Anatomy of a 5-4 Champions League Thriller: A Football Data Case Study ← Blog 15 May 2026•11 min read<br>football analyticsdatabuilding in public<br>Anatomy of a 5-4 Champions League Thriller: A Football Data Case Study<br>A football data case study on PSG vs Bayern, expected threat, metric choice, and why the real value is when you can ask the 2nd, 3rd, and 4th question cheaply.Ray Kameda<br>Co-Founder & Chief Product Officer

Champions League knockout games aren't usually 5-4 thrillers. April's PSG vs Bayern semi-final first leg was. Bayern had 57% possession, an xG of 3.06 to PSG's 1.90, and six big chances to PSG's two. PSG won anyway. I dug into the data to see what might be going on.<br>TL;DR<br>PSG beat Bayern 5-4 despite trailing on xG (3.06 to 1.90), possession (57%), and big chances (6 to 2). Expected Threat (xT) places Marquinhos as the most dangerous player on the pitch. What an analysis like this really requires is the ability to ask the 2nd, 3rd, and 4th question cheaply, which is what changes the patterns you can see.Even the data agrees that the game was the second most exciting Champions League knockout match across the past six seasons. When I scored all 147 UCL knockout matches since 2020 in the dataset by total goals, total xG, and how close the scoreline stayed, only the Manchester City 4-5 Real Madrid match in April 2024 ranked higher.Going into the second leg, my question wasn't really how good PSG were. It was whether the first-leg result was repeatable, or whether Bayern would simply convert their chances next time. Was something structural giving PSG an edge, or was the 5-4 high-variance just noise that wouldn't survive a second round of 90 minutes? To answer that I needed to look at where PSG's attack actually came from, who was carrying it, and whether any of it was reproducible.What it took to answer those questions surprised me more than the answers themselves.Round one: what the first pass of numbers said<br>My first analysis went up on LinkedIn the day before the second leg. Five things stood out:PSG over-performs xG by +0.59 goals per knockout match. Across the 16 matches in the sample at that point, they were scoring more than the underlying numbers said they should, consistently, for over a year.<br>Three players carry 75% of the over-performance. Kvaratskhelia +5.07, Doué +4.31, Dembélé +2.69, for a combined +12.07 across the knockout sample. The rest of the squad collectively under-performs by −2.81. The "we beat better teams on xG" pattern was really "two or three of our finishers are on fire."<br>Marquinhos is PSG's highest cumulative xT contributor. Ahead of Vitinha, Nuno Mendes, Hakimi, and the wingers. A centre-back leading the threat-creation chart was the kind of finding that sticks out, and a friend's pushback later in the analysis would make me take it apart properly.<br>Pressing varies by matchup. PSG's PPDA (Passes per Defensive Action, a measure of how high and intense the press is) against Monaco was 5.3, an intense press with the ball recovered high. Against Bayern in the first leg it was 17.1, the deepest sit-back in the entire 16-match sample. Enrique seems to choose the press level by opponent.<br>Bayern leak from counters and corners specifically. Their defensive profile across 21 UCL knockout matches looked roughly league-average overall, but they concede at twice the baseline rate from counter-attacks (19% conversion against them) and corners (18%). The vulnerability was concentrated rather than general.<br>The leg-2 prediction I posted was that PSG's most likely goal pattern would be open-play volume from Kvaratskhelia or Doué, or a counter exploiting Bayern's set-piece weakness.In the second leg, Bayern drew 1-1 at home, outperformed PSG's xG once more (1.41 to 1.06), and went out 6-5 on aggregate. Across both legs combined, Bayern outplayed PSG on the underlying numbers and lost the tie anyway, which was the same pattern PSG had been pulling off all season.If the analysis had stopped there I would have called it a good outcome. Then a good friend Benjamin Turk read the LinkedIn post and pushed back on the most surprising claim.My friend was right to push back, the answer was more nuanced<br>His objection, paraphrased was: "Marquinhos has two goals and zero assists this campaign. He doesn't show up like a creator on the pitch, so how is he number one in expected threat?"My first instinct was to defend the metric, which is exactly the wrong move. When a number conflicts with what someone watching closely sees, the number is what needs explaining. So I went back to the data and asked three different questions of it.Round 1: Where on the pitch is the xT coming from?<br>I split Marquinhos's xT contribution by the zone where each action started:Defensive third (start_x under 35 metres): 3.2% Middle third (35 to 70 metres): 18.8% Attacking third (over 70 metres): 78.0% Whatever Marquinhos was doing, it wasn't line-breaking passes from deep. The heatmap of his xT origins clustered tight on the...