What this research found
Arsenal won the 2025-26 Premier League with 89 points, and this retrospective asks how much of that margin was repeatable. Cumulative points and expected-goal difference were tracked against the last five English champions, average passing networks were drawn for each half of the season, and the title margin was decomposed into four separate sources. What emerges is an ordinary points curve sitting on top of a top-three expected-goals season, plus a mid-season change of shape that lifted set-piece goals from 9 in the first half to 15 in the second.
- The points pace was unremarkable for a champion until late. Arsenal reached matchweek 19 on 43 points, two behind Manchester City's 2020-21 winning curve and eight behind City's 2021-22, then took 46 points from the final 19 matches and sealed the title with two games to spare.
- The underlying numbers ran ahead of the table. Expected-goal difference finished at +40.3, third-best among the last five champions, even though Arsenal were not the league's highest-expected-goals side: City posted 80.0 against Arsenal's 78.0.
- The +55 goal-equivalent edge over a league-average team splits four ways: +40.3 from underlying play, +16.0 from set pieces at 24 scored against 8 conceded, +16.0 from finishing above expectation with 94 goals from 78.0 expected, and −17.3 from conceding 55 goals against an expected 37.7.
- The mid-season change was structural rather than stylistic. Possession share barely moved, from 51.9% to 51.2%, but the passing networks show Zubimendi becoming a single pivot with the squad's highest workload at 100.6 touches per 90 and Calafiori pushing into the left half-space to form a 3-2-5 in build-up.
- Two of the four components regress in opposite directions, leaving roughly two-thirds of the margin repeatable on the write-up's reading. If goals conceded converge on an expected 38 that is 12 to 15 goals saved, worth about 6 to 7 points, while the set-piece edge is projected to fall back to around 18 goals scored, costing about 3 points as rivals copy the approach.
How it was done
Per-match data for all 380 fixtures of the 2025-26 season and the five preceding seasons was compiled from public aggregators, covering shots, expected goals for and against, set-piece outcomes, defensive actions and possession share for all 20 clubs, with several expected-goals models ensembled where more than one was available. Three analyses followed. Cumulative points and expected-goal difference were plotted by matchweek against the title-winning seasons of Manchester City from 2020-21 to 2023-24 and Liverpool in 2024-25. Average passing networks were built from per-90 completed passes among the eleven most-used outfield starters, split at matchweek 19, with node size encoding touches and edge width encoding pass volume on a shared scale. Finally the goal-equivalent edge was partitioned into underlying performance, finishing variance, set-piece impact and defensive structure, checked for consistency against actual goal difference, and written up as a six-page newspaper-style long read closing on three predictions.
Data sources
- FBref — per-match Premier League data for all 380 matches of 2025-26 and the five preceding seasons
- Understat-compatible open-source expected-goals models, ensembled to reduce model-specific bias
- Comparison seasons: Manchester City 2020-21, 2021-22, 2022-23 and 2023-24; Liverpool 2024-25
Limitations
All comparison curves run over the standard 38-matchweek schedule with no adjustment for rescheduled fixtures. The three forward-looking predictions rest on published regression rates for variance and for the diffusion of tactical innovation rather than on anything measured in this dataset.
How this research was produced
K-Dense Web planned and ran this sports investigation end to end — gathering the sources, carrying out the analysis, producing the figures, and drafting the report. The full session transcript, including every intermediate step, is available to view.
