcome does not know the information.

My own rule: fix the window, then look at the number. The order is painful, because the first window is often the weak one and admitting it costs time. Reverse the order and analysis stops being analysis and becomes marketing.

Flag three: the fraction nobody counts the bottom of

The third entry was a fielding metric — catch efficiency. The problem is simple and deep: the numerator is countable, the denominator is not. Catches taken appear in the scorecard; catches dropped do not, because a “chance” is not official statistics anywhere.

Catch efficiency is therefore not a rate but a floor. The fielder who positions well gets more chances, drops more of them, and reads worse. A metric with an unknown denominator belongs to the audience, not the player.

My fix is partial: in every fielding innings I count chances myself, from video, under one definition. It is slow, it is subjective, and without it every other number is decoration.

One player, two markets, two prices

Put the three flags together and a market appears. Take a death bowler. His price on a Kolkata auction floor is set on one set of numbers, his price in Dhaka on another. The first looks at economy, the second at wickets; one denominator is clean, the other is not.

The same man sells for two prices in two cities, and both prices look defensible — inside separate definitions. An international market clears the price gap; it does not clear the definition gap.

This is where role-adjusted work earns its place. Not the raw economy, but the role: powerplay bowler, middle-overs pressure bowler, or death specialist. Skip the role and the popular story about auction price and performance survives without ever being tested.

When I interviewed Soumya Sarkar in 2026, I first understood that one innings returns to two journalists as two stories. My notebook then held adjectives, not numbers. Adjectives build markets. They do not build models.

Where my own scepticism stops me

The easy explanation is that one feed is wrong. Usually both are right and answering different questions. Failing to see that turns every discrepancy into a conspiracy — and that is my own largest risk.

None of the three flags is the product of conspiracy. The data pipeline has three stages: scorer, entry, publication. Each stage makes a decision, and every decision is a definition. Errors happen when nobody writes the definition down.

So I have imposed a rule on myself: at most two custom roles per ten analyses, defined before outcomes are seen. The more roles I invent, the more players look undervalued — and the real discovery stops being the player and becomes my own bias.

I also keep a base rate for my overrides. Over the last two seasons I have gone against consensus on numbers seven times; three were later proven wrong. Publishing that rate is uncomfortable, but without it my scepticism is indistinguishable from any fan’s opinion.

What I will watch in the next three matches

I am writing three thresholds down now. If the left-arm spinner’s control percentage falls below 65 over sixty balls, “in form” does not enter my ledger. If the opener’s powerplay strike rate sits below 125, his overall 146 is a property of the window, not the player. And if my own chance count over ten innings comes in under eight, I publish no conclusion about catch efficiency at all.

Three thresholds, three traps. I will grade them myself afterwards, win or lose.

Let the ledger breathe before the narrative does.