The Death of the 12th Man: How Home Advantage Is Fading in the Premier League

I analysed 9,380 Premier League matches across 25 seasons to trace the slow death of home advantage and found that the real story is not about stadiums getting quieter. It is about away teams getting smarter.

Something Has Changed

In the early 2000s, hosting a Premier League match meant something. Over the first five seasons of this dataset, home teams won about 46% of their matches. By 2020/21, that number had dropped to 38%. And the drop was not sudden, a smoothing technique called LOWESS (think of it as a flexible trend line that follows the data more closely than a straight line would) shows the decline has been gradual and persistent across the full 25-season window.

The average home goal difference tells the same story. In the early 2000s, home teams were outscoring opponents by nearly half a goal per match. By the mid-2020s, that margin has halved. The trend line now brushes the zero line - the point where playing at home offers no scoring advantage at all.

This is happening in one of the most commercially stable, consistently competitive leagues in the world. So the obvious question: why?

The Fortress Myth

Part of the answer is that home advantage was never distributed evenly. There is a romantic idea in football that every ground is a fortress. The data says otherwise.

Only six teams in the Premier League era have won more than 58% of their home matches: Man United (67.8%), Arsenal (67.1%), Liverpool (65.7%), Man City (65.1%), Chelsea (63.8%), and Tottenham (58.6%). These are, of course, the Big Six. After Tottenham, there is a 10-point drop to Newcastle at 47.4%, barely above coin-flip territory.

At the bottom, Norwich (29.8%), Watford (30.1%), Sunderland (30.5%), and West Brom (30.7%) have home win rates that would embarrass most away records. For these clubs, "home advantage" is a statistical fiction. They lose at home nearly as often as they win. The aggregate 45.8% home win rate is an average of two very different realities: elite clubs that genuinely dominate at home, and everyone else for whom the advantage has always been marginal.

This heterogeneity matters because the league-wide decline we are measuring is not evenly distributed. To understand who is losing their edge and why, we need to look at what happened when the crowds disappeared.

Three Trends Killing Home Advantage

1. Away teams are getting better

The Big Six's away win rate has risen from 45.2% pre-COVID to 47.2% post-COVID. These teams now win nearly half their away matches. The rest of the league's away win rate has also climbed, from 21.2% to 25.7%.

Period	Big Six Away Win %	Rest Away Win %
Pre-COVID	45.2%	21.2%
COVID era	50.0%	29.1%
Post-COVID	47.2%	25.7%

The gap between Big Six home performance (65%) and the rest (37%) has held roughly constant at about 27 percentage points for a quarter century. What has changed is the floor, not the ceiling. Everyone is slightly worse at home than they used to be. The middle and bottom of the table have lost more of their edge (albeit rising in the last year or 2). The big clubs have not suddenly become fragile; the rest just do not get the same boost from playing in familiar surroundings anymore.

This is consistent with a tactical explanation. The revolution in pressing systems, video analysis, and opponent-specific preparation has made it easier for well-coached teams to neutralise home environments. If you know exactly how Brighton will build from the back at the Amex, it is less intimidating. The clubs with the resources to invest in analytics infrastructure are the ones whose away form has improved most.

2. COVID accelerated a trend that was already underway

COVID provided an accidental experiment as stadiums emptied for the better part of two seasons. If home advantage were purely about fans, it should have vanished entirely during COVID and returned to its pre-pandemic level once they came back. Neither happened cleanly.

During COVID, the home win rate dropped from 46.5% to 41.6%. The drop during COVID was statistically significant. In plain terms, if you compared thousands of hypothetical versions of these seasons, you would see a gap this large by chance less than 1% of the time (p = 0.009). That makes us confident the empty stadiums genuinely hurt home teams.

But when fans returned, the bounce-back only got to 44.7%. So COVID did not cause the long-term trend; it accelerated it. And some of that acceleration appears to have stuck.

Note for the statistically inclined: An OLS regression with robust standard errors gives a COVID-era coefficient of -0.031 and a post-COVID coefficient of -0.018. Neither is individually significant at the 5% level given the limited COVID-era sample size, but both are pointing in the same direction: down.

3. COVID revealed whose home advantage was real and whose was borrowed

This is the finding that the aggregate numbers completely hide.

When the fans disappeared, Fulham's home win rate collapsed from 44.4% to 10.5%. They won 2 of 19 home matches. Craven Cottage - compact, atmospheric, perched on the Thames - was a competitive strategy which evaporated overnight.

Man United dropped from 72.3% to 50%. Old Trafford, the most formidable ground in the dataset, lost a fifth of its advantage. Arsenal fell from 68.7% to 47.4%. Newcastle went from 47.2% to 31.6%. But Man City went from 61.4% to 73.7%; Liverpool from 62.5% to 73.7%. Leicester surged from 39.7% to 52.6%. These teams got better at home without the crowd.

The explanation is structural. City and Liverpool, under Guardiola and Klopp, had built systems - tactical approaches, squad depth, coaching infrastructure - that do not depend on atmosphere. Their home advantage comes from pitch familiarity, reduced travel, and preparation routines, not from 50,000 people screaming at the referee. When the atmospheric component was removed, only the structural component remained. And for City and Liverpool, the structural component was enough.

For clubs such as Fulham and Norwich, the structural component was close to zero. The crowd was an advantage.

This has implications for the broader decline. As the Premier League becomes more tactically sophisticated and analytically driven, the atmospheric component of home advantage, the part that comes from noise and intimidation, is being diluted. What is left is the structural part, which favours clubs with resources. The death of home advantage is really the death of atmospheric home advantage, and it disproportionately hurts clubs that relied on it most.

Derbies: The Last Place the Crowd Still Matters

If atmosphere is becoming less relevant, high-intensity derbies should be the exception, matches where crowd intensity is at its peak and emotional stakes override tactical preparation. The data is instructive, if mixed.

The Merseyside derby has a 38% draw rate - over 50% higher than the league average of 25%. The intensity doesn't help the home side; it neutralises both teams. Neither Liverpool nor Everton can impose itself in this fixture.

The North London derby is the opposite: 51% home wins and only 14% away wins. Playing at the Emirates or the Tottenham Hotspur Stadium is a massive advantage in this fixture. The crowd demonstrably matters here.

The Manchester derby has converged to near-parity: 39.6% home, 41.7% away. Two elite squads, two elite coaching setups. The crowd is noise, not signal.

Across all 732 Big-Six vs Big-Six matches, home advantage is actually higher than the league average: 47.1% home wins with a goal difference of +0.44. When the elite play each other, the familiar surroundings appear to matter more than when a top team hosts a relegation candidate; they would beat anywhere.

The Draw Problem: Football's Stubbornly Random Middle

One of the most humbling findings in football analytics: draws are nearly impossible to predict, and understanding why is equally informative.

A prediction model trained on five-match rolling form differentials can distinguish home wins from non-home wins with 61% accuracy, a 16 percentage-point improvement over simply always guessing “home win.” But extend this to three outcomes (home win, draw, away win) and accuracy drops to 50%. The model predicts zero draws across 1,100 test matches. Not a single one.

This is not a modelling failure. It reflects something real about football. Look at how draw rates respond to the form gap between teams:

Form Gap	Home Win	Draw	Away Win
Much Better Away	25.5%	24.9%	49.6%
Even Form	47.9%	25.5%	26.6%
Much Better Home	67.1%	19.6%	13.3%

Home and away wins swing from 25% to 67% as form shifts. Draw rates barely move; they sit between 19% and 26% regardless of form differential. There is no observable pre-match signal that reliably separates a draw from a narrow win. The 1-1 (44.6% of all draws) and 0-0 (29.9%) are emergent outcomes of 90 minutes of play, not predictable from the form table.

Brighton draw 35.8% of their home matches, the highest in the dataset. Man City draw just 16%. The gap is informative: Brighton's style creates tight, contested matches; City's creates decisive ones. But even knowing this, a model cannot reliably forecast which specific Brighton match will end level.

Beyond football, this exercise is a clean illustration of how predictive models can identify directional tendencies in complex systems without being able to pin down which specific outcomes realise those tendencies. In simple terms, basing your betting decisions solely on a fancy machine learning model may provide as many benefits as risks.

Where the Model Was Right, and Where Football Happened

When the model says 75% chance of a home win, home teams actually win about 71% of the time. When it says 25%, they win 20%. The model’s most confident correct calls are unsurprising: Arsenal 2-1 Brentford (89% predicted), Man City 2–1 Leeds (88%), Man City 4–1 Ipswich (87%). Form, squad quality, and home advantage all aligned.

The interesting cases are where the model was confident and wrong:

• Arsenal 0–2 Aston Villa, 14 April 2024

Model: 79% home win. Arsenal were chasing the title, flying at the Emirates. Villa went there and won 2–0. This is the kind of result that separates sport from spreadsheets.

• Brighton 1–5 Everton, 8 May 2023

Model: 72% home win. Everton, fighting relegation, went to the Amex and scored five. Desperation created its own form.

• Bournemouth 2–0 Newcastle, 11 November 2023 (The biggest underdog win in the test set)

Model: gave Bournemouth a 17% chance at the Vitality. Newcastle were in Champions League form. If you could systematically identify these outcomes, you would break every bookmaker in the country. You cannot. That is the point.

Note: The flagship model uses 5-match rolling form differentials (goal difference, points, goals scored, goals conceded, red cards) and season fixed effects, trained on 8,102 matches till the 2021 season and tested on 1,100 matches from 2022 onwards.

So Is Home Advantage Dying?

Yes, but slowly, unevenly, and not for everyone.

The aggregate numbers are clear: home win rate has fallen roughly 1 percentage point every 3-4 years since 2000. COVID accelerated this, and the post-pandemic recovery was partial. The trend predates analytics, predates COVID, and shows no sign of reversing.

But the decline is not uniform. The Big Six remain genuine fortresses - Man United still win 68% at home across 25 years. The clubs losing their home edge are the ones in the middle and bottom of the table, the clubs whose advantage was primarily atmospheric rather than structural. As the league becomes more tactically sophisticated and away teams arrive better prepared, the noise of the crowd buys less than it used to.

The 12th man is not dead. But for a growing number of clubs, he is retired.

Appendix

• Team fixed effects: Adding unique team identifiers to the form features barely changes accuracy (+0.1pp), which means the form features are not simply proxying for "Man City is good." They capture something about momentum, the difference between a team in rhythm and the same team in crisis.

• Model Stability: Rolling cross-validation across 5 time-ordered folds confirms stability: 60.9% plus or minus 1.5%. Unfortunately, the limited accuracy stems from the lack of public form data.

• The calibration chart hugs the diagonal, AKA the model is honest about its uncertainty. The slight overconfident in the 60–70% range is worth noting, but well-calibrated elsewhere.