Gameweek 2: What Our AI Got Wrong — And the One Thing It Got Right

FootyPulse AI called 5 of Gameweek 2's 10 Premier League matches correctly. The BBC's pundits called 7. This is the week to talk about that honestly — because the whole point of publishing our record is that you get to see the bad weeks too.

Fifty per cent. Five right, five wrong. On its own that is a coin toss, and against the people we benchmark ourselves on it was a clear defeat: BBC Sport's AI and Chris Sutton both landed 7 of 10, three ahead of us. Only the BBC's guest pundit, on 4, finished below our model this week.

We publish the full accuracy record and the fixture-by-fixture pundit comparison precisely so we can't quietly bury a week like this. So here it is, in full, with the misses explained rather than glossed over.

The five the model got right

Credit where it is due first, because the wins were mostly the sensible ones.

The model backed Manchester City at Crystal Palace (60%) and City duly won 1-4 at Selhurst Park. It backed Arsenal at Aston Villa (58%) and the Gunners won 1-0 against a Villa side missing most of its first-choice attack. It had Manchester United to beat Ipswich, and United won 5-2. It had Sunderland past Fulham (1-0) and Chelsea to edge Brighton — and Chelsea's 4-3 win was the pick of them, a seven-goal game the model called from a modest 38% edge.

Five correct outcomes, and no complaints about any of them. The problem was the other five.

The five the model got wrong — and why

Four of the five misses share a single, uncomfortable theme: the model backed a home side, and the home side dropped points.

• Liverpool 2-2 Nottingham Forest — the model had Liverpool at 58% to win at home. Forest took a point.
• Bournemouth 1-1 Everton — backed Bournemouth (36%, its least confident call of the week, to be fair). A draw.
• Coventry 0-1 Hull City — backed promoted Coventry at home; promoted Hull won it.
• Tottenham 0-2 Newcastle — backed Spurs at home (48%). Newcastle won comfortably.
• Leeds 1-1 Brentford — backed Leeds at home (48%). Another draw.

Three of those five were draws the model didn't see coming, and that is the honest headline of Gameweek 2 for us: the model under-weighted the draw. It backed a home win in every one of Liverpool-Forest, Bournemouth-Everton and Leeds-Brentford, and all three finished level. Draws are the hardest result in football to predict — nobody, including the BBC's pundits, backs many of them outright — but three in one round is where our week was lost.

The Tottenham-Newcastle miss is the one that stings most, because it wasn't close. The model had Spurs favoured at home; they lost 2-0 without scoring. Interestingly, our own half-time re-prediction had already flagged the danger — more on that below.

The BBC comparison, in full

We are not going to pretend this was anything other than a loss on the scoreboard that matters most for this feature:

	Correct (of 10)
BBC AI	7
Chris Sutton	7
FootyPulse AI	5
BBC Guest	4

Sutton and the BBC's model both caught results our model missed — notably, both had the sense to hedge toward draws in a couple of the games that tripped us up. Over one gameweek that is a real, and deserved, three-point gap. What it is not is a verdict on the season: this is our second scored gameweek, the samples are tiny, and one round of football is nowhere near enough to separate a good model from a lucky pundit — in either direction. We said the same when the numbers flattered us; it has to hold when they don't.

The one thing that worked: the half-time re-prediction

Here is the week's genuine bright spot, and the one number only FootyPulse can show you.

Our model doesn't just predict before kick-off — for Pro members it re-predicts at half-time, using the first-half stats. In Gameweek 2 the half-time call went 6 of 10, one better than the 5 of 10 the pre-match model managed. It also landed two exact scorelines at half-time: Sunderland's 1-0 win over Fulham, and the 1-1 draw between Bournemouth and Everton.

The clearest example was Bournemouth-Everton. The pre-match model backed a Bournemouth win — wrong. By half-time, with the first-half stats in, the re-prediction had switched to a 1-1 draw. That is exactly what finished. The pre-match call missed it; the half-time call caught it.

Now the mandatory honesty: this is one gameweek. Ten half-time calls beating ten pre-match calls by a single result is a promising signal and nothing more — it is nowhere near a sample from which anyone should conclude "half-time re-prediction works." It might be noise. What we can say is that it is being tracked publicly and honestly, gameweek by gameweek, and we will report it whether it keeps beating the pre-match model or stops. Ask us again in ten gameweeks, not now.

What it means

Gameweek 2 was a bad week for our pre-match model and a good one for the BBC's pundits, and we would rather tell you that plainly than pick a flattering angle. The specific lesson — the model under-priced the draw in a round that produced three of them — is one we can watch for. The specific encouragement — the half-time re-prediction edged the pre-match call and nailed two exact scores — is one we will keep measuring before we make any claim about it.

The record is public for a reason. You can hold us to all of it — the 50% weeks as well as the good ones.

FAQ

How accurate was FootyPulse's AI in Gameweek 2?

5 of 10 correct outcomes (50%), its weakest of the two scored gameweeks so far. The full record is public on our accuracy page.

Did the AI beat the BBC pundits in Gameweek 2?

No. BBC Sport's AI and Chris Sutton both scored 7 of 10; FootyPulse AI scored 5. It finished above only the BBC guest, on 4.

What did the AI get most wrong?

It under-backed the draw: three of its five misses (Liverpool-Forest, Bournemouth-Everton, Leeds-Brentford) finished level after it backed a home win.

Does the half-time re-prediction work?

Too early to say. It scored 6 of 10 in Gameweek 2, one better than the pre-match call, but that is a single gameweek and not yet a meaningful sample.