The Premier League table tells you where everyone finished. An expected-goal-difference table tells you whether they deserved to, but it needs shot-by-shot xG, and the free public xG tables are no longer something we can pull. Shot difference and shots-on-target difference need nothing but box scores. We built both for every club in 2025-26 from the 380 match box scores this site serves, then tested them: after 19 games, the shots-on-target table predicted each club's last 19 games better than the points table did.

A shot-difference league table for the 2025-26 Premier League, and how well it predicts

Expected-goal difference is xG for minus xG against, where each shot is weighted by how likely it was to go in. Without shot locations we cannot weight shots, so we count them: shot difference is a club's shots minus the shots it faced, and shots-on-target (SOT) difference is the same for shots on target. A shot on target is a crude quality filter (a shot that never threatened the goal is dropped), which is why SOT difference is the closer stand-in for xG difference of the two.

The data is ESPN's public box score for all 380 matches of 2025-26, pulled on 7 September 2026 and served at data_layer/epl_2025_26_boxscores.json. Over the full season, points correlated with goal difference at 0.97, with SOT difference at 0.91 and with shot difference at 0.84. Goal difference is closest because goals make points; the question a shot table answers is which clubs' goals and points ran ahead of, or behind, their shots. The last column below does that job, the job expected points did in the old version of this page: actual points minus the points a straight line through the 20 clubs gives for that SOT difference (51.8 points, plus 0.31 for every shot on target of difference).

Premier League 2025-26 by final position, with season shot and shots-on-target (SOT) difference. SOT rank 1 = best SOT difference; tied clubs share a rank. Last column: points minus the points the season-wide line gives for that SOT difference. Data: ESPN public box scores, pulled 7 September 2026.
Pos Club Pts GD Shot diff SOT diff SOT rank Pts vs SOT line
1Arsenal85+44+243+971+3.3
2Manchester City78+42+225+782+2.2
3Manchester United71+19+155+753−3.9
4Aston Villa65+7−7+39+12.3
5Liverpool60+10+154+225+1.4
6AFC Bournemouth57+4+44+156+0.6
7Sunderland54−6−144−3917+14.2
8Brighton & Hove Albion53+6+53+284−7.4
9Brentford53+3−100−1614+6.1
10Chelsea52+6+94+68−1.6
11Fulham52−4+36−912+3.0
12Newcastle United49−2+22+87−5.3
13Everton49−3−110−1614+2.1
14Leeds United47−7+1−1213−1.1
15Crystal Palace45−10−3−2016−0.6
16Nottingham Forest44−3−30−711−5.6
17Tottenham Hotspur41−9−31−610−9.0
18West Ham United39−19−162−5318+3.5
19Burnley22−37−281−9820+0.4
20Wolverhampton Wanderers20−41−159−5619−14.5

Two clubs stand out. Sunderland finished seventh on 54 points with the 17th-best SOT difference (−39), 14.2 points above the line; Aston Villa finished fourth with a SOT difference of +3, ninth-best, 12.3 points above it. At the other end, Wolves finished bottom, 14.5 points below what a −56 SOT difference usually brings, and Tottenham's 17th place sat 9.0 points under a SOT difference that ranked tenth.

Does the shot table predict? Reading a table like this as a forecast only makes sense if it beats the plain table at forecasting. So we split each club's 38 matches in date order into its first 19 (which ended on 30 December 2025 or 1 January 2026) and its last 19, built four tables from the first half, and asked which best predicted points per game in the second half. For each we report the correlation with second-half points per game, the rank correlation (do the orderings agree), and the typical miss of a straight-line projection, fitted on all 20 clubs and again leaving each club out of its own fit.

Premier League 2025-26: each club's first 19 matches (per game) against its points per game in the last 19. Typical miss = mean absolute error of a straight-line projection, in points per game. Leave-one-out fits the line on the other 19 clubs.
Table after 19 games Correlation Rank correlation Typical miss, fitted Typical miss, leave-one-out
Points0.600.460.2700.300
Goal difference0.570.410.2570.285
Shot difference0.710.680.2420.272
SOT difference0.770.690.1940.217
Every club at the league average––0.2890.289
Every club repeats its first-half points per game––0.3630.363

The shots-on-target table won on every measure. Its leave-one-out miss of 0.217 points per game is about 4.1 points over 19 games, against 5.7 for the points table. The points table and the goal-difference table, projected honestly, did no better than giving every club the league average (5.5 points over 19 games), and simply assuming each club would repeat its first-half rate was worst of all. Manchester United show why: 30 points from their first 19 but a SOT difference of +32, so the SOT line projected 1.74 points per game against 1.44 from points; they took 41 points, 2.16 a game. The SOT table's biggest miss was Chelsea, projected at 1.65 and finishing the run on 1.16.

Two horizontal bar charts comparing four first-half Premier League 2025-26 tables as predictors of second-half points per game. Left, rank correlation: points 0.46, goal difference 0.41, shot difference 0.68, shots-on-target difference 0.69. Right, typical leave-one-out miss in points per game: points 0.300, goal difference 0.285, shot difference 0.272, shots-on-target difference 0.217, against 0.289 for projecting every club at the league average.
Each club's first 19 matches predict its points per game over the last 19. Gold marks the best predictor on each measure.Data: ESPN public Premier League box scores, pulled 7 September 2026.

Limits. This is one season and 20 clubs. With 20 clubs, a correlation's standard error on the Fisher scale is 0.243, so the gap between 0.77 and 0.60 is suggestive, not proof; another season could reorder the middle two rows. Shot counts treat a 30-yard hopeful and a tap-in the same, which is exactly what real xG fixes, and ESPN's total-shots count includes blocked shots. Nothing here adjusts for opponents, game state or red cards. A similar result at the five-game mark, using shot counts, is in the table after five cannot see a shot, and how quickly shot numbers settle is measured in which stats are real after three games.

Build the table yourself in Python

Download epl_2025_26_boxscores.json into a folder, save this next to it, and run it with Python 3.12 or later. Every statistic in the file is stored as a display string ("19", "61.2"), so the code converts shots with int() before subtracting.

import json
from collections import defaultdict

data = json.load(open("epl_2025_26_boxscores.json", encoding="utf-8"))
games = defaultdict(list)                      # club -> one row per match
for m in data["matches"]:
    h, a = m["stats"]["home"], m["stats"]["away"]
    for team, gf, ga, me, them in ((m["home"], m["hs"], m["as"], h, a),
                                   (m["away"], m["as"], m["hs"], a, h)):
        games[team].append({
            "date": m["date"], "gf": gf, "ga": ga,
            "pts": 3 if gf > ga else 1 if gf == ga else 0,
            "shd": int(me["totalShots"]) - int(them["totalShots"]),       # stats are strings
            "sotd": int(me["shotsOnTarget"]) - int(them["shotsOnTarget"]),
        })

def total(rows, key):
    return sum(r[key] for r in rows)

table = sorted(games, key=lambda t: -total(games[t], "sotd"))
print(f"{'club':<24}{'pts':>4}{'GD':>5}{'shots':>7}{'SOT':>6}")
for t in table:
    g = games[t]
    gd = total(g, "gf") - total(g, "ga")
    print(f"{t:<24}{total(g, 'pts'):>4}{gd:>+5}{total(g, 'shd'):>+7}{total(g, 'sotd'):>+6}")

It prints the clubs sorted by SOT difference, starting with Arsenal: 85 points, +44, +243, +97, as in the table above. To run the half-season test, add this to the bottom of the same file:

from statistics import correlation, linear_regression, mean

clubs = sorted(games)
for t in clubs:
    games[t].sort(key=lambda r: r["date"])     # no club plays twice on one date
after = [mean(r["pts"] for r in games[t][19:]) for t in clubs]   # last 19, points per game
for key in ("pts", "shd", "sotd"):
    before = [mean(r[key] for r in games[t][:19]) for t in clubs]
    slope, icpt = linear_regression(before, after)
    mae = mean(abs(y - (icpt + slope * x)) for x, y in zip(before, after))
    print(f"{key:<5} r {correlation(before, after):.2f}  "
          f"rank {correlation(before, after, method='ranked'):.2f}  MAE {mae:.3f}")

It prints the fitted rows of the test table: pts r 0.60 rank 0.46 MAE 0.270, shd r 0.71 rank 0.68 MAE 0.242 and sotd r 0.77 rank 0.69 MAE 0.194. Those are in-sample misses; the leave-one-out column needs the fit repeated 20 times, once without each club, which our full script does.

Prefer a spreadsheet?

Add these lines to the first script to write the season table to a CSV file, then open it in any spreadsheet:

import csv
with open("shot_table.csv", "w", newline="", encoding="utf-8") as f:
    w = csv.writer(f)
    w.writerow(["club", "pts", "gd", "shot_diff", "sot_diff"])
    for t in table:
        g = games[t]
        w.writerow([t, total(g, "pts"), total(g, "gf") - total(g, "ga"),
                    total(g, "shd"), total(g, "sotd")])

The clubs land in rows 2 to 21, with points in column B and SOT difference in column E. In F2, put the gap between each club's points and the line through all 20 clubs, and copy it down to F21:

=B2-(INTERCEPT($B$2:$B$21,$E$2:$E$21)+SLOPE($B$2:$B$21,$E$2:$E$21)*E2)

Column F is the "Pts vs SOT line" column of the table above. Sort by it to rank clubs from furthest above their shots to furthest below. If you have xG figures from a provider whose terms allow it, put xG for minus xG against in column E instead: the same formula then gives the xG-difference table this page was first written about.

Why shots, and what xG adds

Expected-goal difference counts chances by their quality: a penalty counts for far more than a header from a corner. Shot difference counts them all as one. The two agree on the broad order of a league because good teams both take more shots and take better ones, and they disagree on teams that live on volume from distance or on a few clear chances. The usual claim made for xG difference is that it predicts future results better than goal difference or points. We cannot test xG on this data, but the shot-based version of that claim held in 2025-26, with the limits above.

Expected points (xPts) go one step further: they turn each match's chances into the points the team would average if the match were replayed many times. The gap between points and xPts is the over- and under-performance signal. Our last column is a rougher version of the same idea, built on shots on target instead of chance quality. For the full xG version in code, see expected points from xG in Python; for what xG is and is not, expected goals explained.

What an expected-goals table said about 2023-24 (historical figures, June 2026)

When this page first ran in June 2026 it was built on Understat's 2023-24 Premier League table, retrieved on 4 June 2026. Understat's robots.txt now asks every automated client to stay out, so we no longer pull it and cannot re-check these figures. They are kept here, labelled, because they show what a real xG table adds.

Manchester United were that season's largest outlier: eighth on 60 points from an expected-points total of 44.4, a gap of +15.6, with an xG difference of −14.4. Aston Villa finished fourth on 68 points, 12.6 above expectation, on an xG difference of only +2.3, and West Ham banked 10.7 points above theirs despite an xG difference of −23.4. At the other end, Nottingham Forest took 36 points against an expectation of 50.2 (−14.2) with a near-neutral xG difference of −4.1, and Chelsea's +18.3 xG difference was the fifth best in the league (behind Arsenal, Manchester City, Liverpool and Newcastle) while they finished sixth on 63 points, close to their 64.2 expected.

Making the table useful over time

A single season is a start. The table earns its keep when you rebuild it as a season runs: point the script at a newer season's box scores, or add each round's rows to the spreadsheet, and watch which clubs drift away from the SOT line and which drift back. A club well above the line at Christmas is not doomed, but in 2025-26 the line told you more about the next 19 games than the points did. A club below it with a positive SOT difference is the candidate to climb.

Two extensions are worth adding once the basic table works. First, a rolling ten-game SOT difference, which tracks form more closely than the season total; build a rolling xG form chart shows the method. Second, conditional formatting on column F: green above +5, amber from −5 to +5, red below −5. For how long the plain table takes to settle in the first place, see the table lies until October.

Reproduce it

Every figure in the 2025-26 section comes from our analysis script (rewritten November 5, 2026; it no longer fetches anything). It reads data_layer/epl_2025_26_boxscores.json, ESPN's public Premier League box scores for all 380 matches of 2025-26, pulled on 7 September 2026, through our archive loader, which turns the display strings into numbers. It builds the season table, splits each club's matches in date order, fits the projections with and without each club, writes every number to data/stats/build-an-xg-difference-league-table-espn.json and draws the chart. No network, nothing hand-entered.

Sources & data

  • 2025-26 table and half-season test (added November 5, 2026): all 380 Premier League matches of 2025-26 with ESPN's per-team box scores, /data_layer/epl_2025_26_boxscores.json. ESPN public scoreboard and summary feeds (soccer/eng.1), pulled 7 September 2026.
  • Understat, historical source only: the 2023-24 expected-goals and expected-points figures in the June 2026 section were retrieved on 4 June 2026. We no longer pull Understat (its robots.txt disallows all automated clients) and have not re-checked them.
  • FBref: advanced squad stats and historical league tables, with xG supplied via Opta.
  • StatsBomb: methodological writing on how expected-goals models are built and what drives chance quality.

Published by SoccerAnalytics from public data and the sources named in this piece. How this site is made.