The Dead Column in the Middle Overs: The Model That Kept Lying About Bangladesh's Batting

Hook

At the end of the 14th over the board at Mirpur read 92 for 3. Two columns sat open on my laptop. The left-hand column — built from seventeen months of ball-by-ball data running from April 2026 to November 2026 — said this innings would finish on 158. The right-hand column counted only actual runs: 92. The model needed 66 off the last six overs to survive. Seven balls later the score read 104 for 6. The left-hand column did not move. Not by a millimetre.

That was the night I understood the problem was never the forecast. The problem was my xRA column — Expected Runs Added, the projected run value of each ball in the middle overs. Across seven consecutive matches that column displayed almost the same number, while actual runs drew a sawtooth. A column that shows zero variance inside a high-variance environment stops being a column. It becomes a corpse.

Once, in Rajshahi, the xG column stopped being a number for me and became a confession. I am standing in the same place again, in a domestic format. The admission is uncomfortable, because in a World Cup cycle a wrong model costs the most — every miscalculation here lands on four years of selection policy.

Context: What the Model Measured, and Why It Measured It

When I launched the 'Expected Truth' blog in Rajshahi in 2026, the whole method rested on one rule: put the number on the table first, then let it confess. In football that number was xG and PPDA. Cricket has no direct replacement for xG, because cricket is a game of discrete events — every ball is a fresh start, and before every ball the wickets already fallen are still standing there. Fallen wickets mean the price of risk has shifted.

So I work in three layers.

Layer one — the powerplay platform. The ratio of runs to wickets across the first six overs, and the velocity of the collapse if that platform breaks before the second powerplay.

Layer two — middle-over xRA. Venue-normalised expected runs per ball from overs seven to fifteen. Four inputs build it: the batter's shot map, the bowler's line-and-length cluster, the depth of the field setting, and innings weight (wickets in hand, balls remaining).

Layer three — death-over conversion rate. What percentage of the middle-over savings actually converts in overs sixteen to twenty.

From football's PPDA I borrowed 'pressure delivery share' — the proportion of deliveries a batter faces with fewer than two seconds of decision time. The borrowing was not decoration. It changed one concrete conclusion. In the Euro 2026 final, Italy's PPDA was 10.2 against England's 15.6, meaning Italy pressed far higher up the pitch. Transplanted into cricket, that number shows Bangladesh's bowling unit generating the same high-pressure architecture while the batting unit cannot absorb it. Creating pressure and absorbing pressure are two separate skills, and because we have successfully acquired the first, the deficit in the second has been hidden.

For this cycle we tagged eleven matches — group stage, Super Eight, warm-ups. Three analysts, working from Dhaka and Rajshahi, on the same ball-by-ball files. To each match we attached four environmental variables: flight hours, rest days between matches, boundary-size variance across venues, and an estimate of evening dew point. When the stadiums emptied in 2026, I learned that environmental variables are not decoration; home advantage is a ghost variable, and if you do not measure it, it lives inside your residuals.

The question is now simple. What did my xRA column show across those eleven matches, and why did it die in the middle?

Core: The Anatomy of a Dead Column

Let us put the baseline on the table first.

Between April 2026 and November 2026, across seventeen months, Bangladesh's middle-over xRA stood at +0.086 runs per ball. In neutral conditions, Bangladesh's batting in that phase was converting roughly nine-hundredths of a run per ball above par. On home soil, particularly on Mirpur's slow, turning track, the tempo of the wicket depended less on wickets falling and more on balls bowled. In that frame, 7.9 runs an over and a wicket every 22 balls was the structural rhythm.

Across the first six matches of the current World Cup cycle, the same column read −0.041. The difference is 0.127 runs per ball. Someone will say that is only six or seven runs — spread across 54 middle-over balls, 6.86. And that is precisely where the model opened its mouth and admitted it was lying.

Because Bangladesh lost by 37 runs. Once by 41. Once by 26. The average gap is 6.86. The actual deficit sits in the high teens. The problem was never in the mean. The problem was in the tail. That distinction is the centre of this analysis.

A linear model cannot hold the tail.

My xRA is a linear model. Wickets are assumed to fall at a constant rate — every ball carries equal risk. In cricket, wickets fall in clusters. One wicket raises the probability of another across the following three overs, because a new batter cannot read the innings and cannot rotate strike to disturb the bowler's rhythm. A linear model forgets this second-order effect. The same phenomenon appears in football: you cannot predict the xG drop in the opening minutes after a red card, because a ten-man side drops into a shell and the game state changes entirely.

So three separate events were happening simultaneously in the middle overs this cycle.

First event: platform decay.

Bangladesh averaged 47 for 1 in the powerplay through 2026 and 2026. Across the first six matches of the World Cup cycle, 41 for 2. A five-run shortfall is not a tragedy on its own. The extra wicket is the tragedy. When a number four is exposed in the powerplay, and that individual is a consolidator by instinct, the innings-weight input for middle-over xRA sinks to the floor. The model knew how many balls remained. It did not know who was batting and what that batter was refusing to do.

This is where I found a gap in my own method. xRA assumes a batter willing to take risk. But Bangladesh's middle-order architecture has followed one rule for five years: when the powerplay breaks, the second-most experienced batter is held back for number seven so that the death overs contain a familiar face. A boardroom instinct for safety. That instinct has no place in the model, because the instinct is never written in numbers — it is written in the minutes of a meeting.

Second event: venue turn and boundary size.

Mirpur's average boundary is 64 metres. Several venues in this cycle — Caribbean and American grounds — sit six to ten metres wider. My model was trained on Mirpur shot maps. In those maps, chipping a six over the line is a mandatory stroke; on a 64-metre ground it clears. On a 70-metre ground it lands in long-on's hands.

How much does that translation cost per innings? Two to four boundary-bound attempts, of which one or two become catches — and every catch is an extra wicket in the middle overs, which inflates risk across the next three. The model cannot trace this sequence because it does not measure the distinction between stroke play and stroke translation.

Here I borrowed something from football I had not borrowed before. In basketball's half-court offence, two skills are measured separately: creating space and using space. A venue, a pitch, a boundary dimension creates space. Using it is a different craft. Bangladesh's batters learn to create space in domestic leagues and fail to use it on larger international grounds. That is not a mental failing. It is a technical mapping problem.

Third event: the bowling that concealed the batting.

This is where my editors wanted to make a mistake. After four defeats, the board meeting discussed changes to the death bowling. On paper it looked ugly. The data says the opposite.

Bangladesh's pressure delivery share this cycle was 38 per cent, against a tournament average of 31. Nearly four balls in every ten reached an opposition batter with under two seconds of decision time. Taskin Ahmed's economy was 6.8, Mustafizur Rahman's 7.1, Rishad Hossain's 7.4. That eight-over block ranked inside the tournament's top five.

The problem is not the death bowling. The problem is that when the first innings yields 118 to 125, every death-over error carries six times the weight. Two extras in a Rishad over make it an eight-run over; defending 170, the same two extras still make an eight-run over, but their weight in the outcome is halved. Same result, different value. That is the most undervalued truth in cricket.

And this is exactly where the Alexis Sánchez lesson applies. In January 2026, when he moved to Manchester United, I watched his xG per 90 fall from 0.61 to 0.43. The market was paying for character, demand and history. On-pitch output was secondary. Bangladesh's death bowling is being priced the same way — an economy of 7.1 looks clean, so nobody asks what that 7.1 was defending. An economy of 7.1 while protecting 120 is a completely different performance from an economy of 7.1 while protecting 180. The scorecard writes both in the same colour. The model would too, if I had not attached a match-state charge.

Franchise Market Value Note: What the Market Buys and What the Model Sees

A franchise auction room is an open market, and in that market two kinds of batter command roughly the same price. One strikes at 142, adds +0.03 xRA per ball, and rarely gets out before reading the pitch. The other strikes at 152 but carries −0.01 xRA, because his 152 is compiled in the easiest overs and never when the team needs it.

The second batter is paid more. Why? Because the market measures the most visible number, and strike rate is the most visible. A transfer fee is a story the market tells about its own fear, a narrative built from demand, scarcity and expectation, in which actual on-pitch production arrives last. The IPL auction room, the BPL auction, ILT20 contracts — the same rule everywhere. From a youth development perspective, this means the big franchises can buy talent without running academies, and young talent in smaller leagues becomes a satellite asset built only to be sold. Inside that structure, a twenty-one-year-old who anchors the middle overs gets locked into a five-year specialist role and is never permitted to take death-over risk. If the model measures only outcomes, it will record that containment as a virtue.

Small Column, Large Signal: The Night the Model Was Plainly Wrong

Data is a monastery: you sweep the floors before you see the vision. In this cycle my model was plainly wrong in four places, and those places need to be written down, because a method rots the moment its errors are buried.

First error: in a high-scoring match where the winning score was 182, my model called Bangladesh's 164 'neutral'. Dew fell so heavily at that venue that the spinners could not grip the ball in the second innings. We estimated dew from the conditions at seven in the evening; the dew actually arrived around nine. Two hours of error pulled our defend-the-ball projection twelve to fifteen runs too low.

Second error: fielding. A run-out and two long-boundary saves carry zero weight in the model. A good fielding side saves eight to eleven runs per innings in the middle overs. That sits outside the model because it is the final act of a chain — and Bangladesh did not lose that act once in eleven matches. The tournament's top three save statistics belong here. A model that does not measure its own best asset cannot produce its own best decision.

Third error, and the most important: human weight. In a World Cup cycle a middle-over ball is not merely a ball — it is weighted with the hopes of a hundred million people on a school bus radio. In football, xG stays fixed before a penalty kick; the hand shakes anyway. There is no camera on it. This variable cannot be modelled, and any attempt to model it is sleight of hand. Better to state it plainly: under pressure, Bangladesh's decision quality drops below the model's expectation, and that payment has to come from someone who has played twenty matches down the order.

Contrarian: The Fault Is Not Strike Rate, It Is the Order of Risk

After the tournament, in a Dhaka hotel room, I saw a draft of the board review. One sentence kept returning: 'Middle-over strike rate is too low, we need to raise intensity.' The sentence sounds reasonable and is wrong.

Our middle-over strike rate was not actually poor — 116.4, against a tournament median of 119.8. A gap of 3.4. Yet the xRA gap was −0.127 per ball. How does a 3.4 run-rate deficit become a 0.127 runs-per-ball deficit? Because the mean is similar while the distribution is not.

Look at the strike-rate distribution. First twelve balls: 102. Next twelve: 148. Next twelve: 94. This sawtooth has a specific cause — a new batter spends his first ten balls playing for himself, attacks for two overs, then breaks again. The result: 60 per cent of the team's middle-over shots come from 30 per cent of the balls. Damage is not caused by the absence of attack; it is caused by taking extra risk with the attack, because the risk is attached to the wrong ball.

In other words, the batting order is trapped in a fixed triangle: what should be in the middle is missing, and what should be on the right is missing. In statistical language, Bangladesh's problem is not skill. It is sequence.

Travel variables corroborate this. In the group stage, Bangladesh covered four cities in five days — the heaviest travel load in the tournament since 2026. At the Tokyo Olympics in 2026, Elaine Thompson-Herah ran 10.61 in the 100m and 21.53 in the 200m; between those two events a score-integrity variable was visible, namely recovery time into the next round after a 100m final. In our game, travel load is a recovery tax — roughly two to four runs per match per city change. What the middle overs might have looked like with a softer cushioning competence upon returning to Dhaka is only guesswork. The natural explanation for what happened on the field is simpler.

So the biggest question is this: should the model be taught, or should reality be bent to fit the model? The first is ruthless. The second is self-deception. I chose ruthless. A written message from a coach friend stays with me — an academy coach in Mirpur who was on the ground for all eleven matches. My numbers are an abstraction next to his notebook; his observation will be the foundation of my next design. For that reason I have abolished the holding-on role for six months.

Takeaway

I did not rebuild the model because it failed. I rebuilt it because the world changed. The ground and the environment in which a linear xRA band worked seventeen months ago no longer exist — not the boundary dimensions, not the pitch baking, not the dew point. I have added two things. One, a cluster-aware term: when two wickets fall in quick succession, risk across the next ten overs rises asymmetrically. Two, a shot-translation rigidity index, mapping small-ground shot maps onto large grounds.

The signal is patient; the noise is always in a hurry. The tournament is not eight months away — two Super Eight matches remain. Those two matches are not ordinary for this side. If xRA does not touch +0.05 again across them, the fundamental batting-order template must be broken before the 2026 bilateral series: the anchor at the top goes, and the best six batters own overs seven to fifteen regardless of competition. That is not self-sabotage. That is arithmetic.

The World Cup did not create value; it simply turned the lights on. The places still in the dark are where the next four years of selection argument will be fought.