General article(s) on Polls

[hullenedge]

This good chap has just updated his website comparing all the MRP outcomes:-

inglesp.github.io/apogee/

[manchesterman]

Brilliant bit of work!

Interrogating the stats further, I note the following numbers of seats for each party where all 11 extrapolations show a win for their party:

Labour 334
Con 16
Lib Dem 18
SNP 6
PC 2

So, only 376 of the 631 seats are absolutely nailed on, based on this data, however this isnt necessarily good news for the Tories if they were hoping it would mean only a small majority for Labour.

Of the 11 extrapolations, there are only 223 seats where they are given at least 1 prediction. For Labour, the figure is 541!  78 for the Lib Dems and 48 for the SNP. 

Reform and Greens have no seats with 100% extrapolations for them, but have 15 and 4 seats with at least 1, respectively.

So this gives (FWIW) the following ranges:-

Lab 334 - 541
Con 16 - 223
LD  18 - 78
SNP 6 - 48
PC 2 - 4
Reform 0 - 15
Green 0 - 4
Other 0 - 2

[islington]

Jun 26, 2024 17:10:17 GMT manchesterman said:

Brilliant bit of work!

Interrogating the stats further, I note the following numbers of seats for each party where all 11 extrapolations show a win for their party:

Labour 334
Con 16
Lib Dem 18
SNP 6
PC 2

So, only 376 of the 631 seats are absolutely nailed on, based on this data, however this isnt necessarily good news for the Tories if they were hoping it would mean only a small majority for Labour.

Of the 11 extrapolations, there are only 223 seats where they are given at least 1 prediction. For Labour, the figure is 541!  78 for the Lib Dems and 48 for the SNP. 

Reform and Greens have no seats with 100% extrapolations for them, but have 15 and 4 seats with at least 1, respectively.

So this gives (FWIW) the following ranges:-

Lab 334 - 541
Con 16 - 223
LD  18 - 78
SNP 6 - 48
PC 2 - 4
Reform 0 - 15
Green 0 - 4
Other 0 - 2

I agree, fascinating stuff, with a huge variation between MRPs.

It would be very interesting, once the election is done and dusted, to compare these results with the outcome.

For instance, in the 376 seats where all 11 agree, did they prove correct?

And in the 255 other seats, who got closest? It should be possible to rank all 11 from best to worst.

Is there a list somewhere of constituency-specific polls? Again, it would be interesting to compare them with the actual result.

I write this from a position of strong scepticism about the reliability of MRPs and single-constituency polls.

[manchesterman]

If you click on the hyperlink for each constituency, it opens up a new table of all the projections for each seat; but compiling them all is a bit of a job, but if you're game
For example, I clicked on Bristol Central and the projections ranged from a 40% Labour win to a 20% Green win!!

[islington]

Jun 26, 2024 17:25:39 GMT manchesterman said:

If you click on the hyperlink for each constituency, it opens up a new table of all the projections for each seat; but compiling them all is a bit of a job, but if you're game
For example, I clicked on Bristol Central and the projections ranged from a 40% Labour win to a 20% Green win!!

Well, if we are going to measure the MRPs against the outcome, we need to be clear what we're measuring.

My gut feeling about MRPs is that they are designed to produce a winner for each seat, so if I carry out such an exercise (I might, but I'm not promising) then that's what it will be based on. Even if the forecast in a particular seat is a Labour win by 40%, and in the event Labour scrape home by half a dozen votes after four recounts, I'll still count that as a 'win' for that MRP. Likewise, if in another seat the MRP forecasts the LD to win by a tiny fraction over the Tory, but the actual result is an eye-wateringly tight Tory win over the LD, then that's actually a very good prediction but, harshly perhaps, I'd still call it a 'lose' for the MRP.

[islington]

Jun 26, 2024 17:10:17 GMT manchesterman said:

Brilliant bit of work!

Interrogating the stats further, I note the following numbers of seats for each party where all 11 extrapolations show a win for their party:

Labour 334
Con 16
Lib Dem 18
SNP 6
PC 2

So, only 376 of the 631 seats are absolutely nailed on, based on this data, however this isnt necessarily good news for the Tories if they were hoping it would mean only a small majority for Labour.

Of the 11 extrapolations, there are only 223 seats where they are given at least 1 prediction. For Labour, the figure is 541!  78 for the Lib Dems and 48 for the SNP. 

Reform and Greens have no seats with 100% extrapolations for them, but have 15 and 4 seats with at least 1, respectively.

So this gives (FWIW) the following ranges:-

Lab 334 - 541
Con 16 - 223
LD  18 - 78
SNP 6 - 48
PC 2 - 4
Reform 0 - 15
Green 0 - 4
Other 0 - 2

Just at a quick glance through I've spotted one seat (Exmouth) where the 11 MRPs come up with four different winners, and 29 where there are three.

I can't help but feel that some (conceivably all) of these self-appointed seers are going to have a lot of explaining to do on 5 Jul.

Con/Lab/LD/Ref (1)

Exmouth

Con/Lab/LD (12)

Aylesbury

Brecon

Chelmsford

Chichester

Didcot

Epsom

Frome

Harpenden

Newton Abbot

St Neots

N Shropshire

Mid Sussex

Con/Lab/Ref (13)

Basildon

Basildon S

Boston

Castle Point

Clacton

Gt Yarmouth

Hornchurch

Mid Leics

NW Leics

Louth

Newark

NW Norfolk

S Suffolk

Con/Lab/SNP (3)

W Aberdeenshire

Berwickshire R & S

Dumfries & G

Con/Lab/Grn (1)

Waveney Valley

[hullenedge]

Twitter thread about MRPs and the wide range of outcomes...uniform vs proportional vs beyond proportional swing:-

[hullenedge]

Thread showing when the disconnect between UNS and MRP began:-

[Davıd Boothroyd]

I was looking back at the YouGov MRPs in the 2019 election, as I had a vague memory it wasn't always very accurate.

There were two rounds. The first, from 27 November, was very good on seats - all within 10 of the actual result.

The final one was much less accurate - the result was at the end of the range and the seat totals for Conservative and Labour were out by about 30.

[hullenedge]

Jun 28, 2024 9:40:06 GMT Davıd Boothroyd said:

I was looking back at the YouGov MRPs in the 2019 election, as I had a vague memory it wasn't always very accurate.

There were two rounds. The first, from 27 November, was very good on seats - all within 10 of the actual result.

The final one was much less accurate - the result was at the end of the range and the seat totals for Conservative and Labour were out by about 30.

Sadly I can't locate the source but Iain Warren mentioned that another private (very final) MRP was commissioned in 2019. He did release the figures, but seems to have deleted much of his work, and that MRP did show an upswing to the Tories. Seat totals were out by around ten if I've remembered correctly.

[hullenedge]

The very final YouGov MRP for 2019 (not publicly released until November 2019):-

[The Bishop]

Jun 28, 2024 11:23:12 GMT hullenedge said:

The very final YouGov MRP for 2019 (not publicly released until November 2019):-

Eh?

[hullenedge]

Jun 28, 2024 12:21:36 GMT The Bishop said:

Jun 28, 2024 11:23:12 GMT hullenedge said:

The very final YouGov MRP for 2019 (not publicly released until November 2019):-

Eh?

Oops! It was released during 2020.

Makes you wonder how many MRPs have been privately commissioned this time. Bound to be costly but lucrative for the clients?

[hullenedge]

Nigel Marriott, best forecaster last time, has a top down method but same overall result as the bottom up MRPs:-

[hullenedge]

The Economist's final forecast based upon polls and MRPs:-

[manchesterman]

Jun 26, 2024 17:10:17 GMT manchesterman said:

Brilliant bit of work!

Interrogating the stats further, I note the following numbers of seats for each party where all 11 extrapolations show a win for their party:

Labour 334
Con 16
Lib Dem 18
SNP 6
PC 2

So, only 376 of the 631 seats are absolutely nailed on, based on this data, however this isnt necessarily good news for the Tories if they were hoping it would mean only a small majority for Labour.

Of the 11 extrapolations, there are only 223 seats where they are given at least 1 prediction. For Labour, the figure is 541!  78 for the Lib Dems and 48 for the SNP. 

Reform and Greens have no seats with 100% extrapolations for them, but have 15 and 4 seats with at least 1, respectively.

So this gives (FWIW) the following ranges:-

Lab 334 - 541
Con 16 - 223
LD  18 - 78
SNP 6 - 48
PC 2 - 4
Reform 0 - 15
Green 0 - 4
Other 0 - 2

If you take the mid-range of these figures (with the exception of LD/Green -high end-  & SNP - low end) this isnt a million miles away from the actual figures.

[Clark]

So pretty much every poll had Labour consistently higher than their actual result - I don't recall any poll having Labour as low as 34%.

I guess the Tories did a bit better than expected too at 24%

[The Bishop]

Isn't it 35% in GB-speak? And some polls had the Tories there - the campaign highest was 28%.

[Clark]

Oh I don't know, Wikipedia had Labour receiving 33.8% and the Conservatives 23.7%...

[The Bishop]

Those will almost certainly be UK-wide figures (thus including NI) Most, though not all, polling is done on a GB-only basis.