Leagues

The hardest football league to predict: what actually decides it

By · Published · 6 min read

Ask which is the hardest football league to predict and you will get answers built on memory: the division with the famous upsets, the one where the champion changes often, the one that produced a result nobody saw coming. None of that is evidence. Predictability is a measurable property of a competition, it varies more between divisions than most people expect, and the same model doing the same work will score visibly better in some leagues than others without anything about the model having changed.

That last point is the one worth internalising. If difficulty is a property of the league, then any accuracy figure quoted without naming the leagues behind it is uninterpretable — and comparing two prediction services on such figures is comparing two different questions.

Predictability is a property of the league, not the model

Fit one model, hold every parameter fixed, and run it across several divisions. The scores will differ. A weaker model working a structurally easy league can outscore a stronger model working a hard one, which means a raw score is a statement about the fixture list at least as much as about the forecaster.

So “which league is hardest” decomposes into a better question: what structural features of a competition push probabilities toward the middle, where they carry less information and score worse?

Four forces that make a football league hard to predict

Dispersion of team strength. This is the dominant term. A league with a wide gap between its best and worst sides generates a steady supply of genuine mismatches, and mismatches are where a model earns its score: a well-founded 75% is worth far more than a well-founded 40%. Compress the quality range — as promotion economics, spending caps or simple depth of competition can — and the same model produces probabilities clustered near the three-way even split, which is a mathematically worse position to forecast from even when every number is correct.

Draw frequency. Draws are the structural tax on football forecasting, because a draw is rarely the single most likely outcome even in the divisions that produce the most of them. A draw-heavy league therefore hands the forecaster a slice of the fixture list that is very hard to call at the top of the distribution, which depresses hit rate mechanically. This is the piece most people already sense, and it sets a ceiling on hit rate that no amount of modelling removes.

The size and stability of home advantage. Home advantage is free signal: it is one of the few effects available before a ball is kicked. A league with a large and consistent home edge gives every prediction a reliable head start. A league where the home edge is small, or drifts, removes that anchor and pushes more of the work onto team ratings that are themselves uncertain.

Squad churn and structural turnover. Ratings are built from past results, so they degrade whenever the team that produced those results stops existing in any meaningful sense. Divisions with heavy summer rebuilds, aggressive mid-season trading, or a high share of promoted clubs with little or no history in the dataset will carry more stale and more provisional ratings at any given moment. A promoted club is not merely a weaker input — it is often an absent one.

Two secondary factors modulate all of this: matches per season, since a longer calendar lets ratings converge faster within the season, and data availability, since some divisions simply have deeper, cleaner historical coverage than others.

Why accuracy is the wrong yardstick for ranking leagues

Hit rate confounds two things that need separating: how hard the league is, and how the fixture list happened to fall. A season with an unusual number of lopsided fixtures inflates accuracy for everyone in it.

The fix is a proper scoring rule — a Brier score or a ranked probability score — read against a league-specific baseline rather than against zero. The baseline is the score you get by predicting that league’s long-run base rates for every single match, with no model at all. That number differs from division to division, so an absolute score means nothing until you know what beating nothing looks like there.

Take two hypothetical, illustrative cases. In League A, the base-rate baseline scores 0.213 and a model scores 0.205. In League B, the baseline scores 0.241 and the same model scores 0.219. The raw scores say League A is where the model does better. The comparison against baseline says the opposite: 0.008 of improvement in the first case against 0.022 in the second. The second league is harder in absolute terms and yet is where the model is genuinely adding more. Both readings are true, and only one of them answers the question you asked.

This is also why the honest answer to “which league is hardest” is conditional rather than a name. It depends on the scoring rule, the seasons in the window, and whether you are ranking by absolute difficulty or by how much room a model has to beat the naive benchmark.

The comparison trap this creates

Two prediction services can quote accuracy figures that differ by ten points while performing identically, purely because one covers draw-heavy, tightly-packed divisions and the other covers top-flights with wide quality gaps. Neither number is a lie. The comparison is still meaningless.

Anyone taking a published figure seriously should therefore ask three things before anything else: which competitions are in the denominator, whether every prediction in those competitions is included, and whether the score is reported per league rather than pooled. Pooling is where league mix hides. A single headline figure over a mixed portfolio can be moved substantially by adding or dropping one division, and nothing in the number itself reveals that it happened.

How to find the hardest football league to predict yourself

The procedure is unglamorous and it works. Fix one source of predictions and one market. Choose a season window and apply it identically to every league. Record the full probability distribution for every fixture before kick-off, not the pick. Score each league separately with the same rule. Compute that league’s base-rate baseline from its own historical outcome frequencies, and report the gap, not the raw score. Then check the sample is large enough that the gap survives ordinary variance — a few dozen matches will produce differences that mean nothing.

Do this and you will have something almost nobody publishes: a defensible per-league difficulty ranking, with the model’s contribution separated from the competition’s structure.

Scored per league, never pooled

Our own track record is scored per league and per market rather than pooled into one banner figure, precisely because pooling would let an easy division flatter a hard one. Every prediction is stored before kick-off with its model version and settled after full time, including the leagues where the model reads the football badly — those are the informative ones. The model’s known weaknesses are set out on how our predictions work, and each competition we cover has its own hub, such as the Championship, where the daily probabilities live.

The league that is hardest to predict is, in the end, whichever one your metric and your window say it is. What matters more is that the question is answerable at all — and it only becomes answerable on a record that was published first and never edited afterwards.