Most football stat sites show you "corners per match" as a finished number. Computing it yourself takes about ten lines once all matches live in one table. Here is the exact approach we use for four leagues, with the real output for 2025/26 against 2026/27 so far. It is the same pipeline that produces our goal averages for every league, by TohaProBets, just pointed at corners and cards instead of goals.
The table
One row per match. Nothing is stored per team, because every team split can be derived later.
CREATE TABLE matches (
league TEXT, -- 'APL', 'La Liga', 'Serie A', 'BR Serie A'
season TEXT, -- '2526', '2627', or '2025', '2026' for Brazil
date TEXT,
home, away TEXT,
hg, ag INTEGER, -- goals
hc, ac INTEGER, -- corners won by each side
hy, ay INTEGER, -- yellow cards
referee TEXT
);
A real row looks like this: ('2026-10-24', 'Arsenal', 'Everton', 2, 0, 12, 3, 1, 3, 'Tony Harrington'). Arsenal won 12 corners, Everton 3, and four yellows were shown.
Per-league averages in SQL
SELECT league, season, COUNT(*) AS n,
ROUND(AVG(hc + ac), 2) AS corners,
ROUND(AVG(hy + ay), 2) AS yellows,
ROUND(100.0 * AVG(hc + ac > 9.5), 1) AS corners_over_9_5,
ROUND(100.0 * AVG(hy + ay > 4.5), 1) AS yellows_over_4_5
FROM matches
WHERE hc IS NOT NULL
GROUP BY league, season;
The WHERE hc IS NOT NULL line matters: friendlies and some cup games arrive without corner data, and averaging over NULLs silently shrinks the sample. In this dataset every league row has corners, so n equals the full match count.
The same in pandas
df["corners"] = df.hc + df.ac
df["yellows"] = df.hy + df.ay
out = (df.dropna(subset=["hc"])
.groupby(["league", "season"])
.agg(n=("corners", "size"),
corners=("corners", "mean"),
yellows=("yellows", "mean"),
c95=("corners", lambda s: (s > 9.5).mean() * 100),
y45=("yellows", lambda s: (s > 4.5).mean() * 100))
.round(2))
The output
European data runs to 8 November 2026, Brazil to 18 November.
| League | Season | Matches | Corners | Over 9.5 | Yellows | Over 4.5 |
|---|---|---|---|---|---|---|
| APL | 2025/26 | 380 | 10.00 | 56.3% | 3.75 | 33.9% |
| APL | 2026/27 | 100 | 9.50 | 47.0% | 4.01 | 39.0% |
| La Liga | 2025/26 | 380 | 9.67 | 48.2% | 4.41 | 44.2% |
| La Liga | 2026/27 | 119 | 9.74 | 50.4% | 4.44 | 48.7% |
| Serie A | 2025/26 | 380 | 8.82 | 40.3% | 3.70 | 30.8% |
| Serie A | 2026/27 | 110 | 9.76 | 45.5% | 3.22 | 26.4% |
| Brazil | 2025 | 380 | 9.92 | 50.8% | 4.93 | 56.3% |
| Brazil | 2026 | 345 | 10.54 | 57.7% | 4.58 | 49.0% |
Three things jump out. England lost half a corner per match (10.00 to 9.50) but gained a quarter of a yellow (3.75 to 4.01). Serie A moved the other way on both: almost one extra corner (8.82 to 9.76) and half a card fewer (3.70 to 3.22). Brazil 2026 is the highest of all eight rows on both counts. The same aggregation for goals gives Serie A 3.11 per match against 2.43 last season.
Home and away splits
Split the sum back into its two columns and you get the home edge:
df.groupby(["league", "season"])[["hc", "ac", "hy", "ay"]].mean()
For 2026/27: APL home corners 5.00 against 4.50 away, La Liga 5.31 against 4.43, Serie A 5.03 against 4.74, Brazil 6.08 against 4.46. Away teams collect more yellows in England (2.23 against 1.78), while in Serie A the split is exactly level at 1.61 each.
Team level with one filter
The same table answers team questions. Newcastle's home corners this season:
SELECT AVG(hc), AVG(ac) FROM matches
WHERE league = 'APL' AND season = '2627' AND home = 'Newcastle';
Result: 3.6 won, 5.2 conceded over five games, down from 6.79 and 4.95 last season. A team-level number like this, set against the league baseline above, is the starting point for how Xpect turns match history into probabilities.
Pitfalls
- Sample size. Ten rounds is 100 matches in England; a 0.5 shift is meaningful at league level but not for one team.
- Season keys. Brazil runs on calendar years, Europe across two years. Keep the key as text and never sort it as an integer.
- Corner direction. Store corners won by each side, not "for" and "against", and derive the latter per team. Otherwise the home and away views disagree.
Top comments (0)