Method

The data

Projections are built on ball-by-ball match data from Cricsheet, covering men's international cricket, the IPL, the Big Bash, the T20 Blast and The Hundred. Every first-innings total in that archive feeds the venue baselines.

Format buckets are kept separate

A ground's T20 average tells you nothing useful about an ODI there, and The Hundred is a 100-ball competition rather than 120, so its totals sit on a different scale entirely. Each format keeps its own bucket. Blending them — which is easy to do by accident — produces projections that are wrong by wide margins.

Recent matches are weighted more heavily

Scoring levels drift. A ground that averaged 160 five years ago may average 190 now, and an all-time average buries that. Where there is enough recent sample, the recent figure is used; where there isn't, the model falls back to the full history and the preview says the sample is thin.

The pitch setting — what it is for

Every projection starts from what that ground has historically produced. But a ground is not one surface. Curators prepare a fresh strip for most matches, and two pitches at the same venue a week apart can play very differently — one holding firm and true, the next gripping and turning. The historical average cannot know which one is out there today.

The pitch control is where you tell the model what you can see and it cannot. It runs from pace-friendly at one end to spin-friendly at the other, with neutral in the middle.

What the ends mean

• Pace end. Hard, true surface. The ball comes on to the bat, strokes carry, and spinners get little grip. Totals go up.
• Spin end. Dry, worn or gripping surface. The ball holds up, timing is harder, and spin becomes the threat. Totals come down.
• Middle. Use the venue's own historical character, with no override. This is the right setting when you have no specific information — guessing is worse than leaving it alone.

How much it moves the projection

The effect is deliberately bounded, because a pitch read is a judgement and should not overwhelm several hundred recorded matches. Across the full range of the control:

• Twenty20: about 20 runs from end to end
• One-day internationals: about 40 runs
• The Hundred: about 16 runs

The longer the innings, the more a surface can compound, which is why the ODI range is wider.

It also changes the player markets

The setting does not only shift the total. Moving towards spin raises the projected wicket share of the spinners in each attack and lowers the seamers'; moving towards pace does the reverse. A side that depends on spin is worth less on a surface that offers none, and the player tables reflect that.

When to move it

Move it when you have a specific reason: a pitch report describing the surface, heavy recent use of the square, an unusually dry or green strip, or a curator's own comments. Do not move it to make a number look better. If you find yourself adjusting it until the projection agrees with what you already believed, the setting has stopped being information and become a way of arguing with the data.

Home advantage is measured, not assumed

Across 246 men's T20 internationals since 2023, the batting side averages about 19 runs above the venue baseline at home and roughly nothing above it away. That measured figure is applied rather than a guessed one. One-day internationals since 2019 show the same gap, so the same figure is applied there too. It is a global average: some sides have a far larger home effect than others, and the model does not yet separate them, because the per-team figures do not hold up when tested on matches outside the period they were fitted on.

Recent batting form

Team ratings move slowly, so in October 2026 the model began adding a correction from each side's T20I first innings over the past twelve months, measured against the model's own projections and shrunk towards zero when there are few matches. It was tuned on 2024–25 and kept because it also improved 2026 matches it had not seen: average error 27.5 → 25.6 runs, bias −4 → 0 (both figures from before the November 2026 leak fix, so lower than today's like-for-like numbers). It applies to T20 internationals between the main teams; for other matches there is too little recent data.

What we check but do not weight

Before every preview we publish we check the schedule, the ground's latest results, data freshness, squads and injuries, the weather and the market line. Not all of that enters the number:

• Playing XI. Tested on every T20I line-up since 2024: a side stronger or weaker than its usual XI changed the average error by less than one run. Not weighted; line-ups are used for the player markets.
• Weather. There is no reliable historical weather record for these matches, so its effect cannot be measured and is not guessed. Rain-shortened innings are left out of the accuracy figures.
• Market lines. Never an input. A gap of more than about 15 runs is a reason to look for something we missed before publishing.

Changes are tested out-of-sample

Any change to the model is fitted on older matches and then measured on newer ones it has never seen. This matters: a recent revision looked like a clear improvement on the data it was built from — every team's bias fell close to zero — and then performed worse on fresh matches. It was discarded. An improvement that only shows up in-sample is not an improvement.

Where the model is weak

• Test cricket. First-innings totals in Tests are not normally distributed — a side bats until it is dismissed, so the spread is wide and asymmetric. First-innings projections are not published for Tests.
• Thin samples. New grounds, rarely used venues and debutant players all get flagged rather than quietly projected.
• In-play. Projections describe the match at the start. They are not updated ball by ball and should not be read that way.
• Variance. A central projection is not a narrow prediction. England's home first-innings totals since 2023 span 175 to 304.