Mirpur's Own Ghosts: The BPL Phase Model and an Open Ledger

Mirpur's Own Ghosts: The BPL Phase Model and an Open Ledger

I begin with a number, because it keeps returning to my notebook. Across the last three rounds of the current BPL regular season I logged eight innings in which a side struck at above 180 in the final five overs. Not one of those same eight innings reached a powerplay strike rate of 135. In the commentary box this is called "death-overs ability." In my ledger, five of those eight innings arrived at over fifteen with six or more wickets in hand.

Four of the eight were lost.

So the question is not who the best finisher is. The question is whether runs in the last five overs are simply a deposit made in the previous fifteen. Every round the answer shifts slightly, and this piece is an attempt to open the ledger: how I built a local phase model, where it held, where it broke, and where it still proves nothing.

Dew, the toss, and an incomplete dataset

Night matches at the Sher-e-Bangla National Stadium in Mirpur must be read as environment first. An evening start, a second innings finishing after ten at night, dew settling on the ball, the spinner's grip loosening—the side batting second gains a technical edge, and my log agrees. The problem is what sits beside that observation: no public Hawk-Eye feed, no fielding-placement maps, no release-point data. Imported European workflows do not survive the trip. Since 2026 I have logged ball by ball—bowler, a rough line and length class, shot type, where the fielder stood, whether dew had settled, wickets falling by over. Reaching version 2.1 took years, and every version carries a list of its own earlier errors. A dataset that hides its mistakes stops being a dataset.

My first work was in football, building a grassroots xG model because the Bangladesh Premier League deserved its own ghosts. Moving to cricket, I found the same problem: borrowing international benchmarks without reconciling them to local reality. A run rate of 8.5 means one thing in Bangladesh and another in Australia—different pitches, different outfield speeds, different dew.

I also keep 2026 in mind. Empty stadiums were the laboratory where home advantage stopped performing. Home benefit roughly halved, and players covered ground differently. Crowd noise unsettles a batter's concentration, but it does not physically move a fielder; what rises, my model suggests, is the reflex to pull the hand away late. That remains a hypothesis, not a finding.

The phase model: what the last five overs actually are

My model splits each innings into three phases: powerplay (overs one to six), middle (seven to fifteen), and close (sixteen to twenty). Within each phase I track four primary variables—dot-ball share, boundary share, rotation (ones and twos), and wickets lost—plus two context variables: a dew proxy for the second innings and a recent-form proxy for the side.

Roughly two-thirds of the variance in close-phase strike rate is explained by just two inputs: wickets in hand at over fifteen, and whether a set batter is at the crease. When I then added a separate "finishing ability" variable—each batter's historical close-phase strike rate—it added almost nothing. Much of what we call finishing talent is a mathematical shadow cast by wickets in hand and a set batter.

The strategic lesson follows. Sides that look devastating at the death are largely collecting interest on middle-overs restraint. A side that gambles for tempo in the middle phases looks helpless at the close, because nobody is left at the crease.

Middle-overs dot balls: the real lever

The strongest single predictor was not close-phase boundaries but the middle-phase dot-ball rate. A side that pushes dot balls below 42 percent between overs seven and fifteen will almost reliably strike above 180 at the death, whatever the group average. The explanation is simple; the implication is uncomfortable. A middle-overs dot ball blends two things—a good delivery and a batter's failure—and the scorecard shows them identically. About 60 percent of the dots in my log came from deliveries landing inside six metres; the rest from the batter's decision. That classification is subjective, and I print that limitation beneath the table.

Mirpur's home advantage: how much belongs to the ground, how much to the toss

Most cricket talk about home advantage blends into the toss. On dewy nights at Mirpur the chasing side gains an edge, and chasing usually means losing the toss. So "won at home" and "won batting second" can become the same sentence. I therefore split the calculation: raw home win rate, then residual advantage after controlling for toss outcome and the dew proxy. Where the dew control enters, home advantage contracts; where it does not, it inflates. I am not arguing home advantage is absent. I am arguing that a large share of it is unreadable until toss and weather are priced in.

What an open ledger means

In version 2.1 every delivery is a row with a timestamp, a classification standard, and a note on what changed. The structure is append-only: old rows are never deleted, only annotated in later versions. In franchise cricket that matters more than elsewhere, because money, ownership, and performance evaluation move together. A six-to-ten match sample decides contracts, and two successes plus three failures produce the label "big-match player." An append-only, time-stamped log gives the argument a floor.

What the model cannot prove

Correlation is not causation. I show that wickets in hand and close-phase strike rate move together; I do not show that keeping wickets causes the runs. "Set batter" is itself a selection bias—the batter still there at over fifteen usually received easier bowling. My dew proxy is an estimate drawn from start time and venue, not a hygrometer reading, so dew and second-innings advantage are difficult to separate because they move together. And a residual is a story the model did not expect; I read it slowly. When the model predicted 112 and reality produced 158, those 46 runs usually carry a fielder's wet gloves or one failed over, and that is where the next variable lives.

What I will watch next round

Three signals. First, middle-phase dot-ball share: above 42 percent, a strong closing record is a facade awaiting correction. Second, whether the side fielding first after losing the toss gains measurably once dew settles, tracked by how spin grip changes after over six. Third, sides that lose despite wickets in hand at over fifteen—where exactly the dots between overs seventeen and twenty are spent.

A league does not manufacture its own ghosts; borrowed ones do not last. With our own ledger open, the argument at least stands on a table rather than an instinct.