How cricket teams use data: analytics behind the game

Somebody sits in a room and watches every delivery twice. Once at full speed, once frame by frame, and then they type a row: who bowled, who faced, where it pitched, what line it held, which shot was attempted, whether the batter meant it, where the ball went, what it cost. That row is the whole of cricket analytics. Every matchup graphic on television and every projection on a coach's tablet is a pile of those rows, grouped and counted.

So the honest way to ask how cricket teams use data is to ask about the row. What is in it. Who put it there. How many of them you have once you have filtered down to the question you actually care about. And which of the numbers in it were measured rather than produced by a model. Most coverage of the subject skips all four questions and starts at the output, which is why the output so often sounds more certain than it has any right to.

Analysis in cricket is not primarily about prediction. It is about reducing the number of things a coach or a captain has to guess.

Where cricket teams get the data in the first place

Three separate pipes feed a modern analysis department. They are not equally trustworthy, and the difference between them is the single most useful thing a spectator can learn about the subject.

The scorer's record is a human being logging every ball as a structured fact. Match, innings, over and ball number, striker, non-striker, bowler, runs off the bat, extras with their type, wicket with its method and the fielder involved. Nothing in that list is a judgement. Either the ball was a wide or it was not, and the Laws decide which. This is the layer that has existed in one form or another for as long as anyone has kept a scorebook, and it is the layer that public archives distribute.

Coded video is a human being making judgements against a written taxonomy. A coder watches the same delivery back and tags it: which length band it pitched in, which line relative to off stump, which shot was played, whether the batter controlled it, which foot he moved onto, which fielding region the ball travelled into. None of that is in the scorer's record and all of it is opinion, disciplined by a definition document rather than by physics.

The taxonomy is where good departments separate themselves. Access to data is broadly similar across the top teams; the discipline of applying a definition identically across two hundred matches and four coders is not. If two coders disagree about where a good length ends, every conclusion drawn downstream is noise wearing a decimal point.

Tracking is the machine layer. Several high-frame-rate cameras, calibrated against a surveyed model of the ground, find the ball in each frame and fit a trajectory through the points. Out of that comes release point, speed, the bounce point on the pitch, deviation off the surface, movement through the air, and the height and position at which the ball reached the batter. Edge detection adds a synchronised sound and vibration trace. Increasingly there is player tracking too, giving fielder positions and batter movement.

The difference between a measured number and a modelled one

This is the distinction that almost no article about cricket data explains, and everything downstream depends on it.

The cameras can only fit a trajectory through positions where they actually saw the ball. Everything on that path is measurement, subject to the ordinary limits of a camera system: resolution, frame rate, calibration, occlusion by a body or a bat.

Then the ball hits the pad, and the measured part of the story stops.

On an lbw review, the segment the decision turns on is the segment that never happened. The ball did not continue to the stumps; it hit a leg. What the system shows after the point of interception is a projection. It takes the ball's fitted velocity and the deviation it was undergoing at the moment of impact and extrapolates forward under a physical model of how a cricket ball behaves in flight. That projection is not a photograph of anything. It is a prediction, rendered in the same confident colour as the part that was observed.

Ball-tracking predicts as well as it measures, and the two halves of its output are printed identically.

The uncertainty in the projected part is not uniform either. It grows with how far the ball still had to travel, so a batter struck a long way down the pitch has a longer extrapolation between him and the stumps than one struck on the crease. It also grows when the ball is intercepted very soon after pitching, because the system then has only a short post-bounce segment to estimate the deviation from, and the deviation is precisely the quantity that decides where the ball was going. Those are the two situations in which a projected path is at its shakiest, and both are common.

The review protocol handles this by building a tolerance around the projection rather than treating it as fact. That is the entire logic of an on-field decision standing when the projected impact is marginal: the system is being asked how confident it is, not merely where the ball went. Anyone who understands that the projected segment is modelled understands why the margin exists, and stops being annoyed by it.

The same measured-or-modelled question applies to everything else on the screen. Speed at release is measured. Deviation off the pitch is derived, being the difference between two fitted segments of one trajectory. Control percentage is coded judgement. Expected runs and win probability are models trained on other people's rows. All four get displayed in the same typeface.

29%29%29%14%

• Logged as fact by the scorer6
• Coded by human judgement6
• Measured by the tracking system6
• Produced by a model3

A count of the fields listed in this section. Facts are logged by the scorer, judgements are coded from video against a taxonomy, tracking fields are fitted from camera positions, and model fields are outputs of something trained on other deliveries.

Show the numbers

What is in one delivery record, sorted by how it was obtained

Item	Value
Logged as fact by the scorer	6
Coded by human judgement	6
Measured by the tracking system	6
Produced by a model	3

The ball-by-ball row is the atomic unit of cricket analysis

Strip away the presentation and the archive is one very long table with one row per delivery. The public version of that table, the kind distributed in open ball-by-ball archives, holds roughly this per row: the match and innings it belongs to, the over and ball number, the striker, the non-striker, the bowler, the runs scored off the bat, the extras with their type, and if a wicket fell, the player out, the method of dismissal and the fielder credited.

That is it. Nine or ten fields, most of them names.

From that alone you can rebuild every number on a scorecard, because every conventional cricket statistic is a group-by over that table and nothing more. Batting average is the sum of runs off the bat for one striker divided by the count of rows in which he was dismissed. Strike rate is the same numerator over the count of legal deliveries he faced, times one hundred. Economy is the runs charged to a bowler over the overs he sent down. The full set of definitions and the traps inside them matters more than most people assume, because the traps live in the denominators: which extras are charged to the bowler, which deliveries count as faced, what happens to a not-out innings.

That is also why the argument between average and strike rate is not a matter of taste. The two are different group-bys over identical rows, one dividing by dismissals and one by balls, and they answer different questions about the same innings. The same is true of the three bowling numbers, which are three denominators applied to one numerator.

The team version of the row adds the coded and tracked fields, and it is perhaps three times as wide. Length band, line, shot played, control, foot movement, the fielding region the ball travelled to, release point, speed, bounce coordinates, deviation, movement through the air, impact position. That extra width is the entire competitive advantage, and it is why public analysis and team analysis are not really the same activity.

Two unglamorous properties of the table decide whether any of it works.

The first is identity. The same human being appears in different sources as three different strings, with initials, without them, transliterated two ways, sometimes with a suffix because two players share a surname. Joining a domestic record to an international one requires a register that assigns each player a stable identifier and maps every spelling onto it. Without that, a career splits silently into fragments and the aggregate is wrong in a way nobody notices.

The second is that errors in the row are invisible in the aggregate. A dismissal credited to the wrong fielder, a bowler recorded for someone else's over, a boundary logged as four when five were signalled: each propagates into every figure computed from it, and none of them announces itself. Analysis departments spend far more time on data integrity than any outside account suggests, and the ones that skip it produce fluent conclusions from a corrupted table.

1. 1The ball is bowledA left-arm spinner bowls over the wicket to a left-handed batter in the eleventh over. It pitches on middle and turns away towards slip. The batter pushes at it, gets an inside edge onto the pad and survives.
2. 2The scorer logs the factsOver eleven, ball three, this bowler to this striker, no runs off the bat, no extras, no wicket. Nothing here is an opinion, and this is the only part that reaches a public archive.
3. 3Tracking fits the flightThe cameras find the ball in each frame and fit a trajectory: release point, speed, where it pitched, how much it deviated off the surface, how high it was when it reached the bat. Every one of these is derived from observed positions.
4. 4A coder adds the judgementsGood length, middle stump line, defensive push, not controlled. Four fields no machine produced, entered against a written taxonomy so that they mean the same thing as the same four fields entered last season.
5. 5The row is joined and storedPlayer identifiers, venue, innings phase and match conditions are attached, so this delivery can later be pooled with every other delivery of its kind rather than sitting alone.
6. 6The row is pooled into a questionOn its own it says nothing. Grouped with several thousand deliveries of left-arm spin turning away from left-handers in the middle overs, it becomes a false shot rate, a scoring map and a length band worth attacking.
7. 7The finding is reduced to an instructionNot a report. One sentence and one clip for the bowler: pitch it on middle and let it turn, keep a slip and a short cover, and expect the push. If it cannot be said in a sentence a player will accept, it will not be used.

The sequence a single ball goes through in a professional set-up. Steps one to five happen within seconds. Steps six and seven happen at an interval or between matches, because live coaching from the dressing room is restricted in most formats.

Matchups, done honestly, and the sample size nobody puts on screen

Matchup analysis is the most discussed and the most misused part of the discipline, and the textbook case is worth working through properly, because it shows both the real mechanism and the real problem.

A left-arm orthodox spinner bowling over the wicket to a left-handed batter is the classic example. The mechanism is geometry. The angle of release brings the ball in towards the batter and the turn takes it away from him, towards the slip cordon. That combination is difficult for the same reason off-spin to a right-hander is difficult: the eye is being pulled one way by the angle while the ball goes the other. The edge, when it comes, goes to the keeper or first slip, and the batter cannot simply cover the line by playing across, because the ball that grips less will hit off stump.

Reverse the handedness and the whole thing inverts. The same bowler to a right-hander is turning the ball into the pads, which is a different problem, defended differently, with a different field.

That is a real mechanism. It exists whether or not anyone has data on it, and a captain who understands it will bring a left-arm spinner on to a left-hander for entirely sound reasons.

Now the honest part.

The number usually displayed alongside that decision is one batter's record against one bowler. Those two players have very likely faced each other for a few dozen deliveries across their careers, and often far fewer. A sample that size is compatible with almost any claim you would like to make from it. A batter dismissed twice by a particular bowler in one series is not thereby vulnerable to him; he is a person to whom two things happened.

Watch what filtering does to the data you have.

All deliveries faced in the format10000

Against left-arm orthodox spin1400

And in the middle overs620

And on turning surfaces180

And against this specific bowler30

A constructed illustration with round invented numbers, chosen to show the shape of the collapse rather than to describe any real player. Each filter is defensible on its own. The point is that they multiply.

Show the numbers

What is left of a batter's record after four filters

Item	Value
All deliveries faced in the format	10000
Against left-arm orthodox spin	1400
And in the middle overs	620
And on turning surfaces	180
And against this specific bowler	30

Every one of those filters is reasonable. Handedness matters, phase matters, the surface matters enormously, and the individual bowler is the thing the captain is actually choosing. Apply them one after another and a career becomes thirty balls, which is five overs, which is a figure that would move by a wide margin if one shot had found a fielder instead of a gap.

Four separate problems then sit on top of the sample.

Selection. The matchup exists because a captain engineered it, usually in circumstances that already favoured it. Observational data of this kind records what teams already believed as much as what is true, and the belief and the evidence are then presented as though they were independent of each other.

Truncation. Batting sequences do not end at random. They end when the batter is dismissed. Long runs of deliveries against a given bowler are therefore drawn disproportionately from occasions when things went well for the batter, and the short runs are the failures. The rows are not independent draws from a fixed distribution, and treating them as though they were flatters the batter in exactly the cases where you most want the truth.

Non-stationarity. A record of how a batter played the short ball three years ago describes a player who no longer exists. Technique, fitness, role and confidence all move. Career aggregates assume a stability that is not there.

Conditions. The same delivery is a different delivery on a seaming surface, at altitude, under lights, with a wet ball, on a pitch prepared to turn from the first session. Pooling across conditions hides the effect; splitting by conditions runs straight back into the sample problem.

The version of matchup analysis that survives all this aggregates by type rather than by name. Instead of asking how this batter goes against one left-arm orthodox spinner, ask how he goes against left-arm orthodox spin in general: how quickly he scores, where he scores, how often he plays a false shot, whether he uses his feet, whether he sweeps, which length band he attacks. That gives a usable sample and a transferable conclusion, because bowlers of a type share the essential problem they set.

The cost of that fix is specificity, and it is a real cost. Two left-arm spinners are not the same bowler. One tosses it up and turns it a long way, one fires it flat at the stumps and relies on the arm ball, and pooling them averages two different questions. Buying sample means spending precision, always, and no setting of that dial is correct in general.

What a matchup number is genuinely good for is generating a hypothesis. It should send you to the video, not to the team sheet. A captain who takes it as an instruction ends up bowling an off-spinner in the last over because the numbers said so, having ignored the fact that the off-spinner is nineteen and terrified.

Phase analysis survives what matchups do not

Phase analysis is the least fashionable part of cricket analytics and by a distance the most dependable, for a reason that is structural rather than clever.

The phases are defined by the playing conditions. They are not a modelling choice, they apply identically to every team in every match, and every ball in the archive is therefore already labelled by phase at no cost whatsoever. A batter's powerplay record runs to thousands of deliveries, not thirty. The cells are large because the definition is external to the analyst.

• 120Legal deliveries in a full innings
• 6Overs in the powerplay
• 2Fielders permitted outside the circle in the powerplay
• 5Fielders permitted outside the circle thereafter

Structural counts from the standard playing conditions for a full twenty-over innings, before any additional restriction imposed for a slow over rate.

Those four numbers set the economics of the entire innings.

In the powerplay, with only two fielders permitted outside the ring, boundaries are relatively cheap and wickets are relatively expensive, because the batting side has the most balls left in which to lose the game. In the middle overs the ring opens up, singles become easy and boundaries become hard to find, so the contest turns on whether the fielding side can force the batters into risk to beat the ring. At the death the value of a wicket falls sharply, because few overs remain in which an incoming batter could hurt anyone, and the value of a dot ball climbs to match.

Fifty-over cricket has the same structure at a different scale, with a further restriction band through the middle of the innings, and the effect is identical: what a run is worth and what a wicket is worth both change on a schedule everyone can read in advance.

Out of this comes the most usable product of the whole discipline, which is a set of phase benchmarks for a ground and a match type. What a competitive score looks like at the end of the powerplay here. What a chase needs at halfway. How much a wicket costs in each phase at this venue. These have to be ground-specific, because grounds differ enormously, and the same total means something completely different on a small square with true bounce and on a slow surface with long straight boundaries.

The practical output is not a target score. It is a set of thresholds that tell a captain when the game state has changed. A side behind the benchmark with wickets in hand is in a materially different position from a side level on runs having lost its top order, and quantifying that is genuinely useful, because human judgement under pressure systematically overweights the last two overs.

Phase thinking also drives bowling allocation, which is where it earns its keep. If a bowler is at his best in the powerplay and his worst at the death, and the captain has been holding him back out of habit, the split makes that visible in a way a season of watching does not, because the eye does not aggregate and the memory is dominated by the last time something went badly.

The honest limits are worth stating. A phase is not homogeneous either; the first over of a powerplay and the sixth are different overs against a different ball. Ground benchmarks are themselves samples, and a venue that has staged a handful of matches gives you a handful of matches. Grounds change too, when a square is relaid, a drop-in surface is used, a rope is brought in or a new stand alters the wind. A benchmark is a statement about a ground's past, and grounds move.

How cricket teams use data to build a bowling plan

A pitch map plots where the ball landed. Two axes, both simple: distance down the pitch, and line relative to the stumps. Every professional bowler has seen his own, many times.

The first thing worth knowing is that a length band is not a fixed distance. Good length is defined relative to the batter and the surface, not by a mark on the ground. On a pitch with steep bounce the good length moves back, because a ball that would be driveable on a low surface arrives at chest height. On a slow, low pitch it moves forward. The band that is unplayable at one venue is a half-volley at another, which is why a pitch map read without knowing what kind of surface produced it is close to meaningless.

A map on its own tells you density: where this bowler puts the ball, and how tightly. That is useful for coaching consistency and useless for planning, because it says nothing about outcomes. Overlay the runs conceded and the false shots induced and the map starts to speak. Most bowlers have one band where they are close to unplayable and a neighbouring band where they are ordinary, and the distance between the two is frequently a matter of inches. That is a coaching conversation, and it is one video alone rarely produces, because the human eye cannot aggregate two hundred deliveries into a density.

From the batting side, the same map is read in reverse. Which lengths does this batter attack, which does he defend, and does that change when the ball is moving. Where is his scoring concentrated when the ball is full and straight, and where when it is short of a length.

Turning any of that into a plan requires three things that are not data at all.

A target box, small enough to be specific and large enough to be achievable, because a plan the bowler cannot execute is worse than no plan.

A cue the bowler can feel rather than think about. Nobody has ever bowled a better length by concentrating on a coordinate. They bowl it by aiming at a spot, or by a change of wrist position, or by a thought about the front arm, and converting a finding into a cue is a coaching skill rather than an analytical one.

A fallback, because the plan will fail at some point in the spell, and the decision about what to do next is far better made in the morning than in the seventeenth over.

Two derived measures do most of the work in monitoring whether a plan is working, and both are process measures rather than outcome measures. Control percentage counts the proportion of deliveries a batter met as he intended, judged by a coder, and it is a better short-term guide to how a player is batting than runs are, because it depends far less on luck. False shot rate is the same idea from the bowler's side, counting how often he induced an error. A bowler with a high false shot rate and poor figures has usually been unlucky. A bowler with good figures and a low false shot rate has usually been lucky, and will be found out.

Both rest entirely on a coder's judgement about what counts as control. That is the caveat, and it is the reason these numbers travel badly between organisations that code them differently.

How cricket teams use data to set a field

A wagon wheel plots where a batter scored, usually as lines radiating from the middle of the ground. It is the most reproduced figure in cricket and one of the most misleading, for two reasons.

The first is that it is a projection which throws information away. It shows direction and sometimes whether the ball reached the rope. It does not show how hard the ball was struck, whether it went along the ground or over the ring, or which fielder it beat by how much.

The second is more serious. A wagon wheel is entirely conditional on the bowling the batter received. A player who scores heavily square on the off side may be an outstanding cutter, or may simply have been fed width all season. The map cannot tell those apart, and a field set from it will be a field set against a bowling plan that no longer exists.

The pitch map is what rescues it. Filter the wagon wheel by length and by line and it starts to say something usable: this batter scores square of the wicket against anything short of a length, and almost nothing when the ball is full and straight. Now there is a plan and a field that belong together.

Field setting is the point at which analysis becomes a physical thing standing on grass, and the arithmetic underneath it is a trade rather than an optimisation. A fielder moved to one place is a fielder taken from another. There are nine of them and rather more than nine places worth covering, so every field concedes something, and the work is about which concession is least bad against this batter under this plan.

Three principles come out of it repeatedly.

The field must match the bowling plan. A field set to protect the leg side is worthless if the bowler misses his length and offers width, and most fields that look stupid on television are fields that were sensible until the bowler stopped executing. That is the most common failure of all, and it is not an analytical failure.

Boundary riders should be placed by scoring density, not by convention. The traditional straight boundary rider is often standing where very few balls actually go, and moving him ten yards squarer changes the arithmetic of an over more than any change of bowler does. The conventional names and positions are a vocabulary, not a prescription, and the vocabulary predates the data by roughly a century.

Catching positions are a probability question. A slip is worth having when the chance of an edge carrying, multiplied by the value of a wicket in this phase, exceeds the runs that fielder would otherwise have saved in the ring. That calculation moves with the age of the ball, the pace of the surface, the batter and the phase, and it is one of the few places where a good analyst straightforwardly improves a captain's default setting.

The limit here is the one that runs through the whole discipline. Description is not prescription. Knowing where a batter scores does not tell you how to stop him, because the region you push him into may be one he is better in, and because he adapts. Every field change is an experiment run on a system that responds to being experimented on.

Expected runs and win probability, and what they were trained on

Expected runs is the idea that has changed cricket analysis most in recent years, and the principle is simple. For a given delivery, described by its length, line, speed, movement, the phase of the innings, the venue and the match situation, ask what happened historically across a very large number of similar deliveries. The average of those outcomes is the expected value of this ball. Aggregate over an innings and you have an expected score. Aggregate over a spell and you have what that spell should have cost.

The value of this is that it separates process from result, and cricket needs that more than most sports, because its outcome variance is enormous. A bowler can put six deliveries in exactly the right area and be hit for four boundaries by a batter who took risks that came off. Under conventional figures he had a bad over. Under expected runs he bowled a good over and got a bad result, and the sensible response is to give him the ball again.

The same logic applies to batting, and it is markedly less popular. A batter who plays a run of high-risk shots that come off has scored more than the expected value of what he did, and the model says so, quietly, before the coach picks him again on the strength of one innings.

The construction is not unique to cricket. It is the same move as golf's expected strokes baseline, where the value of a shot is the difference between two averages minus the shot itself, and the same as run expectancy in baseball. State in, average outcome out, credit assigned by the change in the average. What differs between sports is only how well the state can be described.

Win probability is that machinery applied to the match rather than the ball. The state is runs scored, wickets down, balls remaining, the target if there is one, and usually the venue. The model finds completed matches that reached a comparable state and reports the proportion that were won from there.

Here is the part that gets misunderstood.

The same nominal score gives different win probabilities at different grounds, and it should. A model that conditions on the venue is fitted to a different sub-population at each one, so 140 for 4 at the halfway stage of a chase is a genuinely different proposition at a small square with a fast outfield and at a slow ground with long straight boundaries. That is the model working correctly.

The cost of conditioning is sample. Every variable added to the state vector cuts the number of historical matches that match it, and there is a point at which the model is averaging over so few real comparisons that its confidence is manufactured. A model that does not condition on venue avoids that problem by giving one answer everywhere, which is wrong everywhere, just in a different direction at each ground. Specificity or sample. Not both, and the choice is made by whoever built the model, usually without saying so on screen.

Three further things a win probability model does not know.

It does not know the conditions changed mid-match. Dew arriving in the second innings makes the ball difficult to grip and the outfield quicker, so a chase that looked hard at the toss becomes materially easier, and a model trained on a pooled archive prices the position as though the ball were dry.

It does not know about the personnel. It cannot see that one side's best bowler has a hamstring problem and has two overs left that he will get through at three-quarter pace.

And it is anchored to an era. Scoring rates in the shortest format have moved substantially over the life of any long archive, so a model fitted across all of it is partly describing a game no longer being played. Restricting the training set to recent seasons fixes the anchoring and shrinks the sample, which is the same trade as before arriving in a new costume.

A win probability is a statement about a population of past matches that resembled this one. It is not a forecast about this one, and the number on the screen never says which it is.

Auctions, recruitment, and the role-slot problem

Recruitment is where analytics meets money, and it is a harder problem than match analysis because the question changes shape.

Match analysis asks what happened. Recruitment asks what would happen if this player, whose record was generated somewhere else, occupied a specific slot in your side, in your conditions, alongside these ten other people. That is a projection into a counterfactual, and the counterfactual is where all the difficulty lives.

The core difficulty is the role-slot problem. You are not buying the best available players. You are filling a small number of slots subject to hard constraints, and the constraints interact. In a franchise competition there is a cap on overseas players in the XI, a fixed purse, a squad size, a need for a wicketkeeper, a need for enough bowling options to cover one expensive spell, and usually a preference for a mix of left and right-handers so that the fielding side has to keep changing its field. The auction mechanics then add a live constraint on top, because the price is discovered in the room rather than set in advance.

The consequence is that value is not a property of a player. It is a property of a player, a slot and a squad taken together. The same batter is worth a great deal to a side with no top-order left-hander and much less to a side that already has two, and no model producing a single number per player can express that.

Underneath sit three specific traps.

Records are generated in roles. An opener's numbers are produced against a hard new ball, with the fielding restrictions at their tightest and a full complement of overs ahead. Those numbers do not transfer to number five, where the ball is older, the field is spread and the job is to score quickly against spin. A player evaluated on his aggregate record is being evaluated in a role he may never occupy for you.

The purse is shared. Overpaying for the second slot raises the effective price of the last one, because whatever is left has to cover everything still missing. The right question about a bid is never whether the player is worth it in isolation. It is what the squad looks like at the end if the money goes here rather than somewhere else, which means the analyst's real output is a walk-away price rather than a valuation.

Transfer across competitions is where the models are weakest. A player's numbers were produced against a particular distribution of bowling, on particular surfaces, under particular conditions. Move him to a competition played on slower pitches against more spin and the distribution he faces changes, and there is no reliable general method for translating a record across that gap. Anyone who tells you there is has usually fitted the translation to the players who succeeded, which is the survivors talking.

Sample bites here too, and hard. A middle-order batter in a short-format league may face only a few hundred deliveries in a season. A strike rate computed from that is an extremely noisy estimate, and the difference between two players that a franchise pays a large premium to secure is frequently smaller than the noise in the numbers used to justify the premium.

Workload and injury data, where the biggest decisions are actually made

The most consequential analytical work in cricket has nothing to do with matchups, and it receives a fraction of the coverage.

Fast bowling breaks people. The action loads the lower back and the lower limbs in a way that accumulates, the accumulation is largely invisible until something tears, and the players it destroys are usually the ones a side most needs. Every serious set-up now runs a workload programme, and that programme is data-driven in a far more literal sense than field placement ever is.

What gets collected: bowling counts, in matches and crucially in training; movement and acceleration data from wearable units; session duration and intensity; self-reported wellness and perceived exertion; and over longer horizons, screening and medical data. Some of that is measured by a machine and some of it is a number a tired player types into a phone at nine in the evening, and the difference between those two matters in exactly the way described at the top of this article.

The counting problem here is easy to miss and important. The scorer's archive counts match deliveries only. A fast bowler's match overs are a minority of the balls he sends down in a week once nets, warm-ups and rehabilitation work are included. Any workload analysis built from public ball-by-ball data is therefore working from a systematic undercount, and the undercount is not even proportional, because the players who bowl the fewest match overs are often the ones bowling the most in training trying to get back into the side.

The standard framework compares recent load against habitual load. The idea is that the body adapts to what it is used to, so a sharp increase relative to a player's own baseline is riskier than a high absolute load he has been building towards for months. The thresholds people attach to that comparison are contested, and the research literature is a good deal more equivocal than the confident presentation in team meetings suggests, so it is worth saying plainly that this describes the shape of the idea rather than endorsing any particular ratio.

What the data actually decides is availability. Rest weeks. Rotation across formats. Which tournaments a bowler skips entirely. Whether a young quick has a hard cap on overs per spell and per week. How a bowler is reintroduced after injury, at what volume, over how many weeks. Those are selection decisions taken on the basis of numbers, made weeks in advance, and they shape more results than any in-match analysis ever will, because a side's best bowler being fit in March is worth more than every field placement in the tournament.

The limits are severe and honest practitioners say so.

Injuries are rare events. A squad produces very few per season, and a model trying to predict a rare outcome from many candidate variables will find patterns in the training data that do not repeat. There is a real risk of building something that looks impressive in retrospect and forecasts nothing.

Causation is tangled. Load is one input among technique, age, growth, previous injury, travel, sleep, playing surface and plain accident, and separating those from observational data collected on a few dozen people is not a solved problem in any sport.

And this is data about a person's body, gathered by their employer, used in conversations about selection and contracts. The governance around that is genuinely difficult, and it is a live issue across professional sport, cricket included.

Where cricket analysis is genuinely weak

Being specific about the holes is more useful than another paragraph about potential.

Fielding is barely measured at all. There is no consistently maintained, publicly comparable record of ground fielding: balls fielded cleanly, misfields, runs saved or conceded relative to an average fielder in that position, direct hits attempted and missed. Catches and drops are recorded unevenly and the definitions vary between sources. Worse, a fielder's apparent value depends heavily on where the captain told him to stand, so raw counts confound the fielder with the field he was placed in. Cricket has nothing comparable to the tracking-derived fielding measures other sports have built, and the gap is one of data collection rather than of ideas.

Pressure is believed in and not measured. Every player will tell you that a run of dot balls creates something that produces a wicket two overs later. The problem is that pressure and difficulty are the same variable in the data. A sequence of dots occurs because the bowling was good or the conditions were hard, and the wicket that follows is attributable to the same cause. Isolating a pressure effect means separating two things that always arrive together, and the sample available for doing that within one innings is tiny.

Captaincy is close to unmeasurable. A captain makes a decision, one outcome is observed, and the counterfactual never runs. There is no second universe in which he kept the slip in. Across a career, his decisions are hopelessly confounded with the quality of the bowlers he had available to make them with. This is the largest visible variable in cricket and the analysis of it barely exists.

Everything between deliveries is untracked. Field adjustments made mid-over, a word to a bowler at the top of his mark, the decision to take the pace off after a conversation with the keeper, over-rate management, the two minutes of persuasion that gets a tiring bowler through one more over. Cricket's records are a record of deliveries. A great deal of the captaincy and most of the coaching happens in the gaps between them, where nothing at all is written down.

The archive does not reach back evenly. This one constrains every historical claim made anywhere on the internet. Comprehensive ball-by-ball records exist for recent decades and thin out progressively going backwards. Men's international cricket is covered better than domestic cricket, which is covered better than women's cricket, which for much of its history has scorecards and no delivery-level record at all. Any analysis requiring a ball-by-ball row therefore quietly begins in whatever year its archive begins, and comparisons across eras are either impossible or are made on scorecard aggregates that lack every field this article has been discussing. When somebody shows you a chart of how the game has changed over time, the first question is where the data starts and why it starts there.

The gap between the analyst's desk and the dressing room

The day-to-day job is less glamorous than the concepts, and the skill that separates good analysts from good modellers is not modelling.

Before a series, an analyst prepares opposition dossiers from coded footage and aggregated data, and then reduces each one to something a player will actually read. The reduction is the work. A batter does not want fourteen pages on an opposition attack. He wants three sentences about the one bowler who will genuinely trouble him and a clip reel of that bowler's stock ball. The other eleven pages exist so that the three sentences are right, and then they stay in the folder.

During a match, communication is restricted. There is no live coaching from the dressing room in most formats, so the work has to be done before play and delivered at intervals. Between innings, at drinks and at the break, what the analyst supplies is mostly description rather than prescription: which lengths have gone for runs today, whether the ball is holding in the surface, which side of the ground is playing shorter, what the last four overs of the powerplay actually cost.

Afterwards there is coding, review and the unglamorous business of keeping the database honest. Over a few seasons the analyst also becomes the institutional memory of the side, the person who knows that a plan was tried against a particular batter two years ago and precisely why it failed, which is knowledge that would otherwise leave with whoever happened to be captain at the time.

Communication swallows most of the effort for a simple reason: a finding nobody accepts is worth exactly nothing. The dressing room is not a peer review panel. It is a group of people under public pressure who already hold beliefs about what happened, formed at high speed while it was happening to them, and a spreadsheet arriving at ten to eleven on the morning of a match will lose that argument every time regardless of whether it is correct.

The analysts who succeed do three things. They frame findings as questions rather than verdicts. They deliver at the moment a player can act on it rather than the moment the analysis finished. And they say plainly when a sample is too small to support a conclusion, which builds the credibility that gets spent later on the one finding that really matters. An analyst who is right about everything is not being listened to. He is being humoured.

This is why both popular pictures of the job are wrong. The analyst is not a laptop oracle telling the captain where to field, and not a clerk cutting video. He works across a boundary where one side deals in distributions and the other deals in a specific afternoon, and most of the value is created in the translation rather than in the computation.

The questions to ask of any cricket data claim

Seven, in the order that saves the most time. They work equally on a television graphic, a chart on social media and a briefing document.

How many balls is this? Not how many matches, not how many innings. How many deliveries, after every filter has been applied. If the answer is not stated, assume it is small, because a large number is the sort of thing people mention.

Was it measured or modelled? The path to the pad was measured. The path to the stumps was projected. Expected runs and win probability are outputs of something fitted to other matches. All three arrive in the same graphics package.

Who coded the judgement fields, and to what definition? Length, control and false shot are human decisions taken against a written taxonomy. Numbers from two organisations that define them differently should not be compared, and routinely are.

Which conditions produced it? Data pooled across surfaces, grounds and eras hides the effect you care about. Data split by them may have nothing left to say. Ask which of those two you are looking at.

Why does this matchup exist? If a bowler's record against a batter is being quoted, remember that a captain deliberately created every one of those deliveries, usually when circumstances suited. The record contains the captain's beliefs as well as the batter's ability.

Does the archive cover the period being claimed? Ball-by-ball data thins going backwards, and is thinnest in the parts of the game that are least covered generally. A chart that starts in a particular year usually starts there because that is where the data starts, not because anything happened that year.

What would change your mind? This is the one that matters. The teams that get real value from analysis are not the ones asking what the data says. They are the ones asking what the data would have to say for them to abandon a plan, and then checking honestly whether they have enough of it to be worth abandoning anything.

If you want to build a feel for this without a database, the cheapest exercise costs one innings of your attention. Pick a batter, write down before he starts where you think he will score and which length will trouble him, then check the scoring map afterwards. The gap between what you saw and what actually happened is precisely the gap that analysis departments exist to close, and it is generally wider than anybody expects, which is the entire reason the professional game now employs people to close it.