The Expanded Europa League: When 18 Clubs From 17 Domestic Leagues Collide Inside One Spreadsheet

02:47 in Shanghai, and a Fixture List With No Source

The clock in my study in Shanghai read 02:47 when I opened the fixture list. Outside the window, the city was as quiet as a spreadsheet that had not finished filtering. On the screen were eighteen names: Real Sociedad, Bournemouth, Crystal Palace, Lech Poznan, Celtic, Ferencvaros, Besiktas, Marseille, Juventus, NEC, Levski Sofia, Red Bull Salzburg, Lillestrom, Torreense, OFI, Hoffenheim, Plzen, Union Saint Gilloise.

Eighteen clubs. Seventeen domestic leagues. Seventeen countries. And the detail that held me longest: not one line stated where the list came from, who published it, or on what date.

I have worked as a sports data analyst for twenty-eight years, nearly a decade of it spent cross-checking fixture lists against probabilistic models. My first reflex in front of any dataset is not to read the team names but to ask about the source. A list without a source is like a lab result with no patient name: the numbers may be perfectly correct, but they do not belong to anyone, are not tied to any moment, and therefore cannot inform a decision.

I stayed at the desk anyway. The structure of the list, rather than the content of any single tie, was telling a more interesting story.

The Format Expanded. The Reference Frame Did Not.

Since the 2026-25 season, the UEFA Europa League has used a league phase of thirty-six clubs, each playing eight matches with no return legs. The Conference League did the same a tier below. This structure produces a new kind of fixture list: mid-table clubs from Europe's strongest leagues mixed with domestic cup winners from second and third tier leagues, and with sides that have only just escaped qualifying.

The list on my screen reflects that mixture exactly. At one end sit Juventus, Marseille and Real Sociedad, clubs with revenues in the hundreds of millions of euros. At the other sit Torreense, a Portuguese club currently playing in the national second division, and Levski Sofia, an icon of Bulgarian football that has spent years away from European group stages.

The distance between those two ends is not a distance in football quality. It is a distance in reference frames.

I say this from a specific place. In 2026, working as a senior analyst for a sports platform, I published a piece using a model built on expected goals for Shanghai SIPG against Shandong Luneng on matchday 18 of the Chinese Super League. The model produced 2.8 against 0.4 in favour of the home side; I predicted 3-1 while the traditional pundits all picked a draw. The final score was 3-1. The article drew fifty thousand views in twenty-four hours.

Then I abandoned the series to experiment with a basketball betting model. My editor took three days to calm down. I tell this story not to boast but to be precise: what I did in 2026 was not prophecy. It was the act of calibrating a model to the actual distribution of the Chinese Super League — shot volume from outside the box, the share of set-piece goals, the distribution of goalkeeper ability. The same formula dropped into another league without recalibration produces answers that are wrong in a very confident voice.

At the 2026 World Cup I was lead analyst for a betting company. A model built on passes allowed per defensive action and average centre-back height correctly predicted South Korea beating Germany 2-0. I went on social media and told people to bet it. In the round of sixteen the same model said Brazil would beat Belgium. I said so on live television. Belgium won 2-1. A large number of clients lost money because they listened to me. I spent three weeks rewriting the code, adding tournament-specific variables and a noise term.

Since then every piece I write carries a warning: a model is a probability, not a prophecy.

Club Coefficients Are a Compressed File, Not a Ruler

UEFA publishes club coefficients based on five seasons of European results. Media treat them as a measure of strength. Bookmakers use them as an anchor. Fans use them to argue online.

The coefficient is a lagging indicator. It measures an institution's habit over five years, not the quality of a squad today. A club that changes coach, replaces four starters and shifts from a low block to high pressing keeps the same coefficient, because the coefficient cannot read tactics.

The deeper problem is conceptual. A club coefficient is not a measure of quality; it is a compressed file of a domestic reference frame, and every compression loses information.

When Real Sociedad meet Bournemouth, we are not measuring two teams. We are comparing two compressed files encoded by different rule sets and hoping the decoders are compatible. Usually they are not.

Take fouls per ninety minutes, a widely used defensive metric. It depends almost entirely on the referee's threshold, and that threshold varies by league. A defender who commits 3.2 fouls per match in the Polish Ekstraklasa may commit only 2.1 in Europe — not because he defends better, but because European referees allow more contact before blowing. Feed raw foul counts into a card model and you are predicting the referee, not the player.

The same holds for almost every advanced metric in common use. Possession share depends on whether the opponent agrees to concede the ball. Passes allowed per defensive action depends on the league's average tempo. Shot volume depends on whether the club was taught in its youth system to shoot early.

Across the seventeen domestic leagues in my list there are at least seventeen different distributions of all these variables. When two clubs from different distributions meet, the match unfolds in a third distribution — a mixture — for which nobody has calibrated a model.

This is why European prediction models carry more error than their own domestic error. Not because European nights are inherently harder to read, but because we are measuring with a ruler carved in a different workshop.

Data Migration and Decay at the Border

I live in China and work for the Chinese market, but I grew up in Vietnam. The distance between those two football cultures taught me something I have never seen in a technical manual: data does not migrate intact.

It decays at the border. The decay always sits in the denominator.

Data disappearing is not missing data — it is a kind of data. Bringing a European-calibrated expected goals model into the Chinese Super League in 2026, I found three systematic deviations.

The first was long-range conversion. In Europe a shot from twenty-five metres converts at a low, stable rate. In China at that time the rate was markedly higher, because the distribution of goalkeeping ability had a longer tail. Hold the European conversion coefficient and you systematically undervalue every side that shoots from distance.

The second was the share of set-piece goals, considerably higher than in Europe's top leagues. A model without a dedicated set-piece variable credits the attack for work done by the training ground.

The third was the distribution of home advantage. In Europe it is estimated at roughly a quarter to a third of a goal. In leagues with uneven pitch quality the variance between stadiums is enormous. Using one league-wide home advantage coefficient is a structural error.

In Vietnam the problem is sharper still. Amateur analytics groups import European expected goals models and apply them to the V-League without recalibration, producing numbers that look professional and are wrong with discipline. That does not stop them. It makes them more confident, because the numbers look complicated.

I call it misplaced reverence. People do not worship data; they worship the feeling of working with data.

This connects directly to my fixture list. When Crystal Palace face Lech Poznan, or Union Saint Gilloise face Hoffenheim, the model you are using was almost certainly trained mostly on Europe's top five leagues. It has never seen enough samples from the Ekstraklasa, the Czech league, the Greek Super League or Liga Portugal 2. On those clubs the model has no opinion. It has prejudice.

The only honest response is to down-weight the model and up-weight variables that are universal: rest days, travel distance, squad depth, the domestic calendar surrounding the fixture. Those variables do not depend on a reference frame. They depend on a human body.

The Ignored Variables: Distance, Kick-Off Times, Rotation

Some pairings in my list are governed by geography itself.

Torreense, from Torres Vedras north of Lisbon, may travel to Norway to face Lillestrom: roughly two thousand eight hundred kilometres as the crow flies. Levski Sofia to Poznan is about a thousand kilometres. Salzburg to the Oslo region is around fourteen hundred.

Those numbers are unremarkable to an airline logistician. To a player who features on Sunday, flies Tuesday, plays Thursday, flies home Friday and plays again on Sunday, they are not.

Recovery research in elite sport puts the minimum window for near-baseline return after a high-intensity match at roughly seventy-two hours. The European calendar does not respect that number. It was never designed to.

As a result, rotation becomes the single most important variable in the league phase, and it is the variable most models ignore, because models predict match outcomes while rotation is a human decision made before kick-off.

One metric I use for a quick read on rotation capacity is the distribution of minutes given to substitutes over the last ten matches. Platforms such as VangBong.vn publish this as a player depth index. Different names, same idea: measure whether a coach actually trusts his bench. A high-depth side rotates heavily in the league phase and accepts dropped points. A low-depth side plays its strongest eleven until someone tears a muscle.

Kick-off times are similarly undervalued. A match starting at 17:30 and one starting at 22:00 Central European Time are not physiologically equivalent. The early slot grants more sleep before the weekend fixture; the late slot takes it away. Over a league phase this accumulates into a fitness debt no box score records.

Pitches and weather compound it. The European league phase runs from late September to late January. Nordic surfaces shift from wet grass to frozen ground; the Mediterranean gets heavy rain. A club meeting a frozen pitch for the first time in a season will see passing accuracy fall, turnovers rise and win rate drop below its own baseline. That is not tactics. That is physics.

I have reported from eight World Cups, eight Olympic Games and multiple editions of the Tour de France and Giro d'Italia. Cycling taught me what football learns slowly: in a multi-week event the decisive variable is not peak form but tolerance of decline. A club playing eight matches across four months is running a stage race too. Nobody hands out a yellow jersey for the best recovery.

Eight Matches, and the Uninvited Guest

Eight matches is a small sample. Uncomfortably so.

Assume a team with a true win probability of fifty percent per match, the rest split evenly between draw and defeat. Across eight matches the standard deviation of points is roughly two and a half. A side of constant quality can finish the league phase on ten points or on fifteen, and both sit inside normal variation.

Fifteen usually means direct qualification. Ten usually means a play-off. Same team. Same quality. Five points apart, and a completely different calendar, revenue stream and reputation.

This is why I bristle when someone calls a club "vastly improved" after a good league phase. Most of what we call improvement lives in the distribution, not in the club.

All models are wrong, but a few are usefully wrong. What is useful is not the prediction but the way the model forces us to state what we are ignoring. A model without a noise term is a model lying about its own confidence.

After my 2026 miss on Brazil against Belgium, I added a noise term to the code. Not to make the model better — to make it more honest. A noise term does not improve accuracy; it improves integrity. In this trade, integrity has higher commercial value than accuracy, because accuracy does not exist and integrity does.

Football stopped rolling in 2026, but randomness has never taken a lunch break. I wrote that line in 2026, looking back at matches played in empty stadiums, and realised the only variable never locked out was coincidence. It still walks freely between fixtures, still decides which ball hits the post, still picks someone to make a mistake on eighty-eight minutes.

Eight league-phase matches, with a sufficiently large noise term, cannot separate a good team from a lucky one. If you intend to draw tactical conclusions from a league phase, make sure those conclusions also hold across a full domestic season. Otherwise you are reading a short story and believing it is a novel.

The Counterintuitive Angle: The New Format Diluted the Definition of Success, Not the Football

The popular complaint is that expanding the European competitions dilutes the quality. Too many mid-tier clubs. Too many meaningless matches. Too many sides present only to collect the cheque.

I disagree, and I think that reading comes from a modelling error rather than an observation about football.

The new format did not dilute the football. It diluted the definition of a successful campaign.

Under the old format — four teams, six matches, two qualifiers — the objective was clear: first or second. Every rotation decision funnelled toward that. Under a thirty-six-team league phase there are at least four distinct objectives: direct top-eight qualification, a seeded play-off, an unseeded play-off, elimination. Those objectives do not rank on a single scale, and they shift round by round.

So there are moments when failing to win is the rational choice. A side sitting seventeenth, with two fixtures against strong opponents remaining, may reasonably decide that resting key players for a domestic match matters more than chasing three European points. That is deliberate optimisation, perfectly rational, and no probabilistic model accounts for it, because models assume both teams want to win.

This breaks a foundational assumption of most football outcome models: that the two teams' incentives are symmetric. In an expanded league phase, incentives are frequently asymmetric — and asymmetric by design.

Correlation is not causation here either. Everyone knows squad market value correlates strongly with league-phase points. But that correlation runs through a chain of intermediaries: high squad value allows more replacement options; more options allow rotation; rotation sustains intensity across eight matches. Read straight from market value to points and you skip the chain — and you will fail precisely on clubs with expensive but thin squads, or cheap but deep ones.

Of the eighteen clubs in my list, at least three strike me as expensive and thin, and at least two as the reverse. I will not name them, because I am reading a list with no source and no date. But I will say this: if you had to choose a single indicator to predict which side finishes the league phase as a top seed, do not choose squad value. Choose the minutes a coach distributes to players twelve through eighteen in his squad.

Signals for the Next Round

Three concrete signals.

First, the substitution pattern between the sixtieth and seventy-fifth minute at clubs with an important domestic fixture on Sunday. If a side withdraws both central midfielders on sixty-five minutes with the score level, that tells you about priorities, not fitness.

Second, the share of first-half goals from set pieces. In a season of shifting surfaces, set pieces are the shortest route from one reference frame to another, because they depend less on match tempo and more on pitch quality.

Third, the pressing volume of the away side in the opening fifteen minutes. It is the one indicator I trust to compare across leagues, because it measures intent before the match imposes its own structure.

Eight league-phase matches is a small sample, and I have said so plainly. A small sample is still useful, provided we admit it is small. Every spreadsheet is a meditation, except that when the meditation ends you have lost money. And if, after everything above, you still want to ask me who will win this season's Europa League, you have probably read the wrong analyst. The better question is: how many times has your model been wrong in the past, and did you write any of it down?