Gameweek 9 Verdict: Our AI Got 7 From 10 and Beat All Three BBC Columns. Our Own Numbers Say That Was Lucky

FootyPulse's AI called seven of ten Premier League results correctly in Gameweek 9. BBC Sport's AI column got five, and Chris Sutton and the BBC guest each got four. Here is every call, the three misses, why seven is better than our own probabilities expected, and what half-time re-predictions did this week.

Gameweek 9 was a good one for the FootyPulse AI. It got 7 results right from 10 (70%), the best score of the four forecasters. BBC Sport's AI column got 5 from 10, while Chris Sutton and the BBC's guest each got 4 from 10.

The record is public, wins and losses alike. Every FootyPulse call below was locked 12 hours before kick-off and can't be changed afterwards. We score the 90-minute win/draw/loss outcome, not the BBC's exact-score points system.

The scoreboard

Two things to say straight away.

First, this is a swing from last week. In Gameweek 8 we got 3 from 10 and all three BBC columns beat us. A move from 3 to 7 in seven days says more about how noisy ten matches are than about the model changing.

Second, a two-to-three match gap on a ten-match sample is not evidence that one approach is better. It is worth reporting. It isn't a conclusion.

Where we won: the two calls nobody else made

Two results were called correctly by the FootyPulse AI and by none of the three BBC columns:

• Coventry City 3-1 Sunderland. We backed Coventry at home (48%). BBC's AI went for a draw, and Sutton and the BBC guest both picked Sunderland.
• Bournemouth 2-0 Leeds United. We backed Bournemouth (48%). All three BBC columns went for a draw.

We also called Aston Villa 1-1 Fulham as a draw, and only the BBC guest agreed. It was our only draw call of the weekend, with the draw at 38% against 34% for Villa, so it was a narrow lean rather than a conviction.

The rest of our hits were shared: Chelsea 2-0 Manchester United and Liverpool 0-1 Arsenal were called correctly by all four forecasters. Brentford 2-0 Nottingham Forest and Tottenham 2-0 Crystal Palace were hits for us and for BBC's AI too, with the BBC guest also right at Brentford and Sutton also right at Tottenham.

Where we lost: three home calls that didn't land

We backed the home side in eight matches. Five of those won. Three didn't.

• Hull City 2-2 Ipswich Town. We had Hull at 38%, with a draw at 34%. BBC's AI was the only forecaster to call the draw.
• Manchester City 0-1 Brighton. The week's one clean sweep for the upset: all four forecasters backed City and all four were wrong. We had City at 48%. This was the only match nobody got right.
• Newcastle United 0-1 Everton. This was a three-way coin-flip in our locked numbers: 35% Newcastle, 32% draw, 33% Everton. We nudged towards the home side. Only Chris Sutton picked Everton.

One pattern worth watching: all three of our misses were home calls. Two draws turned up this weekend (Hull v Ipswich, Villa v Fulham) and we called one of them. Three of the ten matches ended in away wins, and we backed only one of those (Arsenal at Anfield).

The honest bit: seven is more than our own numbers expected

Here is the caveat that matters most. Adding up the probability we gave to each of our ten calls, we should have expected around 4.4 to 4.6 correct results from the week. Seven is well above that.

If our probabilities are well calibrated, a week as good as 7 or better would come up roughly one time in ten (about 9%). That is our own estimate and it assumes the ten matches are independent, which is only roughly true. It points to a week that ran above expectation, and ten matches can't tell us whether the model is better than its own numbers imply.

One split worth a look, with a heavy sample-size warning. The five calls where our pick had 48% or more (Tottenham, Coventry, Brentford, Bournemouth, City) went 4 from 5. The three calls where our pick was under 40% (Hull, Villa, Newcastle) went 1 from 3. That is the pattern you would hope to see from a sensible model, but eight matches is far too few to build anything on.

Our Friday preview and the locked record didn't match exactly

Our Gameweek 9 preview went out on Friday 30 October, when the model said it expected no draws and backed Everton at Newcastle. It warned that probabilities could move as team news landed, and two calls did move before the 12-hour lock:

• Aston Villa v Fulham: the preview said Villa. The locked call was a draw. It hit.
• Newcastle v Everton: the preview said Everton. The locked call was Newcastle. It missed.

The net effect is zero. Had the Friday calls stood, we would also have scored 7 from 10, with different matches on each side of the ledger. Several other probabilities moved by a point or two without changing a call. The scored record is the locked one, and it is what the accuracy page shows.

Record-keeping note: Chelsea v Manchester United is the odd one out. The accuracy page lists our pre-match call at 58% for Chelsea, while Friday's preview had it at 38%, and the stored record carries a refresh timestamp after kick-off. It also has no half-time re-prediction. The call was Home in both versions and Chelsea won 2-0, so the full-time score doesn't change, but we aren't treating 58% as a clean, locked-in-advance number, and we are checking why that record was refreshed.

Half-time re-predictions: no gain this week

Our half-time calls went 5 from 9, and two of those were exact scorelines: Newcastle 0-1 Everton and Bournemouth 2-0 Leeds. Chelsea v Manchester United is shown on the accuracy page but not scored for half-time, because no half-time re-prediction was locked before the second half. We compare only the nine matches that were scored. On those same nine, the pre-match calls went 6 from 9.

• Improved (1): Newcastle v Everton. The pre-match call (Newcastle) was wrong; the half-time call (Everton) was right.
• Got worse (2): Brentford v Forest and Aston Villa v Fulham. Both pre-match calls were right; both half-time calls moved away from the result.
• Unchanged (6): Coventry, Bournemouth, Tottenham and Liverpool were right both times. Hull v Ipswich and Manchester City v Brighton were wrong both times.

This isn't a like-for-like comparison either way. A half-time call is made with 45 minutes already played, so some improvement would be expected. This week the half-time call didn't beat the pre-match call. Nine matches can't settle which is better. Too early to tell.

Every call, fixture by fixture

What happens next

Gameweek 10 opens on Friday 6 November with Everton v Coventry City. Those predictions will be locked 12 hours before each kick-off and every one goes into the same public record, whatever happens.

Check every GW9 call, probability and result on the accuracy page, see the head-to-head with the BBC on AI vs Pundits, and browse every gameweek on the gameweek hub.

BBC predictions are factual attributions of publicly published forecasts. Source: BBC Sport's Gameweek 9 predictions. We score the 90-minute win/draw/loss outcome, not the BBC's exact-score points system.

Club analysis

Chelsea · Manchester United · Bournemouth · Coventry City · Hull City · Manchester City · Brighton · Aston Villa · Fulham · Liverpool · Arsenal · Newcastle United · Everton

FAQ

How accurate was FootyPulse's AI in Gameweek 9?

It got 7 of 10 results right (70%), the best score of the four forecasters. Its half-time re-predictions got 5 of 9, because one match has no half-time call recorded.

Did FootyPulse's AI beat the BBC's predictions in Gameweek 9?

Yes, by raw count: 7 from 10 against 5 for BBC's AI and 4 each for Chris Sutton and the BBC guest. A gap of two to three matches on ten games is within normal noise.

Which Gameweek 9 result did every forecaster get wrong?

Manchester City 0-1 Brighton. The AI and all three BBC columns backed City.

Which Gameweek 9 results did the AI call that no BBC column did?

Coventry City 3-1 Sunderland and Bournemouth 2-0 Leeds United.

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.