The Invisible Map of the Death Overs: The Column That Said More Than 85

The Invisible Map of the Death Overs: The Column That Said More Than 85

Arnos Vale Ground, Kingstown, 16 June 2026. The second ball of the 19th over of Nepal's innings settles into the keeper's gloves and the board reads 85 — all out. In the stands, the Bangladeshi flags suggest a fine evening of bowling. My right-hand column says something else. At 15 overs Nepal were 67/6. My phase model put their expected runs in overs 16 to 20 at 37.2. They got 18. The gap was 19.2 runs, and the final margin of the match was 21. Almost the entire result of the evening sat in a column the broadcast never shows.

Tanzim Hasan Sakib's 4/7 is the brightest line on the scorecard. It does not explain why Nepal stopped at 85. The explanation lives inside the over numbers, and that is where this piece lives.

I found the match in the columns before I found it on the screen.

A note on data limits, before anything else

I open every piece like this because my first job taught me that confident data and correct data are not the same thing. Some of the figures here are outputs of a retro-run phase model built on public ball-by-ball logs. The spray charts I tagged manually from television frames, so a few boundary-line positions carry roughly ±1.5 metres of error. Seven matches of a World Cup is a small sample. Small samples produce big stories, and I have known that since 2026 — which is why I refuse to publish a conclusion built on fewer than ten matches.

Even so, the question here is not sample size. It is how we divide an innings.

Context: from football xG to cricket xR

In 2026, as a junior data analyst at Brisbane Roar, I built an xG model for the 2026-17 A-League season. Jamie Maclaren scored 19 goals from 16.8 xG. When I took the number to the coaching staff, nobody was excited. So I spent three weeks re-watching every Roar goal to verify shot locations, then published a thread on a football blog. That thread gave me a rule I have never broken: no single metric can carry a conclusion.

At the 2026 World Cup, working remotely as a junior Opta data logger, the Australia-France match changed my reading habits. Aaron Mooy covered 12.3 km, the most of anyone on the pitch. My first read was that Mooy had controlled the game. My PPDA count said Australia pressed at 14.2, and France generated 2.1 xG. I re-watched the match logging every French entry into the final third, and the distance number collapsed into context. That distance was not a stat; it was a map of the game.

In 2026, when the A-League resumed in a NSW hub, I modelled home advantage across 120 matches. Brisbane Roar's home xG differential fell from +0.31 to +0.08. Warren Moon used the report. I warned that the sample was too small for firm conclusions. The empty stadium taught me that atmosphere leaves a data shadow, and the temptation to measure the shadow is where most analysts go wrong.

Crossing back to cricket, I decided the xG logic could not be copied, only its skeleton. In football, xG asks how good a shot was, separately from whether it went in. In cricket, the paired question is what a batter was expected to score from a given delivery, against what he actually did. The rest is phase splitting and field mapping.

Core: the death overs contain a second set of death overs

Modern T20 analysis splits an innings into powerplay (1-6), middle (7-15) and death (16-20). The split is tidy and, to me, insufficient. The 16th over is not the 20th over. In the 16th, fielders are still up, and the captain is still hoarding options. In the 19th, the ring empties and the per-ball maths changes.

The death overs are really two phases — 16-17 as the bridge, and 18-20 as the closing. Blending their economies into one number pushes decisions the wrong way.

In my ball-by-ball log, Bangladesh's bowling split at the 2026 World Cup looked like this: economy under 7.1 in overs 16-17, and above 8.9 in overs 18-20. Their reputation as a death-bowling side was built mostly in the bridge phase. In the closing phase they were ordinary rather than fearsome. Against Nepal there was no closing phase to discuss, because the match was already decided.

Nepal's innings did not break in the 18th over. It broke in the 16th. From 67/6 at 15 overs, Nepal's set batters spent the next two overs pushing into the middle, where Bangladesh's fielders were already waiting. My log shows only two of their shots in overs 16-17 travelling into high-value zones square of the wicket. The rest went straight down the ground to a field set for exactly that.

This is where xR earns its place. The real job of an xR model is not to predict runs. It is to show where the prediction is wrong. A batter who makes 25 off 20 looks slow. xR points out that on those lines, lengths and that field, an average batter would also have made 25 — so 25 is not failure. Staying there is. The reverse holds too: 35 off 20 against an xR of 28 means the batter played well and the bowling plan failed.

For Bangladesh's bowling, my most useful indicator is not economy. It is the share of deliveries bowled into low-xR zones. Against Nepal, four consecutive balls in the 16th over landed outside off at shoulder height. Nepal took 2 runs and lost a wicket. Tracking data put those deliveries between 0.18 and 0.24 xR — near-zero expected runs per ball. That was not accident. Owning that corridor from a left-arm seam angle is a plan.

I am not arguing against the yorker. I am arguing against yorker obsession. Bangladesh's bowlers have been practising fewer yorkers and more hard-length balls into a wide corridor, because a missed yorker becomes a six, and a left-armer's cutter angle increases the miss rate. Reading match state means the risk calculation should differ between the two phases — not taking the risk of winning, but taking the chance of avoiding loss.

Bangladesh's death-bowling strength is not the yorker. It is the field map. Splitting cover and point in the 16th over, placing the sweeper at deep square leg instead of long-on — these decisions never reach a scorecard, but they change a batter's shot selection. A coach once asked me what measuring field positions was worth. I said it shows whether the batter is quietly changing his intention. What football calls off-ball movement is, in cricket, the fielding map: the geometry of pressure applied without touching the ball.

Now the other side. Bangladesh's death-overs batting problem is not technique. It is innings architecture. Their set batters are gone before the 17th over, leaving overs 18-20 to short-term all-rounders with no habit of taking per-ball risk — and putting them there is accepting a low economy. At the 2026 World Cup, Bangladesh's strike rate in overs 18-20 dropped below what their powerplay produced. That is rhythm, not capability, and rhythm is a coaching problem more than a selection one.

The Dallas match against Sri Lanka on 7 June 2026 fits the same frame. Sri Lanka made 124/9, Bangladesh won by 2 wickets. What the eye misses: Bangladesh's overs 16-17 in Dallas told the same numerical story as their overs 16-17 against Nepal, with only the opposition quality changed. The same plan worked twice against very different batting sides, and in both cases Bangladesh conceded in the closing phase. Repetition implies design.

The contrarian angle: correlation is not causation

The easy story is that Bangladesh's phase split works, therefore their death bowling is strong. I will not go there, because I have seen too much analysis with pattern and no mechanism. Bangladesh's plan worked against Nepal, that much is true. But much of Nepal's 85 was self-inflicted: my log counts eleven deliveries where the batter's footwork was static and the shot was pre-selected. The ball was not good. The calculation was bad. Full-member sides do not make those errors at the same rate.

By the same logic, I will not claim Bangladesh's death bowling is now a genuine strength. Good in the bridge phase, middling in the closing phase — that mixed picture is what the model actually says. I trust the model only after it survives a cold Brisbane night. In my personal test set, a phase gap becomes meaningful only when at least two seasons of data point the same way. I have one.

There is another trap, written in my notebook. Contrarian conclusions become a personal brand, and brands defend themselves. "The Nepal match wasn't decided in the 18th over" is an attractive line. If the next six matches produce the opposite data, the attractive line has to go, and that should not feel embarrassing. The method is simple: pre-register the hypothesis, test its robustness, publish what survives.

The next signal

The 2026 T20 World Cup runs in India and Sri Lanka across February and March. My next three months have one task: building a closing-phase xR baseline for those spin-friendly, dew-soaked surfaces. Dew shortens grip, which rewrites the cutter angle entirely. The side that works out first where its field map belongs in overs 18-20 wins the trophy inside that margin. My database currently holds the columns. The pictures arrive in a few months.

Who reads those columns first — is that a question of skill, or of patience?