Expected Runs vs Stadium Truth: Why T20 Models Fail Without Pitch, Dew and Match-State

Hook: Reading 119

On 9 June 2026, at Nassau County International Cricket Stadium in New York, India and Pakistan were playing a T20 World Cup group match. The drop-in pitch had been laid only months earlier. My ball-by-ball model was running on my laptop and said the par score under those conditions should sit between 148 and 156. India were bowled out for 119. Pakistan chased and fell for 113.

That night I wrote a line on the first page of my notebook; it still sits at the top of every brief I produce: the model said one thing; the empty stadium said another. The stands were not empty that night, but the surface was effectively a refused surface — one where batting demanded courage, and a model that cannot measure courage is only guessing.

I have been chasing numbers since the 2026 World Cup in Russia. In a small Sydney bedroom I built my first xG model in Excel and logged 1,248 shots. France scored four goals from 2.1 xG; Argentina scored three from 1.4 xG. The eye test said one thing, the data said the opposite. Since then I have run one habit: a verification loop between a number and what the ground actually shows.

Context: A Fast Cricket Analytics Market, a Slow Reality

The IPL, the Big Bash, The Hundred, the Hundred-ball formats globally — all now use expected runs (xR), expected wickets (xW) and one-ball win probability. Betting markets read them too. But these models carry a structural weakness that football does not.

In football, a shot's value needs roughly four variables: distance to goal, angle, type of assist and defensive pressure. In cricket, per-ball intent is not obvious. The same delivery may be pushed by one batter and slog-swept by another. If the model does not know intent, the ball's “value” becomes a distraction.

I work at a small client desk in Sydney. My job is to produce two T20 briefs a week — a pitch surface report, dew probability, and an xR curve from powerplay to death overs. Over the last two seasons I have manually verified roughly 380 T20 innings at scorecard level. From that bank of work, one conclusion holds: a cricket number is not true on its own — it becomes true only when pitch, dew, match-state and role context are attached to it.

A claim I hold firmly: I do not trust a number I cannot trace to a touch.

Clients often ask me for one number, one side to back. I answer with a question first: which pitch, which session, which dew level? It may sound tiresome to them. To me it is a rule. No number lives in a vacuum.

Core: Pitch, Dew, Match-State, Role

1) True par score is never a single number

Across recent IPL seasons I have seen the same venue produce two completely different decks within days. One night a flat deck, 210 is normal. Three nights later, grass trimmed for a match, that number collapses to 155. A venue-average model picks the wrong side of that line.

My model now holds three layers of par score: baseline, dew-adjusted, and game-state-adjusted. Baseline is the venue's trailing six-match strike-rate average. For the dew adjustment I merge the second-innings spin economy with catching-slip background data. For the game-state layer I add a chasing/defending wicket-probability curve.

An example: in a batting-friendly IPL venue, the first innings finished on 223. Chasing, the side scored only 41 in the last five overs and lost four wickets. The model had the chasing side at 62% win probability before the 50th ball. Reality was 17%. Why? Dew. In the second innings the ball came out slippery, spinners lost grip, and that single variable never entered the model.

Without pitch and dew, T20 model precision is an illusion.

2) Powerplay xR versus Death-over xR

Since last year I have split xR by phase — powerplay xR (overs 1–6) and death-over xR (overs 17–20). This exposes batters whose total xR looks good but whose distribution is uneven.

One number: a top-order batter carries an overall xR of 1.42 per ball. His powerplay xR is 1.71; his death xR is 0.98. If the opposing captain recognises the matchup, he can push this batter into the death overs and reduce his output. A single overall figure gives false comfort.

A finisher, by contrast, sits at 1.30 overall. That looks thin. But his death-over xR is 2.10 with a strike rate of 192 — more valuable in that phase than a top-order batter's 135. The match-state is different, so the value is different.

Without role adjustment an xR figure is an average; role-adjusted, it is a decision.

3) Empty stadiums and the source of home advantage

In 2026, during the global shutdown, I was working on data from the Bundesliga restart. Home win percentage fell from 43.3% to 33.3%. In the A-League Grand Final, Sydney FC beat Melbourne City 1-0 at an empty Bankwest Stadium; comparing PPDA with distance covered, I found the home xG advantage dropped by 0.25.

I ran the same test on cricket. IPL 2026 was played entirely in the UAE behind closed doors. In my dataset, first-innings average and second-innings win rate showed one clear thing — an empty stadium did not erase home advantage; it exposed the advantage's real source. What fell away was not umpire bias but dew and away batters' normal rhythm.

Empty stadiums did not erase home advantage; they exposed its source.

4) The Bumrah factor and misattribution

At the 2026 T20 World Cup, Jasprit Bumrah was Player of the Tournament with 15 wickets at an economy of 4.17. To a model, that is an unlikely number. But the number cannot explain the Bumrah factor on its own.

Bumrah's slog-over sequence, his slowly-bowled yorker, his powerplay line into the batter's inside edge — plug those three variables into a model and his per-ball xW sits roughly 0.015 above an average quick. That looks small, but across a 24-ball spell it is about 0.36 expected wickets. Over a tournament, that is the margin in several games.

I have reached a harder conclusion here. Much of Bumrah's success is Bumrah, yes. But it is also team context — he bowls inside a disciplined plan with fielders drilled to hold their nerve under pressure. Attributing the numbers to the individual name alone causes us to lose the team system.

Contrarian: The Gap Between Correlation and Cause

The largest trap in cricket analytics is treating correlation as causation.

Example: I have noticed over two years that sides taking more powerplay wickets win more often. Sounds reasonable. Look deeper, and wicket-taking capacity is a blend of field-setting, timing of bowling changes and reading opponent intent. Picking a side purely on wicket numbers is a mistake; you must verify the skill alongside those background variables.

Second trap — small samples are loud, large samples are honest. In five matches, a batter keeps a 210 strike rate and the media calls him a star. Over 30 matches, at 148, the five-game burst looks like an outlier. My model brackets strike rate by sample size. A five-innings strike rate gets a different weight from a 30-innings strike rate.

Third trap — format conflation. Test PPDA or strike rate cannot sit in the same table as T20. Scoring trends in the IPL cannot be mapped directly onto the Big Bash — different pitches, different balls, different fielding standards.

My rule is simple: a model that does not declare its signal-versus-noise threshold up front can never be falsified — and a model that cannot be falsified is not a model, it is a belief.

I treat Bumrah's 2026 tournament as a good test. His economy was 4.17 — almost unnatural in T20. But if I strip away his bowling plan and fielders' positioning and draw conclusions from economy alone, I will have mislabeled the Bumrah magic with a single number.

Takeaway: What I Will Watch Next Round

The 2026 T20 World Cup in India and Sri Lanka begins on 7 February 2026 and ends on 8 March 2026. I currently hold an unfinished live xR model with two layers still unlocked — dew probability and travel-fatigue context.

In the coming week I will watch three things. First, the sides that lower their powerplay tempo to save runs for the death overs — the slope of their xR curve is the signal for the next two series. Second, spinners bowling in empty stadiums and what changes in their flight and drift. Third, the workload of fast bowlers returning from medical comebacks — the noise of a small sample is most dangerous here.

I am not afraid of numbers. But I do not trust a number that has never heard the smell of the ground. When someone brings me only an average score next round, I will ask: which pitch? What dew level? And which outfield? Because without context, what remains is not analysis — it is a number's solitude.