Gameweek 4 Verdict: FootyPulse's AI Has Its Worst Week of the Season

FootyPulse's AI picked the correct outcome in just 2 of Gameweek 4's 10 fixtures — 20%, its lowest weekly accuracy of the 2026/27 season so far. It also finished last of four in this week's head-to-head against the BBC's published predictions: BBC's AI went 30%, Chris Sutton 40%, and the BBC's guest pundit 40%.

It's the kind of week the FootyPulse model exists to publish rather than hide. The site's whole pitch is a record you can actually check, wins and losses both, so here's Gameweek 4 in full — where it went wrong, where it got it right, how it stacks up against the BBC's panel, and what happened once the model got 45 minutes of real match information to work with, which is a very different story.

The Gameweek 4 scorecard

Ten fixtures, six home picks, four away picks, zero draws called. Four of the ten results finished level — a detail that matters more than it might look, and one we'll come back to.

Fixture	FootyPulse pick	Predicted score	Actual result	Verdict
Aston Villa v Nottingham Forest	Aston Villa (48%)	1–0	1–2	❌
Bournemouth v Brentford	Brentford (38%)	1–2	2–2	❌
Chelsea v Hull City	Chelsea (68%)	3–1	2–2	❌
Crystal Palace v Ipswich Town	Crystal Palace (48%)	2–1	2–3	❌
Liverpool v Fulham	Liverpool (68%)	3–1	0–0	❌
Tottenham Hotspur v Everton	Tottenham (38%)	2–1	0–0	❌
Sunderland v Arsenal	Arsenal (55%)	1–2	0–2	✅
Coventry City v Brighton & Hove Albion	Brighton (42%)	1–2	0–5	✅
Manchester United v Manchester City	Man Utd (38%)	3–2	0–1	❌
Leeds United v Newcastle United	Newcastle (38%)	1–2	4–1	❌

Two of our most confident calls, two draws nobody saw coming

The two highest-confidence picks of the gameweek were also its two clearest misses. Chelsea at 68% to beat Hull City finished 2–2. Liverpool at 68% to beat Fulham finished goalless. Both were the model's strongest convictions of the week, and both ended level.

That's not a coincidence worth dressing up. The model called a winner in all 10 Gameweek 4 fixtures — it never selected a draw once — and four of those ten games finished drawn. Every single one of those four results was, by definition, a miss the model couldn't have avoided with its current approach to that outcome. Chris Sutton, by contrast, called two of the week's four draws correctly (Bournemouth v Brentford and Tottenham v Everton). BBC's AI called one. FootyPulse's model called none.

The other standout miss was more conventional: Leeds United against Newcastle United. The model backed Newcastle for a narrow 1–2 away win; Leeds won 4–1. That's a scoreline nobody in the AI-vs-pundits panel saw coming — BBC's AI, Chris Sutton, and the BBC's guest all picked against Leeds too — but it's still a heavy miss to carry into the accuracy log.

Where it got it right

Two calls landed, and both were away wins in fixtures with a clear favourite on paper. Arsenal were backed at 55% to win at Sunderland and did so 0–2. Brighton & Hove Albion were picked at 42% away at Coventry City and won emphatically, 5–0 — the model correctly saw the away side as favourites even at odds shorter than even money, and the result went further than the model's own predicted scoreline of 1–2.

Half-time re-predictions: catching four of the week's misses

Here's the other half of Gameweek 4. FootyPulse re-predicts every match at half-time using what's actually happened in the game — and this week that re-prediction turned a 20% pre-match record into 60% at half-time, a 40-point swing.

Fixture	Pre-match call	Half-time call	Actual	Verdict
Leeds United v Newcastle United	Newcastle (38%)	4–0 Leeds win	4–1	✅ Corrected
Manchester United v Manchester City	Man Utd (38%)	2–0 Man Utd win	0–1	❌ Stayed wrong
Coventry City v Brighton & Hove Albion	Brighton (42%)	0–2 Brighton win	0–5	✅ Stayed correct
Sunderland v Arsenal	Arsenal (55%)	0–2 Arsenal win	0–2	✅ Stayed correct, exact
Tottenham Hotspur v Everton	Tottenham (38%)	1–0 Tottenham win	0–0	❌ Stayed wrong
Aston Villa v Nottingham Forest	Aston Villa (48%)	1–0 Aston Villa win	1–2	❌ Stayed wrong
Chelsea v Hull City	Chelsea (68%)	3–3 Draw	2–2	✅ Corrected
Bournemouth v Brentford	Brentford (38%)	2–2 Draw	2–2	✅ Corrected, exact
Crystal Palace v Ipswich Town	Crystal Palace (48%)	1–4 Ipswich win	2–3	✅ Corrected
Liverpool v Fulham	Liverpool (68%)	2–0 Liverpool win	0–0	❌ Stayed wrong

Of the eight pre-match misses, the model corrected exactly half once it could see the actual match: Leeds v Newcastle, Chelsea v Hull, Bournemouth v Brentford, and Crystal Palace v Ipswich all flipped to the right call at the break, and Bournemouth v Brentford landed the exact scoreline too. The other four — Man Utd v Man City, Tottenham v Everton, Aston Villa v Nottingham Forest, and Liverpool v Fulham — stayed wrong at half-time and stayed wrong at full-time.

That last one is the pointed detail. Liverpool at 68% was one of the week's two big misses, and even 45 minutes into a game that was still goalless, the model held its Liverpool-win call rather than shifting toward the draw that actually arrived. Chelsea, the week's other 68% miss, did get corrected — the model spotted the draw brewing against Hull and re-called it as a 3–3, close enough in shape to the eventual 2–2 for the outcome to score correct. Same starting confidence, same magnitude of miss, two different half-time reads.

FootyPulse's site notes one half-time re-prediction was excluded from this gameweek's count — not locked before the second half, so it wasn't scored — a small but deliberate part of keeping this record honest rather than convenient.

How FootyPulse stacks up against the BBC's panel

FootyPulse's AI is compared every gameweek against the BBC's three published predictions — BBC's AI, Chris Sutton, and a rotating BBC guest — on the same 10 fixtures, scored the same way.

Predictor	Gameweek 4	Season to date (40 fixtures)
Chris Sutton	40% (4/10)	50% (20/40)
BBC's AI	30% (3/10)	47.5% (19/40)
BBC Guest	40% (4/10)	37.5% (15/40)
FootyPulse AI	20% (2/10)	35.1% (13/37)*

*FootyPulse's site-wide Premier League accuracy is 13/37 — three fixtures the BBC panel didn't publish a pick for aren't included in the head-to-head total of 40, which is why the two figures don't share a denominator. Full breakdown on the accuracy page.

FootyPulse's AI is currently fourth of four, both for the gameweek and across the season. That's the honest position after four rounds, and it's worth sitting with rather than explaining away.

The honest angle: early season, real numbers

Four gameweeks is 37 scored Premier League fixtures — enough to publish with confidence, not enough to draw hard conclusions from. The week-by-week trend has moved the wrong way for two rounds running: 43% in Gameweek 1, 50% in Gameweek 2, then 30% in Gameweek 3 and 20% in Gameweek 4. Season running accuracy sits at 35% (13/37).

Confidence hasn't been a reliable guide either. Across the season so far, FootyPulse's own high-confidence picks (65%+ win probability) have landed at 43% — better than low-confidence picks, but still wrong more often than right, and this gameweek's two 68% calls both missed. That's a genuine pattern worth watching, not dismissing: a model that's right less than half the time even on its most convicted calls has real room to improve on how it handles matches where a draw is live.

Gameweek 4's 40-point half-time swing isn't a one-off. Across the whole 2026/27 season so far — 37 matches with a locked half-time re-prediction — FootyPulse's model has gone from 35% accurate pre-match to 62% accurate once it re-reads the game at the break, a 27-point lift overall. That's a Pro feature, published in full on the accuracy page, and it's the clearest evidence yet that the model does its best work with real match information in front of it, not before kick-off — even if, as Liverpool v Fulham showed this week, it doesn't catch every miss.

What's next

Gameweek 5 fixtures are already live on FootyPulse, including Brentford v Chelsea, Tottenham v Aston Villa, and Newcastle v Hull City. The record will keep publishing exactly as it has all season — wins, losses, and the gameweeks like this one, in full.

FAQ

How did FootyPulse's AI do in Gameweek 4? It predicted the correct outcome in 2 of 10 fixtures (20%), its lowest weekly accuracy of the 2026/27 season, and finished behind all three of the BBC's published predictions for the round.

Did FootyPulse's AI beat the BBC pundits in Gameweek 4? No. Chris Sutton and the BBC's guest pundit both went 40% (4/10), BBC's AI went 30% (3/10), and FootyPulse's AI went 20% (2/10) — last of the four.

Why did the model miss so many results in Gameweek 4? Four of the ten fixtures finished as draws, and the model didn't select a draw in any of its 10 picks that week — so all four were automatic misses. Its two highest-confidence calls (Chelsea and Liverpool, both at 68%) were also both drawn matches.

Is FootyPulse's AI accuracy declining? Weekly accuracy has fallen for two rounds running — 50% in Gameweek 2, 30% in Gameweek 3, 20% in Gameweek 4 — though the season sample (37 matches) is still early. Season-to-date accuracy is 35.1%. The full trend is on the accuracy page.

Did half-time re-predictions help in Gameweek 4? Yes. FootyPulse's AI went from 20% correct pre-match to 60% correct at half-time (2/10 to 6/10) — a 40-point swing after re-reading the actual match. It corrected 4 of its 8 pre-match misses, though its Liverpool v Fulham call stayed wrong even at the break.