Why 145 Runs Reads as 178 in xR: A Domestic-Season Data Audit on Slow Pitches

A domestic T20 evening at the Sher-e-Bangla National Stadium. Twenty overs gone, the board reads 145/8. The next night, the ball-tracking feed pushes out an expected-runs (xR) figure of 178. In the model's ledger, that batting unit underperformed by 33 runs and the bowling unit leaked 33.

In the same spell, watching from beside the dugout, I noticed the left-arm spinner who conceded 22 in four overs sat bottom of that round's player ratings. I had been tracking him for six months; across four spells at Mirpur his bounce-sensor readings kept landing in the low-bounce zone. On the night the scoreboard said 145, the pitch was really a 135 surface.

That number stopped me. In 2026 I loaded every shot of Croatia's seven World Cup matches and France's seven into a hand-built xG spreadsheet — Croatia's open-play xG was 1.10, France's 2.40. I published before the final that France would win; they won 4-2. The lesson was singular: a scoreline and an expected number are not the same thing. Domestic cricket's xR models now stand exactly where that spreadsheet once did.

Context

Expected runs is not a new idea. Borrowed from football's xG, the metric has spread from franchise broadcasts to domestic-league graphics. The mechanics are simple: stadium ball-tracking cameras record each delivery's pace, bounce, line, length and the batter's shot type. The model matches those against a global outcomes database and returns how many runs such a delivery usually yields.

My work starts from the opposite end. Provenance first, sample size second, model assumptions third, practical decisions last. The hand-built spreadsheet I started in 2026 for the Bangladesh Premier League still sets the rule: before accepting any number, I need to know where it came from, how much data it rests on, and what it assumes.

In a domestic season those questions matter more, because samples are small and pitch variation is wide. An international series gives two sides four or five matches; a domestic league stages three or four games a week on the same square. Pitch fatigue, grass moisture and rolling rhythm all fall outside the model. Yet xR depends on precisely those variables.

Core analysis: where the xR model fails

I sat down with every ball of that innings. Four places kept producing the same error.

First, the pitch baseline. Most xR models train on flat, quick surfaces in England, Australia and India. Slow, low-bounce wickets at Mirpur or Rangpur are a minority in that database. A top-edge delivery worth 1.4 runs in England is a 0.8-run delivery in Dhaka. Feed the model the former and the innings inflates on its own.

Second, outcome leakage. xR is sold as result-neutral, but field placement and shot direction creep into the feature set. On a slow pitch, when a captain pushes deep midwicket and long-on back, the batter is forced into singles. The model scores those singles as low-value, when they are really the pitch talking.

Third, the value of a wicket. xR treats every wicket as an equal terminal event. On a slow surface, a set batter's wicket and a tail-ender's are not worth the same. In that innings, 40 of the 145 came from a seventh-eighth-wicket stand; the model reads it as overperformance without separating it.

Fourth, bowler workload. The model does not know whether the spinner is playing a third straight match or returning from a four-day Test. My ledger logs each bowler's 14-day over-load before every spell; in that Mirpur round the top three spinners averaged 52 overs. Higher workload shortens length, shorter length leaks runs even on a slow pitch — the model cannot see it.

Fifth, sample size. A single domestic season might yield eight or ten innings on one pitch. Eight innings cannot produce a reliable pitch coefficient. I have kept the same threshold since 2026: below 30 matches, no conclusion, only a note.

Contrarian angle

The easy call would be to scrap xR. But the report I built in 2026, comparing 306 pre-COVID Bundesliga matches with 92 post-restart matches, taught me the opposite lesson. Home win rate fell from 43.3% to 33.3%, home xG from 1.54 to 1.31 — yet 92 matches are not enough to rewrite home-advantage theory. I said so, and that caution advanced my career.

The same logic applies to xR. The model errs, but a known error type is a usable one. The opposite trap is real too: if I keep adding context — humidity, wind, time of day, rolling counts — eventually no number survives. So I rank context by materiality; only pitch type and workload are mandatory adjustments, the rest stay in the notes.

Budget matters here. Annual subscriptions to foreign black-box xR platforms are not affordable for our domestic sides. So I pick cheap proxies but never skip validation: core metrics (bounce, length, workload) first, premium features after. Explaining why spinners like Mehidy Hasan Miraz or Taijul Islam stay so consistent at Mirpur needs no expensive model — it needs a correct sample and honest context.

Takeaway

Next round I am watching two things. One, how quickly the domestic feed's xR graphic synchronises with the pitch report — without that, the number is half a story. Two, whether any side keeps a pacer's workload curve below threshold — Taskin Ahmed or Mustafizur Rahman. If xR is not the scoreboard's truth, workload is; and that can be measured, even on a small budget.