Death-Over Entropy: Where Bangladesh's T20 Batting Model Fractures

A BPL match in Chattogram, end of the 16th over. The chasing side needs 74 from 52 balls — 8.5 an over — with eight wickets in hand. The man beside me in the press box says the game is in the bag. I am looking at a different number in my notebook: that team's dot-ball rate in the death overs across its previous three matches sits at 41 percent. Second ball of the 17th, a push to cover against the left-arm spinner for a single — dot. The 18th over: four dots in a row. The 19th: two wickets. They lose by 11. The scorecard prints "middle-order failure." My log prints something sharper: three of those four dot balls were balls that deserved to be hit. The collapse did not begin with the bowler's skill. It began with the batsman's decision.

That night I opened the laptop and pulled the ball-by-ball log I have kept since 2026. A pattern surfaced that no broadcast graphic ever shows.

What the dataset is, and what it is not

Start with the sample. My log holds ball-by-ball records for 487 T20 innings from January 2026 to December 2026 — 294 from the BPL and 193 from international T20. For every ball I record five variables: runs, wicket event, line-and-length bucket, the batsman's shot zone, and the pressure metric in force before that ball, which is required rate minus par rate.

Attached to all of it is a context integrity note, because no number means anything outside its environment. The four venues behave differently. Mirpur is slow, the ball arrives late. Sylhet offers uneven bounce. Chattogram is batting-friendly with short boundaries. Dhaka is flat and does not grip. Same team, same batsman, effectively a different player on a different surface. So every verdict carries three lines ahead of it: which pitch, which season, how many balls.

One thing needs saying plainly, or the rest gets misread. Football's xG logic does not translate one-to-one into cricket. xG grades a shot — position, body part, type of assist, distance from the defender. Cricket's closest equivalent is expected runs per ball, indexed by shot zone and length bucket. The mapping breaks at the nature of the outcome. In football an attack ends in a goal or it does not — binary. In cricket a ball's outcome is never binary. An over ends after six balls, an innings ends after a fixed number of overs, and strike rotation means two batsmen are being accounted for at once. Cricket's clock runs in balls, not minutes. Ignore that distinction and every death-over calculation walks the wrong way.

A model is a monastery: you enter with noise, and you leave with discipline. This log has been no different — half of what I wrote in the first six months has since been deleted.

Where the fracture shows

The conventional picture of Bangladesh's T20 batting is: slow powerplay, serviceable middle, weak finish. My log broadly agrees, but it locates the cause somewhere else.

In the first seven overs — the powerplay plus the over after it — Bangladesh's run rate is 7.8. The average across the world's top eight T20 sides is 8.9. The gap is roughly 1.1 runs per over. Here is the first surprise: Bangladesh's wicket-loss rate in that same window is also low, 1.4 per innings. Slow, but safe. The base is not the problem. The break begins at the 14th over.

Between overs 14 and 20, Bangladesh's run rate is 8.1. The top eight average 10.4. Run rate alone understates it, because a 2.3-run gap can be hidden by a couple of innings. The real signature is in the dots. Bangladesh's dot-ball rate in this window is 37.6 percent. The highest dot-ball rate among the top eight is 31.2. That is a gap of more than six percentage points — roughly four extra dot balls per innings, precisely in the phase where every dot ball costs the most.

Not every dot is a failure. I split them into two types. Structural dots: the ball was too good to hit — yorker, slow cutter, length outside off. Decision dots: the ball was hittable, but the batsman chose the wrong shot or let it go. Of Bangladesh's 37.6 percent, 22.4 percent are decision dots. Roughly three in five dots are the batsman's call, not the bowler's craft.

This is where a metric I use comes in: death-over entropy. In plain terms, how spread out a team's run production is across that window, versus how concentrated it is in a few big hits. High entropy means runs are scattered — six, four, one, two. Low entropy means runs are stacked in a couple of explosions with long silences between. Bangladesh's death-over entropy has fallen steadily from 2026 to 2026. The index was 0.71 in 2026, where 1 means maximum distribution. By 2026 it sits at 0.54.

That means Bangladesh's death-over runs now lean harder on a few big hits. Two sixes land and the score looks fine. They do not, and the collapse is theatrical. It explains the all-or-nothing quality — 55 runs or 28, rarely anything between.

Now down to the batsman level. In the last four overs — 17 to 20 — I looked at Bangladesh batsmen with a minimum sample of 100 balls. Three of the top five are anchor types: they rotate from ball one, wait for boundaries, avoid risk. That shape works in the middle overs, when the required rate is under control. It turns dangerous in the last four. Expected runs per ball rises, while the anchor's risk aversion stays fixed. The result: balls consumed, runs withheld.

I call this the anchor tax. The arithmetic is simple. Say an anchor makes 7 off 6 in the 17th over — not bad personally. The team needed 11 that over. Tax: four runs. Next over he accelerates to 12 off 6, but the requirement has climbed to 14. The tax compounds, and the pressure doubles on whoever is left for the final two overs. This compounding tax explains Bangladesh's death-over innings far more consistently than a shortage of six-hitting.

The 14-to-16 window deserves its own look. Bangladesh's strike-rotation rate there drops below 80 percent — one strike change roughly every five balls. The top sides sit between 85 and 90 in that window. Fail to rotate and one slow batsman consumes 12 to 14 balls, and by then the innings tempo is fixed. World-class sides use this window to build the platform for the last four overs. Bangladesh stalls in it.

Pitch variance is part of this. At Sylhet and Mirpur my log shows death-over entropy naturally lower, because the ball grips and big hitting is hard. But at Chattogram, with short boundaries and a flat deck, Bangladesh's entropy is still low — 0.56. The problem is not the pitch. If it were, the number would climb on a flat deck. It does not.

A side note most people skip. Death-over batting failure is not only a batting fault. Look at bowling workloads in the BPL. The schedule is so dense that frontline death bowlers cannot sustain 25 to 30 death balls across consecutive matches. So the final overs often fall to a third or fourth-choice bowler, and the batsmen's decision quality becomes more visible. Bad batting does not need bad bowling; rather, a gap in bowling-workload management exposes the batsmen's decision weakness.

The numbers can be true and the verdict still wrong

Here is my loudest caveat. Every number above can be accurate and the conclusion still wrong — if I stop at a shortage of talent.

Bangladesh's death-over problem is not talent. It is incentive. Look at the BPL auction structure. A batsman's price is set by his safety. Anchors cost more, because they deliver consistent scores, and consistent scores sell. In the same auction a finisher is priced on his peak six-hitting; his failure rate is never counted, because it is not sellable. The system manufactures two player types: the safe anchor and the lottery-ticket finisher. The middle ground — where death-over risk is managed ball by ball, where decision dots are driven toward zero — nobody buys, because it is not yet measured. What cannot be measured has no market price.

The ghost games of 2026 have a bounded role here. I decided in advance what would count as a 2026-specific effect. Three indicators: whether dot-ball rate rose in empty stadiums; if so, whether it rose in the structural or the decision component; and whether the shift showed up in bowlers' length patterns. IPL 2026 was played entirely without crowds in the UAE. In that window of my log, dot-ball rate did rise, but mostly in the structural component, not the decision component. The absence of a crowd helped the bowlers; it did not fool the batsmen.

Which means Bangladesh's decision-dot problem is not environment-dependent. It is structural. Empty stadium or full, the result holds. I am stopping the 2026 comparison there, because cricket's window is small, and stretching it further means hunting for evidence in the wrong place.

What role does the eye test play? Bounded, but not zero. From the stands I can say a batsman plays fast under pressure. That generates a hypothesis, not a verdict. The verdict comes from the log. When the eye says Liton Das accelerates under pressure and the model says his decision-dot rate sits below the team average, I print both and record the disagreement — I do not rule.

There is a counter-explanation I respect. Perhaps low death-over entropy is not poor batting but deliberate strategy: the team knows its strength is not six-hitting, so it lowers risk and wins another way. That is not a bad argument. But it is testable. If it were strategy, low entropy in the last four overs would coincide with a high win rate. My log says the opposite. In matches where Bangladesh kept entropy below 0.55 in the last four overs, the win rate was 38 percent. Where the index exceeded 0.65, the win rate was 61 percent. Low entropy does not produce wins; it marks finishing failure.

One more addition, or the analysis turns one-sided. Consistency has value too. Bangladesh's powerplay safety means fewer collapses; aggressive sides occasionally lose three wickets for 30. The structure should not be discarded wholesale. The question is sequencing: which overs for safety, which for risk. In the current sequencing the safety window is too long and the risk window too narrow.

What I will watch next season

My headline indicator next season is the decision-dot ratio in the last four overs, not the raw dot-ball rate. If that ratio drops below 22 percent, the team has not merely increased aggression — it has changed the quality of its decisions. If run rate rises without entropy rising, the problem was never the batsmen; it was the run-accumulation plan for overs 14 to 16.

Let one question stay open. The finisher we pay the most for at auction — are we buying his sixes, or his capacity to avoid dot balls? Until that question is answered, next season's log will again collect 41 percent dots, and the broadcast will again say the game was in the bag.