Criclysis

Method

How the numbers are made

Every figure on Criclysis is computed from Cricsheet's ball-by-ball records. This page explains where the data comes from, how often it changes, how strengths and weaknesses are found, and how well the win-probability model does on matches it never saw. Every number below is read from the current build.

Data: Cricsheet (CC BY 4.0), matches up to .

Data and licence

Every delivery comes from Cricsheet: the batter, bowler, runs, extras and dismissal for every ball of every international in its archive. Cricsheet publishes it under the Creative Commons Attribution 4.0 licence (CC BY 4.0), which is why every data view on this site credits it.

Ball-by-ball coverage does not reach back to the start of international cricket. A career that began before coverage starts is counted only from that point, so some totals here are partial and will not match a player's full career record.

Players are keyed by Cricsheet's person id, never by name, so two cricketers with the same name stay separate. In this build, 68 names belong to more than one person; bowling styles are not matched by name for 1 players with such names, so they appear without a style rather than with a guessed one.

Photo credits. Player photos are not Cricsheet data. They come from Wikimedia Commons, matched by id: Cricsheet's people register gives each player's ESPNcricinfo id, and Wikidata links that id to the player's image. Only images released under CC0, into the public domain, or under CC BY, CC BY-SA or the Government Open Data Licence – India are used, and each is shown with its author and licence, linked to the file's page on Commons. An id that leads to more than one Wikidata entry gets no photo rather than a guessed one, and a player without a usable image is shown with their initials.

Coverage in this build

Coverage by format
FormatMatchesPlayersDeliveriesFirstLast
Men's Test8991,0711,730,52319 Dec 200128 Oct 2026
Men's ODI2,6172,0181,388,57527 Jun 20022 Nov 2026
Men's T20I3,6494,814820,25717 Feb 20051 Nov 2026
Women's Test2415446,75122 Feb 200310 Jul 2026
Women's ODI620703324,69822 Jan 200728 Oct 2026
Women's T20I2,2002,795499,82918 Jun 200925 Oct 2026

Sources

  • Cricsheet

    Ball-by-ball match data (primary source for every split on this site)

    Licence: CC BY 4.0

  • Player metadata table

    Bowling style, batting hand, role and country for the matchup splits

    Licence: See docs/DATA_SOURCES.md (not Cricsheet data)

Daily refresh and what “live” means

Criclysis has no live-score API, uses no API keys and scrapes no websites. “Live” means two things:

  1. A daily refresh. A scheduled GitHub Action downloads Cricsheet's public daily files, rebuilds every aggregate and the win-probability models, and updates the database in small, hash-checked steps. A finished international usually appears within about a day. If any step fails, the site keeps serving the previous build.
  2. Replays. Any match with ball-by-ball data can be played back one delivery at a time, with the scoreboard, win probability, projected and par scores, and each batter's record against the bowling. Replays are real matches; nothing is simulated.

Player context in a replay (strengths, weaknesses and matchups) uses career figures as of the build date, so for older matches it includes what happened afterwards.

Data runs to
the latest match in this build
Built
build 2026-11-03T18:18:14+00:00-37e0590
Imported
into the database

Strengths and weaknesses

A raw split is not a finding. A batter's average against one bowling type means little until you know what they average overall, and how other players of the same standing fare against the same bowling.

  1. Split every ball

    Each player's balls are split by bowling type and family, phase of the innings, time at the crease, opposition, venue, home or away, chasing or setting, batting position, and the individual bowlers or batters faced. Batting splits are shown from 60 balls faced and bowling splits from 90 balls bowled.

  2. Choose a comparable cohort

    Players are compared only with players of the same gender in the same format, never across men's and women's cricket. The cohort is everyone with at least 600 balls faced (batting) or 900 balls bowled (bowling).

  3. Shrink small samples

    Each split is pulled towards the player's own overall figure in proportion to how few balls it rests on, so a short spell against one type of bowling cannot produce an extreme claim.

  4. Rank the change, not the raw number

    The ranking is of the change from the player's own baseline, against how much the rest of the cohort changes in the same split. That asks the useful question: relative to how good this player is, are they better or worse against this?

  5. Apply the thresholds

    A claim needs 150 balls faced or 240 bowled, a clear change from the player's baseline, and a rank at or above the 70th percentile (strength) or at or below the 30th (weakness). Too-small cohorts produce no claim at all.

  6. Show the evidence

    Every claim carries its value, the player's baseline, the cohort median, the percentile and the balls behind it. Confidence rises with the sample relative to the claim threshold.

Sample thresholds in this build

Sample thresholds
RuleValueUnitWhat it does
Batting split shown60balls facedA batting split (against a bowling type, in a phase, at a venue) is shown from this sample; below it the page says “not enough balls”.
Bowling split shown90balls bowledThe same rule for bowling splits.
Batting claim150balls facedA split needs this sample before it can become a strength or weakness.
Bowling claim240balls bowledThe same rule for bowling claims.
Batting cohort600balls facedPlayers with at least this many balls in the format form the batting comparison group.
Bowling cohort900balls bowledThe same rule for the bowling comparison group.
Strength70percentile or aboveWhere a split must rank among comparable players to be called a strength.
Weakness30percentile or belowWhere a split must rank to be called a weakness.
Team win %5decided matchesA team's win percentage is shown from this many won, lost, tied or drawn matches.
Venue average3inningsA ground's first-innings average is shown from this many innings.

Cohort sizes

Cohort sizes by format
FormatBattersBowlersMedian batting avgMedian economy
Men's Test46031529.773.23
Men's ODI48340331.085.10
Men's T20I33818026.367.20
Women's Test231336.882.53
Women's ODI13210428.364.62
Women's T20I19712720.305.84

The players who meet the cohort thresholds above, per format; medians are of those players' career figures. A dash means no cohort for that discipline in this build.

Win probability, projected score and par

The model is learned from Cricsheet history; nothing in it is tuned by hand. It knows only the state of the match (runs, balls and wickets), not team strength, conditions or the players at the crease.

  1. Resources. For a side with a given number of legal balls left and wickets down, how many more runs do teams usually score? A Duckworth-Lewis-style curve is fitted for each wicket count from every completed first innings. It gives the projected score, and the spread of real outcomes around it gives the range shown with the projection.
  2. The chase. A logistic regression on how much of the remaining resources the target still demands (the log of runs needed over expected runs), with terms for the share of the innings left and wickets lost.
  3. The first innings. There is no target yet, so the model takes the spread of possible final totals from the resources curve and asks the chase model how often each would be defended.
  4. Par is the first-innings total that gives the side batting first an even chance. A ground's par shifts by how much more or less teams score there, shrunk towards the format par where a ground has few matches.

Win probability, projected score and par are shown only for ODIs and T20Is played over their full scheduled overs and not decided by a rain rule; otherwise the page says why they are missing.

Validation on matches the model never saw

Each model is trained on matches before 1 Jan 2023 and scored on the later matches: every ball state of each chase, and every sixth ball state of each first innings. The Brier score is the mean squared error of the probability, so lower is better. The baseline is a constant forecast: how often the side batting first, or chasing, won in the training matches. Skill is 1 − Brier ÷ baseline. The published model is then refitted on every match, including the holdout.

Ball states within one match are strongly related, so the effective sample is closer to the number of holdout matches than to the number of states. First-innings skill is lower for a real reason: at the halfway point nobody knows how the chase will go.

Win-probability validation by format
ModelHoldout matchesChase BrierBaselineSkill1st-inns BrierBaselineSkillPar
Men's ODIfrom 1 Jan 20234320.1350.25046%0.2210.25112%258
Men's T20Ifrom 1 Jan 20231,8630.1070.25057%0.2010.25020%150
Women's ODIfrom 1 Jan 20232160.1110.25456%0.2020.25320%233
Women's T20Ifrom 1 Jan 20231,3170.1040.24858%0.1990.24920%114

Men's ODI

432 holdout matches from 1 Jan 2023 · trained on 1,790

Chase Brier
0.135vs 0.250 baseline (skill 46%) · 104,131 ball states
First-innings Brier
0.221vs 0.251 baseline (skill 12%) · 20,278 ball states (every 6th)
Par
258runsformat par; model fitted on 2,222 matches
Recency half-life
12yearshow fast older matches lose weight

Calibration on the holdout

Across: the model's predicted chance, in 10% bins. Up: how often that side actually won. On the dashed line, predictions match outcomes.

View as table
Men's ODI calibration by predicted-probability bin
Predicted binChaseFirst innings
StatesPredictedWonStatesPredictedWon
0-10%22,0904%6%6686%10%
10-20%11,29715%21%1,09715%17%
20-30%7,89425%28%1,85225%31%
30-40%7,17535%44%2,80735%40%
40-50%6,75445%50%4,28146%50%
50-60%6,21255%54%3,78755%57%
60-70%5,77665%61%2,79265%63%
70-80%5,26475%72%1,70075%71%
80-90%6,83785%83%1,02284%82%
90-100%24,83298%95%27292%88%

Men's T20I

1,863 holdout matches from 1 Jan 2023 · trained on 1,469

Chase Brier
0.107vs 0.250 baseline (skill 57%) · 185,488 ball states
First-innings Brier
0.201vs 0.250 baseline (skill 20%) · 36,251 ball states (every 6th)
Par
150runsformat par; model fitted on 3,332 matches
Recency half-life
12yearshow fast older matches lose weight

Calibration on the holdout

Across: the model's predicted chance, in 10% bins. Up: how often that side actually won. On the dashed line, predictions match outcomes.

View as table
Men's T20I calibration by predicted-probability bin
Predicted binChaseFirst innings
StatesPredictedWonStatesPredictedWon
0-10%47,7893%2%1,6615%3%
10-20%17,59915%14%2,31315%16%
20-30%14,36025%23%2,91325%25%
30-40%12,13635%33%4,38835%31%
40-50%9,75945%44%7,94746%46%
50-60%9,45355%54%7,24955%57%
60-70%9,43265%65%4,54065%72%
70-80%10,31475%75%3,06075%77%
80-90%12,25285%82%1,60384%86%
90-100%42,39498%98%57793%96%

Women's ODI

216 holdout matches from 1 Jan 2023 · trained on 317

Chase Brier
0.111vs 0.254 baseline (skill 56%) · 50,127 ball states
First-innings Brier
0.202vs 0.253 baseline (skill 20%) · 10,050 ball states (every 6th)
Par
233runsformat par; model fitted on 533 matches
Recency half-life
5yearshow fast older matches lose weight

Calibration on the holdout

Across: the model's predicted chance, in 10% bins. Up: how often that side actually won. On the dashed line, predictions match outcomes.

View as table
Women's ODI calibration by predicted-probability bin
Predicted binChaseFirst innings
StatesPredictedWonStatesPredictedWon
0-10%14,7034%4%3375%5%
10-20%5,17615%17%46016%15%
20-30%3,51825%37%76225%25%
30-40%2,97435%46%1,06736%35%
40-50%2,71145%46%2,03446%44%
50-60%2,82555%56%2,14755%53%
60-70%2,52165%63%1,39465%69%
70-80%2,28375%73%95875%78%
80-90%2,57685%95%70685%90%
90-100%10,84098%99%18593%92%

Women's T20I

1,317 holdout matches from 1 Jan 2023 · trained on 718

Chase Brier
0.104vs 0.248 baseline (skill 58%) · 129,417 ball states
First-innings Brier
0.199vs 0.249 baseline (skill 20%) · 25,748 ball states (every 6th)
Par
114runsformat par; model fitted on 2,035 matches
Recency half-life
Noneevery match weighted equally; the search found no gain from favouring recent ones

Calibration on the holdout

Across: the model's predicted chance, in 10% bins. Up: how often that side actually won. On the dashed line, predictions match outcomes.

View as table
Women's T20I calibration by predicted-probability bin
Predicted binChaseFirst innings
StatesPredictedWonStatesPredictedWon
0-10%41,4023%2%1,0805%4%
10-20%11,47015%15%1,33815%18%
20-30%8,51825%26%1,78925%24%
30-40%6,80135%35%2,43735%36%
40-50%6,50745%39%3,75245%45%
50-60%6,04655%51%6,75954%55%
60-70%5,56965%61%3,70465%67%
70-80%6,34575%65%2,46175%78%
80-90%7,93385%79%1,68685%91%
90-100%28,82698%96%74293%98%

Questions about a number? Every match links to its full scorecard, built from the same records.