Every pre-match table includes some version of "last five matches." It is the most familiar shorthand in football coverage: five coloured letters, a goals column and perhaps a rolling expected-goals average.
Yet two analysts can pull the "last five" for the same club on the same morning and get different numbers. One query filters strictly for domestic league fixtures. The other pulls the last five competitive games across all tournaments—bundling three league games, a midweek Champions League trip and a domestic cup tie against a third-division opponent where nine starters rested.
Neither query is a software crash, and both produce clean decimals. The problem is that an unlabelled rolling window hides which environment generated the numbers.
One team, three different "last five" averages
To see how quickly schedule mixing changes a baseline, consider a constructed five-match sequence for a top-flight club playing across three competitions in a single month:
| Match | Competition | Opponent level & lineup note | Goals for–against | xG for–against |
|---|---|---|---|---|
| 1 | Domestic League | Mid-table league rival (full starters) | 2–1 | 1.65 – 1.05 |
| 2 | UEFA Cup Phase | Elite European away fixture (low block) | 0–2 | 0.55 – 2.10 |
| 3 | Domestic League | Top-six rival (full starters) | 1–1 | 1.35 – 1.25 |
| 4 | Domestic Cup R3 | Third-tier opponent (8 starters rested) | 5–0 | 3.80 – 0.20 |
| 5 | Domestic League | Lower-half league opponent (full starters) | 2–0 | 1.80 – 0.80 |
These figures are hypothetical and designed to isolate the aggregation effect. Look at what happens when downstream code summarizes this club's attacking output:
- All-competitions 5-match mean:
(1.65 + 0.55 + 1.35 + 3.80 + 1.80) / 5 = 1.83 xG Forand1.08 xGA. - Domestic league-only 3-match mean:
(1.65 + 1.35 + 1.80) / 3 = 1.60 xG Forand1.03 xGA. - Single-competition cup lookup before Round 4:
3.80 xG Forand0.20 xGA(a sample ofN = 1).
The all-competitions average is pulled upward by the 3.80 xG cup rout against a lower-division side, while simultaneously absorbing the 0.55 xG defensive shell from a tough European away night. If the upcoming fixture is an ordinary domestic league game, the unadjusted five-match blend describes a team that neither existed in the cup tie (rotated squad) nor faces a Champions League away environment this weekend.
The single-match cup denominator trap
An even sharper distortion occurs when an automated pipeline switches from "all competitions" to "current competition only" without checking the denominator.
Suppose a club enters the second round of a domestic cup. A naive database query asks for the club's season-to-date goals per match and xG per match in competition_id = domestic_cup. Because the team has played only one cup tie so far—say, the 5–0 win in the table above—the query returns a "season average" of 5.00 goals per match and 3.80 xG per match.
If that number is fed directly into a Poisson goal model or an language-model prompt without its sample size (N = 1), the model treats a single ninety-minute cup tie as if it were a twenty-match league baseline.
As we noted in our guide on why missing football data is not zero, a rate without its denominator is an incomplete measurement. In cup competitions and early league rounds, a sound pipeline should:
- Attach the exact match count (
N) to every season or rolling rate. - Enforce a minimum sample threshold (for example, requiring at least 3 to 5 completed matches in the target competition before treating competition-only rates as standalone baselines).
- Document the fallback when the competition sample is too small—such as falling back to the club's primary domestic league baseline while flagging the cross-competition substitution.
When all-competitions windows are the right choice
Separating league and cup baselines for goal-rate estimation does not mean cup fixtures should be deleted from the pre-match record. For two specific questions, an all-competitions timeline is essential:
- Rest days and physical load: A Thursday night knockout tie that went to extra time affects Sunday's league match regardless of competition labels. Filtering out non-league games makes a congested seven-day, three-match stretch look like a full week of rest.
- Lineup continuity and injury context: If a key defensive midfielder picked up a suspension or hamstring injury in a midweek cup game, a league-only match log will not explain why the player disappeared from the weekend starting eleven.
The practical rule is to match the filter to the mechanism:
| Analytical question | Appropriate fixture scope | What to check |
|---|---|---|
| Baseline attacking and defensive strength | Same competition (or primary domestic league with opponent-strength adjustment) | Sample size (N), home/away split, and whether promoted or newly rebuilt squads are involved. |
| Rest, travel and congestion | All competitive fixtures | Hours since last kickoff, travel distance, and how many projected starters actually played midweek minutes. |
| Tactical matchup and recent momentum | Recent competitive matches with lineup context | Whether the coach used the first-choice system or a heavily rotated cup eleven. |
Home and away balance inside a short window
Short rolling windows have a second hidden bias: schedule asymmetry.
In a five-match league window, a club must play either 3 Home / 2 Away or 2 Home / 3 Away—and after postponements, occasionally 4 Home / 1 Away. If home advantage in a given league is worth roughly 0.30 to 0.40 expected goals per match, a five-game window with four home fixtures will naturally look stronger than a five-game window with four away trips, even if the team's underlying level has not changed at all.
Here is a tiny Python check you can run on any fixture list before feeding a rolling average into a model or summary table:
def summarize_rolling_window(fixtures, target_competition):
"""
Inspect schedule mix inside a rolling fixture window.
Each item in fixtures is a dict with:
comp, is_home, xg_for, xg_against, starter_minutes_share
"""
n_total = len(fixtures)
same_comp = [f for f in fixtures if f["comp"] == target_competition]
home_count = sum(1 for f in fixtures if f["is_home"])
rotated_count = sum(1 for f in fixtures if f["starter_minutes_share"] < 0.60)
return {
"window_size": n_total,
"same_competition_count": len(same_comp),
"home_away_split": f"{home_count}H-{n_total - home_count}A",
"heavily_rotated_matches": rotated_count,
"all_comp_xg_for": round(sum(f["xg_for"] for f in fixtures) / n_total, 2) if n_total else None,
"same_comp_xg_for": round(sum(f["xg_for"] for f in same_comp) / len(same_comp), 2) if same_comp else None,
}
sample_window = [
{"comp": "League", "is_home": True, "xg_for": 1.65, "xg_against": 1.05, "starter_minutes_share": 0.95},
{"comp": "Europe", "is_home": False, "xg_for": 0.55, "xg_against": 2.10, "starter_minutes_share": 0.91},
{"comp": "League", "is_home": False, "xg_for": 1.35, "xg_against": 1.25, "starter_minutes_share": 0.93},
{"comp": "Cup", "is_home": True, "xg_for": 3.80, "xg_against": 0.20, "starter_minutes_share": 0.32},
{"comp": "League", "is_home": True, "xg_for": 1.80, "xg_against": 0.80, "starter_minutes_share": 0.96},
]
print(summarize_rolling_window(sample_window, target_competition="League"))
# {'window_size': 5, 'same_competition_count': 3, 'home_away_split': '3H-2A',
# 'heavily_rotated_matches': 1, 'all_comp_xg_for': 1.83, 'same_comp_xg_for': 1.6}
Exposing those four metadata fields—same_competition_count, home_away_split, heavily_rotated_matches, and the split between all-competition and same-competition means—prevents almost every misreading of a short form table.
Checklist for reading any "last five" claim
Before trusting a rolling form comparison in a match preview—whether on XG Mind's team pages, our daily match intelligence briefings, or an external analysis—check four things:
- Competition boundary: Are these five matches from the same league, or do they mix domestic cups, continental phases and friendlies?
- Sample denominator (
N): If a cup or early-season average is quoted, how many completed matches are inside that average? - Venue split: How many of the window's fixtures were played at home versus away, and how are season-long home/away rates behaving by comparison?
- Lineup overlap: Did the players who generated that five-match form actually play in those matches, or was a cup tie contested by a second-string eleven?
Rolling windows are useful because full-season averages react slowly to coaching changes, injuries and tactical shifts. They only work when the matches inside the window are comparable to the match you are trying to understand.
Further reading:
- Hudl StatsBomb: Expected Goals (xG) Explained, on what shot-based xG measures and why single-match totals fluctuate around underlying team strength.
- FiveThirtyEight Club Soccer Predictions Methodology, on how cross-competition ratings adjust for opponent strength and competition tier rather than averaging raw match logs.
- The Backtest That Knew Tomorrow: Football Data Leakage, on ensuring rolling windows strictly respect kickoff cutoffs.