Expected points is a simulation, not a standing
Every international break produces the same graphic: the Premier League table beside an expected-points table, with arrows showing which clubs are running above or below what they “deserved.” It is a genuinely useful exercise and it is presented in the worst possible way. Expected points is not a moral ledger and it is not an alternative standing. It is the output of a simulation built on top of another model’s output, carrying every assumption of both, and the assumptions are load-bearing. Understanding what the number is made of is the difference between using it as a forecast and using it as a grievance.
How the number gets built
Start with a match. Each shot has an expected-goal value from a model trained on historical chances — distance, angle, body part, assist type, defensive pressure where available. Now take both teams’ lists of shot values and ask: given these chances, how often would this match end in a home win, a draw, or an away win? The standard approach treats each shot as an independent coin weighted by its xG and computes the distribution of possible scorelines, either analytically or by running the match thousands of times. Multiply those three probabilities by three, one, and zero points, sum, and you have expected points for that fixture. Repeat across 38 matchdays and you have a table.
Note what just happened. A modeled value became an input to a probability distribution, which became a weighted sum, which got added to 37 other weighted sums. Every step is defensible. None of them is free.
The independence assumption
The simulation almost always treats shots as independent events. Football does not work that way. A rebound off the goalkeeper is not independent of the shot that produced it, and counting both at full value double-books the same opportunity. A team pinned in its box for a ten-minute spell generates a cluster of correlated attempts. A corner routine produces a header from a set position that a scramble in open play does not resemble. Better implementations correct for rebounds and cap sequences, but the correction varies by provider, which is why two reputable xPts tables published on the same Monday can disagree by several points across a season.
Game state is invisible and it is everything
This is the criticism that matters most. Expected goals has no idea what the scoreboard said. A side that scores early and then sits deep will record a low xG total for the remaining hour by design, and the simulation will read that hour as a team failing to create. A side chasing a two-goal deficit will pile up low-value attempts against a packed box and post an xG figure that flatters an afternoon it spent losing. Neither team was doing anything irrational. Both are misdescribed by a model that treats minute five and minute eighty as interchangeable. Game-state-adjusted versions exist and are better; the ones circulating on social media on Monday morning usually are not those.
There is a second-order version of the same problem. Managers who are good at protecting leads are penalised twice by the graphic: once for the low-value second half their approach produced, and again when the arrow labels the club as overperforming. Game management is a skill, it shows up in the real table, and the expected-points table is structurally incapable of crediting it. Any metric that consistently marks a repeatable ability as luck has a specification problem, not a discovery.
Penalties, red cards, and the tails
A penalty carries an expected value around 0.78, which makes it the single largest input any match can contain and gives the referee more influence over an xPts table than most strikers. Red cards break the model in a different way: the chances that follow come from a match with an extra man, which is not the sport the training data mostly describes. A club whose season includes several dismissals and a run of spot kicks will have an expected-points figure assembled from matches the model handles least well, and no arrow on a graphic will mention it.
Three points for a win is a nonlinear payout
Because the scoring system is 3–1–0, small differences in chance quality map to large differences in points, and the mapping is not smooth. Converting a narrow draw into a narrow win is worth two points; converting a narrow loss into a draw is worth one. A club that plays a lot of tight matches sits on the steep part of that curve, where modest variance in finishing produces enormous swings in the real table while barely moving the expected one. That is not a flaw in the model. It is a fact about football that expected points is correctly reporting and that fans reliably read as the model calling their season unlucky.
What it is actually good at
The defensible use of expected points is forecasting, not justice. Through the first ten or twelve matchdays, a club’s underlying chance creation and concession predict its remaining fixtures better than its actual points total does, because actual points carry more finishing and goalkeeping variance than most people assume. That is a real, repeatedly demonstrated result, and it is why analytics departments watch the gap. The correct sentence is “this side’s process suggests it will pick up points at a higher rate from here.” The incorrect one is “this side should be fourth.”
It also matters that the gap narrows as the season runs. Early on, actual points are a small sample and the model has the advantage. By March, a club has played most of the league twice, finishing variance has had time to wash out, and the real table is carrying far more information than it did in September. Expected points is most useful precisely when it is least persuasive to supporters, and least useful in April when everyone starts posting it.
How to read it on StatLine
StatLine carries the live Premier League table alongside player tables with appearances, goals, assists, shots, cards, and minutes. Those are the raw materials for a sanity check on any expected-points claim you encounter. Look at how a club’s goals compare to its shot volume, at whether one forward is carrying an unsustainable conversion rate, and at the cards and minutes columns for the disciplinary and rotation context a model is likely handling poorly. If the finishing looks extreme and the shot volume looks ordinary, the gap between the two tables is probably variance. If the shot volume itself is thin, it is probably the team.
The honest read
Expected points is a forecast wearing the costume of a standing. It is built by stacking a simulation on top of a chance-quality model, assuming shots are independent when they are not, ignoring the scoreline that shaped every shot it counts, and passing the result through a nonlinear points system that magnifies small differences. It still beats actual points as a predictor early in a season, and that is a real achievement worth respecting. Use it to anticipate what a club will do next. Do not use it to argue about what a club deserved. The league table is not a claim about merit either — it is just the only one of the two that pays out.