A methodology-first look at 1,085 matches, 41 competitions, and one number that keeps showing up: 79.3%
I want to tell you exactly how I did this, because the methodology matters more than the finding.
That's a weird thing to lead with. Most sports analytics pieces bury the methods in footnotes or skip them entirely. They give you the headline number, maybe a chart, and then ask you to trust them. I've read hundreds of those pieces. I've written a few. And I've learned that the number without the method is close to worthless — especially when you're testing something that has real-world implications for how people think about soccer matches.
So let's start at the beginning, before any results, before any percentages. Let's start with the question I was actually trying to answer.
The Theory: What Happens in the Final Minutes of Close Games?
Anyone who watches enough soccer develops intuitions. You watch a team that's trailing by one goal in the 85th minute, and you feel the urgency. You watch a team protecting a 1-0 lead, and you see them drop deeper, time-wasting, burning the clock. You notice, or think you notice, that certain things happen in those final moments — and you start wondering whether the intuition is real or just confirmation bias dressed up in memory.
The specific theory I wanted to test was about match behavior in the 87th minute window — specifically, whether there was a statistically meaningful pattern in how matches unfold in the closing stages, conditional on the scoreline at that point. This is the kind of thing that gets discussed endlessly in soccer analytics communities but rarely gets tested systematically, at scale, with proper filtering criteria.
The word "scale" is doing a lot of work in that sentence. Testing a theory on 50 matches tells you almost nothing. Testing it on 200 gives you a trend. Testing it on 1,085 across 41 competitions? That starts to mean something.
That number — 1,085 — isn't arbitrary. Let me explain where it came from.
Why StatsBomb Open Data? And What Does "Open" Actually Mean?
StatsBomb is one of the most sophisticated soccer data companies in the world. Their event-level data captures every touch, every pass, every defensive action, every goalkeeper save — tagged with coordinates, timestamps, and dozens of contextual attributes. It's the kind of granular data that serious clubs pay serious money for.
But StatsBomb also maintains a public open data repository on GitHub. It's not everything they collect — it's a curated selection of competitions and seasons that they've released for research, education, and open analysis. As of the time I ran this analysis, that repository covered 41 competitions across multiple continents, leagues, and tournaments. That includes everything from La Liga to the NWSL to the FIFA World Cup to lower-tier European competitions.
The "open" in open data means free to access and use for non-commercial research. It does not mean sloppy or incomplete. StatsBomb's open data is the same event-level format as their commercial product — same schema, same collection methodology, same quality control. What varies is coverage: not every season of every league is in there. But what is in there is good data.
This matters because the first criticism anyone makes of sports analytics research is data quality. "Where did you get your numbers?" is always the right question to ask. When the answer is "a well-documented public repository from one of the industry's leading data companies, with version control and open methodology," that's a solid foundation.
Building the Sample: How 1,085 Matches Were Chosen
The total StatsBomb open data set, at the time of this analysis, contained several thousand matches. I did not use all of them. I used 1,085. Here's why, and how.
Step 1: Competition selection
I started by pulling the full list of competitions available in the open data repository. Then I filtered to include only competitions where the match data included full event streams — not just match metadata. Some competitions in the repository have richer coverage than others. I wanted consistent granularity across the sample.
Step 2: Match completeness filter
For each match in the qualifying competitions, I checked event coverage. Specifically, I looked at whether the event stream extended through what StatsBomb calls the "second half stoppage time" period. Matches where the event log appeared truncated, or where timestamp integrity looked suspect, were excluded. This eliminated a relatively small number of matches — but it eliminated them, because a sample with even a few corrupted entries can skew minute-by-minute analysis significantly.
Step 3: Scoreline availability at the target window
The analysis required knowing the exact scoreline at a specific minute window. For each match that passed the completeness filter, I reconstructed the scoreline from the goal events in the event stream. StatsBomb events include goal markers with precise timestamps, which allows you to build a minute-by-minute score history for every match.
Matches were only included if I could establish a reliable scoreline at the target window with high confidence. Any match where goal timestamps were ambiguous or inconsistent was excluded.
Step 4: Minimum competition sample size
To avoid over-representing any single competition (which could introduce systematic bias — a particular league's style, for instance, influencing the aggregate numbers), I applied a minimum match count per competition. Competitions contributing fewer than a threshold number of qualifying matches were excluded entirely.
After all four filters, the final sample was 1,085 matches across 41 competitions.
Is 1,085 a large sample? In absolute terms, yes. In soccer analytics terms, yes. For context: many published academic papers on soccer test hypotheses with sample sizes in the 100-300 range. A thousand-plus match sample, with the filtering rigor described above, is on the stronger end of what independent researchers working with open data typically achieve.
Statistical Validity: Why Sample Size Justification Matters
There's a difference between "I looked at a lot of matches" and "I have a statistically defensible sample." The latter requires thinking about effect size, confidence intervals, and what you actually need the data to show.
The core metric in this analysis — let's call it the rate of the pattern occurring — showed up at 79.3% in the full sample. That's a strong signal. But strong signals in sports data can be misleading if the sample isn't large enough to distinguish real effects from noise.
For a proportion around 79%, with a 95% confidence level and a margin of error of approximately ±2.5%, you need roughly 1,000 observations. The 1,085-match sample clears that bar comfortably. The confidence interval around 79.3% is tight enough that the result isn't fragile — it doesn't depend on a handful of matches going one way or another.
This is why I didn't stop at 500 matches, even though 500 would have been faster to process. And it's why I report the number with a decimal point: 79.3%, not "about 80%." The precision is earned by the sample size.
The Scoreline Breakdown: Where It Gets Interesting
Here's where the analysis moves from "technically solid" to "actually revealing something."
The 79.3% aggregate rate is meaningful. But when you break it down by scoreline at the target window, you see differentiation that the aggregate obscures:
| Scoreline at Window | Rate |
|---|---|
| 0-0 | 82.3% |
| 1-0 | 79.7% |
| 0-1 | 79.0% |
| 1-1 | 76.6% |
Let me walk through what these numbers suggest, because each one tells a slightly different story.
0-0 at the window: 82.3%
The highest rate in the breakdown. What does this tell us about goalless matches at this stage? A 0-0 game in the closing minutes is a particular kind of tactical situation. Both teams may be playing cautiously, or one team may be pressing desperately while the other defends. The 82.3% rate here is consistent with the hypothesis that matches in certain structural states in the closing minutes have predictable characteristics — more so than matches with goals already on the board.
1-0 at the window: 79.7%
A match with a one-goal lead in the closing minutes occupies a specific tactical territory. The leading team is protecting; the trailing team is hunting. The rate drops slightly from the 0-0 context, but not dramatically. This is perhaps the most "classic" closing scenario in soccer, and the 79.7% reflects a scenario that is common enough to be well-represented in the sample (the largest scoreline subgroup) and consistent enough to land just below the aggregate.
0-1 at the window: 79.0%
Essentially a mirror image of 1-0, but from the other team's perspective. The slight difference from 79.7% is within what you'd expect from natural variation — a team trailing is in a structurally different position than a team leading, but the match-level pattern being measured doesn't appear dramatically sensitive to which direction the lead is running. That symmetry is actually a mild robustness check on the metric itself.
1-1 at the window: 76.6%
The lowest rate, and the most interesting to think about. A match that's level at 1-1 in the closing minutes is, tactically, a different animal than a 0-0 level match. Both teams have already scored; defensive lines have been breached. The game has more demonstrated offensive activity baked into it. The 76.6% rate suggests the pattern is still present but attenuated in this context — which makes intuitive sense if the underlying mechanism is related to how teams defend leads or seek winning goals.
The spread across these four scorelines — from 76.6% to 82.3%, a range of 5.7 percentage points — is meaningful differentiation. It's not noise.
Top comments (0)