What this dashboard is

This dashboard follows Liverpool's Premier League performance since 2014/15 and puts three things side by side: what happened (results), what the chances said should have happened (expected goals), and what the betting market expected before kick-off (closing odds). The gaps between those three are the interesting part: a team can win without creating much, lose while dominating, or beat a market that had priced it lower.

Two rules run through everything:

Actual and expected are never mixed. Goals, points and wins, draws and losses are whole numbers from the results. xG, xGA and xPts are always labelled "expected" and shown to one decimal. Where a page shows a gap between the two, it says which is which.
Every number is computed once, in Python, and only formatted in the browser. A large automated test suite checks the pipeline and the page against independent calculations. See How accuracy is enforced.

Data sources

Understat (Premier League, 2014/15 onwards): shot-level data with x/y coordinates, situation, shot type and xG for every shot; team match xG, xGA, PPDA and deep completions; player match statistics (minutes, goals, xA, key passes, xGChain, xGBuildup).
football-data.co.uk: results and betting odds (Pinnacle open and close, market average and maximum, Betfair Exchange where present).
Coverage: 461 Liverpool matches across 13 seasons (2014-15 to 2026-27), 12,084 shots including the opposition's, data through 20 September 2026. The current season is partial and is flagged wherever it has fewer than 10 matches.
Price used for each match: Pinnacle closing odds for 439 matches; market-average closing odds for 22 matches.

Politeness and reproducibility. Every raw response is cached, requests are limited to one per second, completed seasons are never fetched again, and the whole site is rebuilt from the cache. There are no runtime calls from your browser to any data source.

Attribution. Shot-level expected goals, player and team match data: Understat. Results and betting odds: football-data.co.uk. Charts: D3. This is an independent portfolio project, not affiliated with or endorsed by Liverpool FC, the Premier League, Understat or football-data.co.uk. FBref advanced statistics are not used (they are no longer available), and there are no defensive event statistics such as tackles or interceptions: the defensive picture comes from xGA, shots conceded, PPDA and deep completions allowed.

How the two sources are joined

Understat and football-data name clubs differently (for example "Manchester United" and "Man United"), so a mapping table translates every name. Each Liverpool match is then matched on home team, away team and date within one day (Understat times are UTC, football-data dates are local). The build fails outright if any match joins zero times or more than once, or if the two sources disagree on the score. There is no silent fallback and no warning that can be ignored.

Metric definitions

Every label, tooltip and number format on the site comes from one registry file, metrics.json, and this table is generated from it, so the definitions here cannot drift from what the charts show. "Better when" drives the colour of change arrows: for example, lower xGA is good and is coloured green.

Results (actual, whole numbers) (9)

Metric	Definition	Unit	Better when	Minimum sample	Source
Matches	Premier League matches played in the selection.	count	no judgement	10	Understat results
Wins	Matches won. Actual result, not expected.	count	higher	10	Understat results
Draws	Matches drawn. Actual result.	count	no judgement	10	Understat results
Losses	Matches lost. Actual result.	count	lower	10	Understat results
Points	Actual league points: 3 per win + 1 per draw.	count	higher	10	Understat results
Goals for	Goals scored, including opposition own goals. Actual, always a whole number.	count	higher	10	Understat results
Goals against	Goals conceded, including own goals by Liverpool players. Actual.	count	lower	10	Understat results
Goal difference	Goals for minus goals against. Actual.	count	higher	10	Understat results
Clean sheets	Matches in which Liverpool conceded no goals. Actual.	count	higher	10	Understat results

Expected values (9)

Metric	Definition	Unit	Better when	Minimum sample	Source
xG (expected)	Expected goals: the sum of the xG of every shot Liverpool took. Expected, not actual goals. Shot-level Understat model.	count	higher	10	Understat shots
xGA (expected)	Expected goals against: the sum of the xG of every shot Liverpool conceded. Lower is better.	count	lower	10	Understat shots
xG difference (expected)	xG minus xGA. Expected, not actual.	count	higher	10	Understat shots
npxG (expected)	Expected goals excluding penalties.	count	higher	10	Understat shots
xPts (xG simulation)	Expected points from the shot-level xG simulation: each shot is an independent Bernoulli trial with p = its xG, goal totals are convolved exactly and P(win), P(draw) give 3P(W) + P(D). Expected, not actual.	count	higher	10	Understat shots
Market-expected points	Expected points from de-vigged closing betting odds: 3 x P(win) + 1 x P(draw). Expected, not actual.	count	higher	10	football-data.co.uk
Goals minus xG	Actual goals scored by Liverpool players (own goals excluded) minus expected goals. Positive means finishing above expectation.	count	higher	10	Understat shots
Points minus xPts	Actual points minus xG-simulated expected points.	count	higher	10	Understat shots
Points minus market expectation	Actual points minus market-expected points. Positive means results beat the market.	count	higher	10	football-data.co.uk

Per-match rates (18)

Metric	Definition	Unit	Better when	Minimum sample	Source
Points per match	Actual points divided by matches played.	rate	higher	10	Understat results
Goals per match	Actual goals scored per match.	rate	higher	10	Understat results
Goals against per match	Actual goals conceded per match.	rate	lower	10	Understat results
xG per match (expected)	Mean expected goals for per match.	rate	higher	10	Understat shots
xGA per match (expected)	Mean expected goals against per match. Lower is better.	rate	lower	10	Understat shots
xG difference per match (expected)	Mean of xG minus xGA per match.	rate	higher	10	Understat shots
Rolling 10-match xG difference	Average xG difference (xG minus xGA) per match over the latest 10 matches, current match included. Expected, not actual. Needs at least 10 matches in the selection.	rate	higher	10	Understat shots
npxG per match (expected)	Mean non-penalty expected goals per match.	rate	higher	10	Understat shots
Shots per match	Mean shots taken per match (own goals excluded).	rate	higher	10	Understat shots
Shots conceded per match	Mean shots faced per match. Lower is better.	rate	lower	10	Understat shots
xG per shot	Mean of per-match xG divided by shots: average shot quality.	rate	higher	10	Understat shots
xG per shot (pooled)	Total xG divided by total shots over the selection: average shot quality, comparable with league clubs (own goals excluded).	rate	higher	10	Understat shots
PPDA	Passes allowed per defensive action in the opposition 40% of the pitch. Lower means a more intense press, so lower is treated as better.	rate	lower	10	Understat shots
Deep completions per match	Passes completed within roughly 20 yards of the opposition goal (excluding crosses), per match.	rate	higher	10	Understat shots
Deep completions allowed per match	Opposition deep completions per match. Lower is better.	rate	lower	10	Understat shots
Open-play xG per match	xG from open-play shots per match (excludes set pieces, corners, free kicks and penalties).	rate	higher	10	Understat shots
Open-play xGA per match	xG conceded from open-play shots per match.	rate	lower	10	Understat shots
Set-piece xGA per match	xG conceded from corners, indirect and direct free kicks and set pieces per match (penalties excluded).	rate	lower	10	Understat shots

Percentages and probabilities (6)

Metric	Definition	Unit	Better when	Minimum sample	Source
Win rate	Share of matches won.	pct	higher	10	Understat results
Clean-sheet rate	Share of matches without conceding.	pct	higher	10	Understat results
Market P(win)	Liverpool win probability from closing odds with the overround removed.	prob	higher	10	football-data.co.uk
xG-sim P(win)	Liverpool win probability from the exact xG simulation.	prob	higher	10	Understat shots
Bookmaker margin	Sum of inverse odds minus 1 on the closing 1X2 market.	pct	no judgement	10	football-data.co.uk
Closing odds (Liverpool win)	Decimal closing price on a Liverpool win: Pinnacle where available, else market average.	odds	no judgement	10	football-data.co.uk

Market lens (17)

Metric	Definition	Unit	Better when	Minimum sample	Source
Brier score (market)	Mean over matches of the sum of squared errors of the three outcome probabilities. Lower is better; a fixed 1/3 forecast scores 0.667.	prob	lower	10	football-data.co.uk
Brier score (xG sim)	Brier score of the xG-simulated probabilities.	prob	lower	10	Understat shots
Log loss (market)	Mean negative log probability assigned to the actual result. Lower is better; a fixed 1/3 forecast scores 1.099.	prob	lower	10	football-data.co.uk
Log loss (xG sim)	Log loss of the xG-simulated probabilities.	prob	lower	10	Understat shots
Brier score (external model)	Brier score of an optional external model supplied by CSV (see README).	prob	lower	10	external CSV
Log loss (external model)	Log loss of an optional external model supplied by CSV.	prob	lower	10	external CSV
Flat-stake P&L (units)	Retrospective profit of a hypothetical 1-unit stake on Liverpool to win every match at the closing odds. Analysis only, not a strategy.	count	higher	10	football-data.co.uk
Flat-stake ROI	P&L divided by units staked (one per match). Retrospective analysis, not a strategy.	pct	higher	10	football-data.co.uk
Points: actual vs expected	For each season, actual league points next to xG-simulated expected points and market-expected points. Seasons with fewer than 10 matches are left out of the chart.	count	no judgement	10	football-data.co.uk
Results vs market expectation	Running total of actual points minus market-expected points over the selection. Rising means results are beating what the closing odds implied; shaded stretches are the mispriced runs.	count	higher	10	football-data.co.uk
Mispriced runs	Stretches where 10 consecutive matches returned at least 4 points more or fewer than the closing odds implied (actual points minus market-expected points). Overlapping windows of the same sign are merged into one run.	count	no judgement	10	football-data.co.uk
Calibration scores	Brier score and log loss of the market probabilities and the xG-simulated probabilities on the selected matches, next to a naive forecast that always predicts the selection own win, draw and loss rates. The xG simulation is built from the match shots, so it describes the chances rather than forecasting them.	prob	lower	10	football-data.co.uk
Reliability plot	Forecast probability against how often the outcome happened. Every match contributes three points (win, draw, loss), grouped into probability bins. On the dashed diagonal, forecasts are perfectly calibrated. Larger markers mean more outcomes in the bin.	prob	no judgement	10	football-data.co.uk
Every match: market vs xG simulation	For each match, the closing-odds win probability and the xG-simulated win probability, with the actual result shown along the top.	prob	no judgement	10	football-data.co.uk
Flat-stake cumulative P&L	Running profit of a hypothetical 1-unit stake on Liverpool to win every match at the closing odds. Retrospective analysis, not a strategy.	count	higher	10	football-data.co.uk
Flat-stake ROI by season	Profit divided by the number of 1-unit stakes for each season. Retrospective analysis, not a strategy: a season is a small sample of about 38 bets.	pct	higher	10	football-data.co.uk
External model	Optional probabilities from your own model, supplied as a CSV with a match key and home, draw and away probabilities. When present they are scored with the same Brier score and log loss as the market and the xG simulation.	prob	no judgement	10	external CSV

Players (20)

Metric	Definition	Unit	Better when	Minimum sample	Source
Appearances	Matches in which the player played at least one minute.	count	no judgement	450	Understat shots
Minutes	Minutes played in Premier League matches.	count	no judgement	450	Understat shots
Goals	Goals scored (no own goals). Actual.	count	higher	450	Understat shots
Non-penalty goals	Goals scored excluding penalties. Actual.	count	higher	450	Understat shots
Assists	Assists. Actual.	count	higher	450	Understat shots
xG (expected)	Sum of xG of the player shots. Expected.	count	higher	450	Understat shots
npxG (expected)	Sum of non-penalty xG of the player shots. Expected.	count	higher	450	Understat shots
xA (expected)	Expected assists: the xG of shots the player passes created. Expected.	count	higher	450	Understat shots
npG minus npxG	Non-penalty goals minus non-penalty xG. Positive means finishing above expectation (Finisher role).	count	higher	450	Understat shots
xG per 90	xG per 90 minutes played.	per90	higher	450	Understat shots
npxG per 90	Non-penalty xG per 90 minutes (Shot Threat role).	per90	higher	450	Understat shots
xA per 90	Expected assists per 90 minutes (Creator role).	per90	higher	450	Understat shots
xGBuildup per 90	xG of possessions the player was involved in, excluding shots and key passes, per 90 (Build-up role).	per90	higher	450	Understat shots
xGChain per 90	xG of every possession the player was involved in, per 90 (Involvement role).	per90	higher	450	Understat shots
Shots per 90	Shots per 90 minutes.	per90	higher	450	Understat shots
Key passes per 90	Passes leading directly to a shot, per 90 minutes.	per90	higher	450	Understat shots
Squad table	Every Liverpool player with minutes in the selection. Per-90 values divide the total by minutes played and multiply by 90, so they are only shown as reliable above the minimum-minutes filter. Click a column to sort and a player to open the profile.	per90	no judgement	450	Understat shots
Role leaders	The top three players for five roles using only the data available: Finisher (npG minus npxG), Shot Threat (npxG per 90), Creator (xA per 90), Build-up (xGBuildup per 90) and Involvement (xGChain per 90), among players above the minimum-minutes filter.	per90	no judgement	450	Understat shots
Season-over-season trends	The player metric in every Premier League season. Hollow points mark seasons below the minimum minutes, where per-90 rates are unreliable.	per90	no judgement	450	Understat shots
Player comparison	Two players side by side for the current selection. Bars share one scale within each row.	per90	no judgement	450	Understat shots

Charts and maps (10)

Metric	Definition	Unit	Better when	Minimum sample	Source
Shot map	Every Liverpool shot in the selection, drawn on a 105 x 68 m pitch attacking left to right. Circle size grows with the shot xG; filled circles are goals. Own goals are not shots and are not drawn.	count	no judgement	10	Understat shots
Shots conceded map	Every shot Liverpool faced. Opposition coordinates are flipped so that Liverpool defend the left goal and attack left to right, as on the attack map. Circle size grows with xG; filled circles are goals conceded from shots.	count	no judgement	10	Understat shots
Shot volume vs shot quality	Shots per match against xG per shot (pooled) for each club in the league, with Liverpool highlighted and dashed league averages. Up and to the right means more and better chances.	rate	no judgement	10	Understat shots
Threat source mix	Share of Liverpool xG by source: open play, set pieces (corners, indirect and direct free kicks) and penalties.	pct	no judgement	10	Understat shots
Open play vs set piece xGA	xG conceded per match split into open play, set pieces (corners, indirect and direct free kicks) and penalties. Lower is better.	rate	lower	10	Understat shots
xG per match trend	Expected goals (bars) against actual goals (dots) in each match, with a rolling 10-match average of xG.	rate	higher	10	Understat shots
xGA per match trend	Expected goals conceded (bars) against actual goals conceded (dots) in each match, with a rolling 10-match average and the league average per match.	rate	lower	10	Understat shots
xG race	Cumulative expected goals for each team over the match, minute by minute. Steps are shots; filled markers are goals (actual), including own goals credited to the other team.	count	no judgement	1	Understat shots
Match shot map	Both teams shots on one 105 x 68 m pitch. Liverpool attack left to right and the opposition shots are rotated to attack the other way. Circle size grows with xG; filled circles are goals.	count	no judgement	1	Understat shots
Market vs xG-simulated probabilities	Liverpool win/draw/loss probabilities before the match from the closing betting odds with the margin removed (proportional and Shin methods), next to the exact simulation from this match shots. The simulation uses what actually happened in the match: it is a post-match view of the chances, not a forecast.	prob	no judgement	1	football-data.co.uk

Style of play (proxy scores, not good or bad) (33)

Metric	Definition	Unit	Better when	Minimum sample	Source
Style of play	Proxy scores for how Liverpool build up, attack and defend, relative to every club in the same season. Not the CIES index: CIES uses Impect event data that is not freely available. Style is neither good nor bad.	count	no judgement	10	Understat team data (proxy)
Defence style	How Liverpool defend, relative to every club in the same season: Defensive line, Pressing and Box protection, combined into the Pressure index. Proxy scores from Understat team data, not CIES scores. Style is neither good nor bad.	count	no judgement	10	Understat team data (proxy)
Build-up style	How Liverpool build up, relative to every club in the same season. Understat has no pass data, so only Build-up complexity is a (weak) proxy; length and direction need pass-level data. Style is neither good nor bad.	count	no judgement	10	Understat team data (proxy)
Attack style	How Liverpool attack, relative to every club in the same season: Attack tempo and Attack spacing (Attack medium needs shot-type data for every club), combined into the Occupation index. Proxy scores, not CIES scores. Style is neither good nor bad.	count	no judgement	10	Understat team data (proxy)
Pressure index	Mean of the available axis scores of the Defence phase (at least two axes; marked partial when an axis is unavailable). Score 0-100 relative to that season's league (50 = league average, 100 = the right-hand end). A proxy built from Understat team data, not a CIES score. Style is neither good nor bad.	score100	no judgement	10	Understat team data (proxy)
Control index	Mean of the available axis scores of the Build-up phase (at least two axes; marked partial when an axis is unavailable). Score 0-100 relative to that season's league (50 = league average, 100 = the right-hand end). A proxy built from Understat team data, not a CIES score. Style is neither good nor bad.	score100	no judgement	10	Understat team data (proxy)
Occupation index	Mean of the available axis scores of the Attack phase (at least two axes; marked partial when an axis is unavailable). Score 0-100 relative to that season's league (50 = league average, 100 = the right-hand end). A proxy built from Understat team data, not a CIES score. Style is neither good nor bad.	score100	no judgement	10	Understat team data (proxy)
Defensive line: Low ◄► High	How high the team defends. Toward High: few deep completions allowed and a larger share of the shots conceded coming from fast attacks (a high line is exposed to quick breaks). A proxy from Understat team data. Score 0-100 relative to that season's league (50 = league average, 100 = the right-hand end). A proxy built from Understat team data, not a CIES score. Style is neither good nor bad.	score100	no judgement	10	Understat team data (proxy)
Pressing: Passive ◄► Active	How actively the team wins the ball high. Toward Active: lower PPDA (passes allowed per defensive action) and more defensive actions in the opposition's 40%. The two KPIs share a source and are correlated. Score 0-100 relative to that season's league (50 = league average, 100 = the right-hand end). A proxy built from Understat team data, not a CIES score. Style is neither good nor bad.	score100	no judgement	10	Understat team data (proxy)
Box protection: Lenient ◄► Tight	How little the team lets in near goal. Toward Tight: fewer shots and less xG conceded from inside the box. The central-rectangle KPI is not computed (needs league-wide shot coordinates). Score 0-100 relative to that season's league (50 = league average, 100 = the right-hand end). A proxy built from Understat team data, not a CIES score. Style is neither good nor bad.	score100	no judgement	10	Understat team data (proxy)
Build-up length: Long ◄► Short	How long or short the team's passes are when building from the back. Score 0-100 relative to that season's league (50 = league average, 100 = the right-hand end). A proxy built from Understat team data, not a CIES score. Style is neither good nor bad. Pass length needs pass-level event data, which Understat does not provide. No honest proxy exists in shot data.	score100	no judgement	10	needs pass-level data
Build-up direction: Vertical ◄► Horizontal	Whether the team builds forward quickly or circulates the ball sideways. Score 0-100 relative to that season's league (50 = league average, 100 = the right-hand end). A proxy built from Understat team data, not a CIES score. Style is neither good nor bad. Pass direction needs pass-level event data, which Understat does not provide. No honest proxy exists in shot data.	score100	no judgement	10	needs pass-level data
Build-up complexity: Simple ◄► Elaborate	How elaborate the team's possession is. Toward Elaborate: a larger share of shots from slow attacks and a larger share of xGChain that comes from build-up (xGBuildup). A weak proxy: there is no pass data. Score 0-100 relative to that season's league (50 = league average, 100 = the right-hand end). A proxy built from Understat team data, not a CIES score. Style is neither good nor bad.	score100	no judgement	10	Understat team data (proxy)
Attack medium: Aerial ◄► Grounded	Whether chances come through the air (headers, crosses) or on the ground. Score 0-100 relative to that season's league (50 = league average, 100 = the right-hand end). A proxy built from Understat team data, not a CIES score. Style is neither good nor bad. Needs shot type and last action for every club's shots (headers, crosses). Understat only gives those per match or per player, about 4,560 requests league-wide. Supply att_header_share and att_cross_share in data/external/style_kpis.csv to enable it.	score100	no judgement	10	needs pass-level data
Attack tempo: Rapid ◄► Placed	Whether attacks are quick or patient. Toward Placed: a smaller share of shots and xG from fast attacks (Understat's own attack-speed classification). Score 0-100 relative to that season's league (50 = league average, 100 = the right-hand end). A proxy built from Understat team data, not a CIES score. Style is neither good nor bad.	score100	no judgement	10	Understat team data (proxy)
Attack spacing: Scattered ◄► Grouped	How concentrated the team's attacking is. Toward Grouped: a larger share of shots from inside the box and more deep completions per shot. Central-lane share and shot dispersion are not computed (need league-wide coordinates). Score 0-100 relative to that season's league (50 = league average, 100 = the right-hand end). A proxy built from Understat team data, not a CIES score. Style is neither good nor bad.	score100	no judgement	10	Understat team data (proxy)
Deep completions allowed per match	Deep completions allowed per match. A higher value pushes toward the left-hand label of its style axis. Source: Understat team history (deep_allowed).	rate	no judgement	10	Understat team history (deep_allowed)
Share of shots conceded from fast attacks	Share of shots conceded from fast attacks. A higher value pushes toward the right-hand label of its style axis. Source: Understat team page (attackSpeed, against).	pct	no judgement	10	Understat team page (attackSpeed, against)
PPDA (passes allowed per defensive action)	PPDA (passes allowed per defensive action). A higher value pushes toward the left-hand label of its style axis. Source: Understat team history (ppda).	rate	no judgement	10	Understat team history (ppda)
Defensive actions in the opposition 40% per match	Defensive actions in the opposition 40% per match. A higher value pushes toward the right-hand label of its style axis. Source: Understat team history (ppda.def).	rate	no judgement	10	Understat team history (ppda.def)
In-box shots conceded per match	In-box shots conceded per match. A higher value pushes toward the left-hand label of its style axis. Source: Understat team page (shotZone, against).	rate	no judgement	10	Understat team page (shotZone, against)
xGA from in-box shots per match	xGA from in-box shots per match. A higher value pushes toward the left-hand label of its style axis. Source: Understat team page (shotZone, against).	rate	no judgement	10	Understat team page (shotZone, against)
Shots conceded from the central rectangle per match	Shots conceded from the central rectangle per match. A higher value pushes toward the left-hand label of its style axis. Source: external only: league-wide shot coordinates.	rate	no judgement	10	external only: league-wide shot coordinates
Share of shots from slow attacks	Share of shots from slow attacks. A higher value pushes toward the right-hand label of its style axis. Source: Understat team page (attackSpeed).	pct	no judgement	10	Understat team page (attackSpeed)
xGBuildup as a share of xGChain	xGBuildup as a share of xGChain. A higher value pushes toward the right-hand label of its style axis. Source: Understat league player table (club totals).	pct	no judgement	10	Understat league player table (club totals)
Share of shots that are headers	Share of shots that are headers. A higher value pushes toward the left-hand label of its style axis. Source: external only: league-wide shot type.	pct	no judgement	10	external only: league-wide shot type
Share of open-play shots from a cross	Share of open-play shots from a cross. A higher value pushes toward the left-hand label of its style axis. Source: external only: league-wide shot last action.	pct	no judgement	10	external only: league-wide shot last action
Share of shots from fast attacks	Share of shots from fast attacks. A higher value pushes toward the left-hand label of its style axis. Source: Understat team page (attackSpeed).	pct	no judgement	10	Understat team page (attackSpeed)
Share of xG from fast attacks	Share of xG from fast attacks. A higher value pushes toward the left-hand label of its style axis. Source: Understat team page (attackSpeed).	pct	no judgement	10	Understat team page (attackSpeed)
Share of shots taken inside the box	Share of shots taken inside the box. A higher value pushes toward the right-hand label of its style axis. Source: Understat team page (shotZone).	pct	no judgement	10	Understat team page (shotZone)
Deep completions per shot	Deep completions per shot. A higher value pushes toward the right-hand label of its style axis. Source: Understat team history (deep) and shots.	rate	no judgement	10	Understat team history (deep) and shots
Share of open-play shots from the central lane	Share of open-play shots from the central lane. A higher value pushes toward the right-hand label of its style axis. Source: external only: league-wide shot coordinates.	pct	no judgement	10	external only: league-wide shot coordinates
Horizontal dispersion of open-play shots	Horizontal dispersion of open-play shots. A higher value pushes toward the left-hand label of its style axis. Source: external only: league-wide shot coordinates.	rate	no judgement	10	external only: league-wide shot coordinates

Baselines and change

The vs-baseline panels compare the selection with a baseline: the mean of per-match rates over the previous two Premier League seasons. The panel always names the baseline it used. Rules that keep the arithmetic honest:

The first season in the data (2014/15) has no baseline, and there is no baseline for a multi-season selection. Both show N/A, never "0%". Season 2015/16 has only one earlier season, so it uses that one and says so.
For counts, rates and per-90 values the change is a percentage: (current − baseline) ÷ baseline. It is N/A if the baseline is missing, zero or negative.
For percentages and probabilities the change is shown in percentage points, never a percentage of a percentage.
Metrics that can be negative (goal difference, xG difference, "minus xG" gaps) show the plain difference instead of a percentage. A percentage of a negative or near-zero baseline produces impossible figures such as "−117%".
The arrow follows the sign of the change. The colour follows whether higher is better: xGA falling is a green down-arrow.
If a selection has fewer than 10 matches (or a player fewer than 450 minutes) a small sample badge appears and takeaways that would need statistical significance are suppressed.

Expected goals and the exact simulation

Expected goals (xG) come from Understat's shot-level model: each shot gets the probability that it becomes a goal, based on its location, the type of chance and how it was created. The site never re-estimates xG; it uses Understat's value for every shot and adds them up. Own goals are not shots and carry no xG.

xG-simulated result probabilities. To turn the chances in a match into win, draw and loss probabilities, each shot is treated as an independent coin flip that lands with probability equal to its xG (penalties are ordinary shots with their own xG). Each team's goal count is then a Poisson-binomial distribution, and it is computed exactly by convolution, not by Monte Carlo sampling: start with the distribution [1] and, for each shot with probability p, combine [1 − p, p] into it. The two teams' distributions are multiplied together to give the probability of every scoreline, and summed to P(win), P(draw) and P(loss). Simulated xPts = 3 × P(win) + P(draw).

Worked example. {esc(cc["name"])} have two shots (xG 0.30 and 0.50) and the opposition one (xG 0.20). The probability of each goal count is:

Goals	{esc(cc["name"])}	Opposition
0	35.0%	80.0%
1	50.0%	20.0%
2	15.0%	0.0%

Combining the two gives P(win) = 55.0%, P(draw) = 38.0%, P(loss) = 7.0%, so xPts = 3 × 0.5500 + 0.3800 = 2.03.

The simulation uses the shots that were actually taken, so it describes how good the chances were, after the fact. It is not a pre-match forecast, and that is why the calibration section treats it as a benchmark rather than a rival model. As a cross-check, season totals of simulated xPts sit within 0.7 points of Understat's own xPts in every season (the test allows 2.0).

Turning odds into probabilities (de-vig)

Betting odds contain a margin (the "vig" or overround), so the implied probabilities 1 ÷ odds add up to more than 100%. The market lens removes it with the closing price (the most informed price before kick-off). Two methods are implemented and both are shown on the match page:

Proportional (the default): divide each implied probability by their sum, p_i = (1 ÷ o_i) ÷ Σ(1 ÷ o_j). Simple, transparent, and spreads the margin evenly.
Shin (1993): assumes a share z of bettors are insiders and solves for the z that makes the implied probabilities sum to exactly 1. It takes proportionally more margin out of long shots than favourites, correcting the favourite-longshot bias. On low-margin books the two agree closely.

Worked example: Arsenal v Liverpool, 8 Jan 2026 (Pinnacle closing odds). The implied probabilities add up to 103.01%, a margin of 3.01%.

Outcome	Decimal odds	Implied (1 ÷ odds)	Proportional	Shin
Arsenal win	1.59	62.89%	61.05%	61.68%
Draw	4.40	22.73%	22.06%	21.81%
Liverpool win	5.75	17.39%	16.88%	16.51%
Sum		103.01%	100.00%	100.00%

Shin insider share z = 0.0151.

Price used. Pinnacle's closing odds are the primary price. Where they are missing (as they are for part of 2025/26 and all of 2026/27 in the source), the market-average closing price is used instead, and the match page states which price it used. Market-expected points = 3 × P(win) + P(draw), using the de-vigged closing probabilities.

Market lens: how the comparisons work

Points: actual vs xPts vs market. For each season: actual points, points expected from the chances (xG simulation) and points expected by the market. Differences are labelled "minus" gaps and coloured by sign.
Calibration. Brier score is the mean over matches of the sum of squared errors of the three outcome probabilities (lower is better; always guessing one third scores 0.667). Log loss is the mean of −ln(probability given to what actually happened) (lower is better; one third scores 1.099). The reliability plot pools every (match, outcome) pair and compares forecast probability with how often the outcome occurred. A naive benchmark that always predicts the selection's own win, draw and loss rates is shown for scale. Because the xG simulation sees the match's shots, expect it to score better than the market: that is what "results follow chances" looks like, not a claim that it beats the market.
Mispriced runs. Every window of 10 consecutive matches whose actual points differ from market-expected points by at least 4.0 is flagged; overlapping windows of the same sign are merged into one run ("beat" or "lagged" the market).
Flat 1-unit stake. A hypothetical 1-unit bet on a Liverpool win in every match at the closing odds: profit is odds − 1 on a win and −1 otherwise, and ROI is profit divided by the number of stakes. This is retrospective analysis of a public price series, not a strategy and not betting advice. A season is only about 38 bets, so ROI swings widely on luck.

Plugging in your own model

To score an external model next to the market and the xG simulation, save a CSV at data/external/model_probs.csv and rebuild. Columns: match_id (the Understat id) or date (yyyy-mm-dd, the local match date) with home_team and away_team (Understat club names), plus p_home, p_draw and p_away. Each probability must be in [0, 1] and each row must sum to 1 (within 0.001); anything else stops the build. Matches without a row are simply not scored. The Market Lens page then adds the model to the calibration table and reliability plot.

Style of play (proxy scores)

The Attack and Defence pages show how Liverpool play, not how well. The idea comes from the CIES Football Observatory's "style of play of teams" work: three phases of the game (defence, build-up, attack), three tactical options in each, every option scored from 0 to 100 against the other teams of the league, and a composite index per phase. What is borrowed is that framework: the nine axes and their end labels, the 0 to 100 league-relative scale, and the idea of a composite index per phase. What is changed is the data and so the measurement: CIES scores come from Impect event data (pressure locations, pass lengths, reception heights, possession phases), which is not freely available. Every score here is a proxy built from Understat team data and is never a CIES score or the CIES index.

Style is neither good nor bad. High pressing is not better than a low block, so none of these components use the green and orange change colours or arrows. Liverpool are drawn in the club colour, other clubs in grey.

How a score is computed, for each season and each axis separately, over the clubs of that season's Premier League:

Each KPI is turned into a z-score: z = (value − league mean) ÷ league standard deviation (population standard deviation, over that season's 20 clubs). Per-match rates are used, so a part-played season compares like with like.
The axis value is the mean of sign × z over its KPIs (sign +1 pushes toward the right-hand label, −1 toward the left). An axis needs at least two KPIs; otherwise it is not scored and shown as N/A, never 0.
Averaging shrinks the spread, so the axis values are standardised again to unit variance across the league.
Score = 100 × Φ(z), where Φ is the normal distribution function: 50 is the league average and 100 is the right-hand end. The rank is among the clubs, 1st being furthest toward the right-hand label; clubs with identical values share a rank and the next rank is skipped.
The index of a phase is the mean of its available axis scores. It needs at least two axes and is labelled partial when an axis is missing. Its change is the difference of the two displayed whole-number scores of consecutive seasons.

Why scores are relative to each season's league. The scale is reset every season, so a club can change its score without changing what it does, because the other 19 clubs moved. A movement between seasons is a change in relative style, not absolute style. Scores cover whole seasons: the league data cannot be split by manager inside a season, so the manager eras shown in the trend lines are bands, not separate scores.

Phase	Axis	KPIs (Understat team data)	Status	Confidence
Defence	Low ◄► High	Deep completions allowed per match (lower pushes toward High) Share of shots conceded from fast attacks (higher pushes toward High)	Proxy	low
Defence	Passive ◄► Active	PPDA (passes allowed per defensive action) (lower pushes toward Active) Defensive actions in the opposition 40% per match (higher pushes toward Active)	Proxy	medium
Defence	Lenient ◄► Tight	In-box shots conceded per match (lower pushes toward Tight) xGA from in-box shots per match (lower pushes toward Tight) Not computed (needs shot-level data for every club): Shots conceded from the central rectangle per match	Proxy	medium
Build-up	Long ◄► Short	none	Unavailable: needs pass-level data	none
Build-up	Vertical ◄► Horizontal	none	Unavailable: needs pass-level data	none
Build-up	Simple ◄► Elaborate	Share of shots from slow attacks (higher pushes toward Elaborate) xGBuildup as a share of xGChain (higher pushes toward Elaborate)	Proxy	low
Attack	Aerial ◄► Grounded	none Not computed (needs shot-level data for every club): Share of shots that are headers; Share of open-play shots from a cross	Unavailable: needs pass-level data	none
Attack	Rapid ◄► Placed	Share of shots from fast attacks (lower pushes toward Placed) Share of xG from fast attacks (lower pushes toward Placed)	Proxy	medium
Attack	Scattered ◄► Grouped	Share of shots taken inside the box (higher pushes toward Grouped) Deep completions per shot (higher pushes toward Grouped) Not computed (needs shot-level data for every club): Share of open-play shots from the central lane; Horizontal dispersion of open-play shots	Proxy	low

The central rectangle. A shot counts as central when, from the shooter's view, x ≥ 0.843 (the 16.5 m line of a 105 m pitch) and the shot lies within 0.13 of the middle line (|y − 0.5| ≤ 0.13). This is an assumption, configurable in config/style.json. It is defined but not computed for any club, because it needs the shot coordinates of every match of every club (about 4,560 requests), which is more than this site fetches. Such KPIs are declared as external-only and can be supplied with a CSV of season,club,kpi_id,value at data/external/style_kpis.csv; axes that use them then show External data instead of Proxy.

Confidence. Each axis carries a confidence from its KPIs: medium for pressing, box protection and attack tempo, which are measured fairly directly by PPDA, in-box shot counts and Understat's attack-speed classes; low for the defensive line, build-up complexity and attack spacing, which are weak indirect proxies. Build-up length and direction are unavailable: they need pass-level event data and there is no honest proxy in shot data. Attack medium (aerial or grounded) needs shot type and last action for every club. Where an axis cannot be scored the page says Needs pass-level data and draws no track. The Control index therefore has one scorable axis and cannot be built, and the Occupation index is partial. Because the build-up phase has almost nothing to show without pass data, the Attack page shows only the Attack phase and the Defence page only the Defence phase; the build-up axes stay defined in the configuration and appear in the table above so they can light up if pass-level data is supplied.

Takeaways

The takeaway sentences are generated by fixed rules with explicit thresholds; nothing is written by hand and every number in a sentence is computed from the data and re-checked by tests. If no rule fires, nothing is shown, and rules that need a meaningful sample switch off for small samples.

Rule	Fires when	Needs
Form	points from the last 5 matches are 10 or more (strong) or 3 or fewer (poor). Between those, nothing is said. A positive claim needs at least 7 points.	any sample
Finishing	goals scored by players minus xG is at least 3.0 goals in either direction.	10+ matches
Results vs market	actual points differ from market-expected points by at least 3.0.	10+ matches
Points vs xPts	actual points differ from xG-simulated points by at least 3.0.	10+ matches
Versus baseline	xG or xGA per match differs from the baseline by at least 10%.	10+ matches and a baseline
Best since / worst on record	a season's xGA per match is the lowest (or xG the highest) since an earlier season, or its xGA is the highest in the data.	3+ earlier seasons

Known limitations and data quirks

Three different xG totals exist for a season in Understat's own data, and the site says which it uses. Team match totals, the sum of match shots and team season statistics can differ: team season statistics count opponent own goals as 1.0-xG "shots" (removed here), and Understat's team match xG is lower than the sum of that match's shots in 17% of team-matches, by up to 0.9 xG. This site uses the sum of shots everywhere so that every map, chart and total reconciles. League-wide reference lines for other clubs come from team-level figures with the own-goal artefact removed.
Own goals are recorded on the side of the player who scored them and credited to the opposition; goals in the record include them, player goals never do.
xG is a model. Understat's model does not know about goalkeepers, defenders' positions or player quality beyond the shot itself, so a finisher who beats xG for years is not necessarily lucky.
PPDA and deep completions are Understat definitions. PPDA is a ratio of the opponent's passes to Liverpool's defensive actions in the attacking 40% of the pitch, so it varies widely from match to match; trend lines use a rolling 10-match average.
The market is one price series. When Pinnacle's price is missing, a different bookmaker composite is used, and its margin is typically larger.
Small samples. A single season is about 38 matches. The site flags samples under 10 matches and players under 450 minutes rather than hiding them.

How accuracy is enforced

The build and the deployed site are guarded by automated tests that must pass before anything is published; a failing run leaves the last good version live.

Pipeline tests: every completed season has exactly 38 matches; every match joins once to odds and the scores agree; every club name maps; no duplicate ids; goals from shots plus own goals equal the final score in every match; player goals plus own goals equal team goals in every season; roster totals equal Understat's own season table.
Maths tests: hand-computed fixtures; the exact xG convolution equals brute-force enumeration; de-vigged probabilities lie in (0, 1) and sum to 1; Shin and proportional agree on low-margin books; the change function returns N/A for missing baselines, points for probabilities, and the right colour for lower-is-better metrics; a regression test locks in that 13.3 against 12.2 is +9.0%.
Page tests (a real browser): every page renders for every season with no console errors; after switching season no other season label appears on screen; displayed records equal the build data; every match page is checked against an independent recomputation of its xG, market probabilities and simulated probabilities; shot positions are compared circle by circle against the source coordinates.