Gameweek 8 verdict: our AI got 3 from 10, and all three BBC columns beat it

FootyPulse's AI called three of ten Premier League results correctly in Gameweek 8. BBC Sport's AI column and Chris Sutton each got five. Here is every call, the three draws that did the damage, and what a ten-match sample can and can't tell us.

Gameweek 8 was a poor one for the FootyPulse AI. It got 3 results right from 10 (30%). BBC Sport's AI column and Chris Sutton each got 5 from 10, and the BBC's guest got 4 from 9, with no call recorded for Hull v Brentford.

We publish this because the whole point of FootyPulse is that the record is public, wins and losses alike. Every call below was locked 12 hours before kick-off and can't be changed afterwards.

The scoreboard

Two things to say straight away.

First, it was a hard week for everyone. The best score on the board was 5 from 10, and no one got more than half right.

Second, the gap between us and the BBC columns is two matches. On a ten-match sample that is well inside normal week-to-week noise. It is worth reporting, but it isn't evidence about which approach is better.

Where we lost ground: the draws

Three of the ten matches finished level: Ipswich 1-1 Nottingham Forest, Hull 1-1 Brentford and Sunderland 0-0 Leeds. Our AI called none of them.

• Ipswich v Forest: we had Ipswich (48% home, 27% draw, 25% away).
• Hull v Brentford: we had Hull (38% home, 34% draw, 28% away).
• Sunderland v Leeds: we had Leeds (50% away win).

The BBC's AI column called draws at Ipswich and Hull. Those are exactly the two matches that separate its 5 from our 3. Chris Sutton called the Ipswich and Sunderland draws.

In our own numbers, a draw was never the single most likely outcome in any of the ten matches this week. Our draw probability ranged from 21% to 34%, and Hull v Brentford was the closest call. Backing the most likely outcome means draws get missed, even when three of them turn up in one weekend. We don't yet know whether that is a flaw to fix or just how a ten-match week can fall. We'll keep tracking it on the accuracy page.

What we got most wrong

Our four most confident calls (58% or higher) went 1 for 4. Arsenal v Everton, at 68%, was the only one that landed.

• Chelsea 0-2 Tottenham. We made Chelsea 58% favourites. Our own match analysis flagged that Cole Palmer and Moisés Caicedo were missing and Wesley Fofana was suspended, and we still backed the home side. This was the only match all four forecasters got wrong, so we weren't alone, but that isn't an excuse.
• Manchester United 0-3 Bournemouth. United were 58% in our model, and Bournemouth won by three. Chris Sutton was the only one of the four to back the away side.
• Aston Villa 2-1 Manchester City. We had City at 58%. Only the BBC guest went with Villa.

A 58% call should still miss roughly four times in ten, so three misses from three is painful but not proof that the model is miscalibrated. With so few high-confidence calls in one week, we can't say more than that. We'll watch how our confident calls perform over a longer run, and we'll publish the answer either way.

For balance, there were three hits. Arsenal 2-0 Everton was our one exact-score call. Crystal Palace v Newcastle (38% to Newcastle) landed too. Liverpool v Brighton also counts as correct, but it was a coin-flip: 38% home, 38% away, with the tie going to Liverpool. Strip that one out and the week looks even softer.

How bad was it by our own measure? Going by the probabilities we published, we should have expected around 5 correct results from 10. A week as poor as 3 or fewer would turn up roughly one time in six if those probabilities are well calibrated. That is our own estimate, and it assumes the matches are independent. It points to a bad week, not to a broken model, though ten matches can't separate the two.

Every call, fixture by fixture

The other two misses were Coventry 0-1 Fulham, where we had Coventry at 48%, and the Sunderland v Leeds draw above.

The half-time box: a better day, with an asterisk

Our half-time re-predictions went 5 from 10, against 3 from 10 for the pre-match calls.

• Improved (3): Hull v Brentford, Manchester United v Bournemouth and Ipswich v Forest. All three went from wrong to right.
• Got worse (1): Palace v Newcastle. Our pre-match call had Newcastle, but the half-time call flipped to Palace, and Newcastle won 2-0.
• Unchanged (6): Arsenal and Liverpool were right both times. Chelsea, Coventry, Villa and Sunderland were wrong both times.
• Two of the half-time calls were exact scorelines: Hull 1-1 Brentford and Arsenal 2-0 Everton.

Here's the asterisk. This isn't a like-for-like comparison. A half-time call is made with the first half already played, so some improvement is expected. What matters over a season is whether it beats a simple "the half-time picture holds" baseline. That test isn't in this piece, and ten matches wouldn't settle it anyway. Too early to tell.

What happens next

Gameweek 9 starts on 31 October with Chelsea v Manchester United, and Liverpool v Arsenal follows on 1 November. Those predictions will be locked 12 hours before each kick-off, and every one will go into the same public record.

You can check every call, probability and result for GW8 on the accuracy page, and see the head-to-head with the BBC on AI vs Pundits. The full gameweek wrap-up is on the Gameweek 8 verdict page.

BBC predictions are factual attributions of publicly published forecasts. Source: BBC Sport's Gameweek 8 predictions. We score the 90-minute win/draw/loss outcome, not the BBC's exact-score points system.

Club analysis

Chelsea · Tottenham · Man United · Bournemouth · Aston Villa · Man City · Arsenal · Everton

FAQ

How accurate was FootyPulse's AI in Gameweek 8?

It got 3 of 10 results right (30%) and hit one exact scoreline, Arsenal 2-0 Everton. Its half-time re-predictions got 5 of 10.

Did the BBC's predictions beat the AI in Gameweek 8?

By raw count, yes: BBC's AI and Chris Sutton got 5 of 10 and the BBC guest got 4 of 9, against 3 of 10. But a two-match gap on ten matches is within normal noise.

Which Gameweek 8 result did every forecaster miss?

Chelsea 0-2 Tottenham. The AI and all three BBC columns backed Chelsea.

How many draws were there in Gameweek 8, and did the AI call any?

Three: Ipswich 1-1 Forest, Hull 1-1 Brentford and Sunderland 0-0 Leeds. The AI called none of them.

Can FootyPulse change a prediction after the match?

No. Predictions are locked 12 hours before kick-off and do not change after that. The full scored record stays public; live half-time re-predictions are a FootyPulse Pro feature.