some random exit polls/election day polls for other ballots

[nodealbrexiteer]

There is far less published on ITN's prior forecasts but piecing together what has been written by Robert Waller and Roger Mortimore's email to me as well as some information from Dr Payne in his general articles on election night forecasting I think I can safely say the following:

Leaving aside February 1974 where they only exit polled in 2 marginal seats ITN from October 1974 to 2001 consistently used exit polls concentrated in the marginals that would be key in that election(using ORC/Harris until 1992 and them MORI 1997 and 2001)-for example Con/Lab,Lab/Con,Con/Lib Dem etc. They generally used a one polling station per seat approach picking polling stations they believed to be representative of the seat and usually revisiting them at subsequent elections if possible.

The exit poll would them be conducted in those key seats to find changes in vote share from the last election and thus used to produce a seat forecasts. From Robert's kind explanation in an earlier post and from Mortimore's email it looks like the changes were applied to the relevant set of seats using a uniform swing (plus some adjustments for special seats like the BBC).

In terms of updating the forecast according to Payne:

'when a result comes in, update the estimate for the relevant category by a weighted sum of the exit-poll based estimates and the average of the changes from all the results declared so far in this category.'

This had the effect on putting great weight on the exit poll(as the forecast only changed due to a declared result in the relevant type of marginal unlike the BBC method which used information from other types of seats) so if an initial prior forecast wasn't that good like 1992 it would slow down getting near the correct result.

I don't know from 2005 how ITN updates the initial prediction. 

[nodealbrexiteer]

Dr Chris Prosser of royal holloway, University of London kindly replied to an email I sent to Prof Colin Rallings re the ITN method:

Colin passed your question over to me as I’ve been running the ITN models for the past couple of elections. At the moment ITN uses two different forecasting approaches:



The ‘old’ model (which has been used for a long time – not 100% sure when it was adopted as it precedes my time as part of the team), which is basically a group swing model – seats are classified into different groups before election night and then a weighted combination of exit poll swing and actual results is used to calculate the average swing for each group, which is then applied to seats that have not been declared yet.
The ‘new’ model (which we ran in the background in 2017 and used for the prediction in 2019, though the ‘old’ model was also running in 2019) is a regression model based approach which uses the exit poll as an initial prediction and then re-estimates the model as actual results come in.


The two approaches are actually more similar than that sounds – you could implement the old approach with a regression model if you only used the seat groupings as the predictor variables – and on the night they rarely disagreed by more than a few seats.



Hope that helps!

[nodealbrexiteer]

I'll leave the last word to Dr Payne-this from his last email:

Thanks for your email. I'm afraid I have not kept records of the individual party predictions. I converted them into the error in the forecast of the lead of the winning party over the second party for the table in the EurAmerica article.

One point I would like to emphasise (I might have already done said this to you) is that the Brown/Payne job was to do a results-based forecast, not a prior forecast. We used whatever polling data the BBC commissioned to start off our forecast to counter declaration-order bias (early results usually strong Lab seats in the North..) , usually by calculating pseudo results for a Con-Lab and a Con-Dem marginal. So our prior forecast error given in the EurAmerica paper is what our prediction program produced, not necessarily the prior forecast given by the BBC. I can't remember which elections any BBC prior differed from ours - but certainly 1970, Oct 1974 and 1997.

[nodealbrexiteer]

I'm told the Political Communications book on the 2019 General Election will have a chapter by Nick Moon with a history of the exit polls in this country

[nodealbrexiteer]

Though on asking Nick Moon himself he was vague on some details, don't know how far he'll go back and apparently little or nothing on vote shares and sample sizes, may be just seat forecasts.(it's an article I'd liked to have written somewhere! )

[nodealbrexiteer]

Noticed on viewing a clip of Jonathan Dimbleby last night giving details of the 2005 exit poll vote shares that he mentioned the 8% other figure included 2% SNP and 1% Plaid. BBC did not break it down like this.

[nodealbrexiteer]

Some rather paranoid people commenting on the lack of vote shares in the exit poll:

www.bbc.co.uk/blogs/nickrobinson/2010/05/hung_parliament_argument.html

www.bbc.co.uk/blogs/nickrobinson/2010/05/exit_poll_guarantees.html

[nodealbrexiteer]

Here is a breakdown of what seats each party got in various regions of Great Britain at the 1983 General Election and how the BBC 10pm forecast for each region compared to the results:

England South: Con 224, Lab 29,Lib 5, SDP 2(predicted 222/31/7/0)

England Midlands: Con 70, Lab 30 (predicted 59/38 and Lib 3)

England North: Con 68, Lab 89, Lib 5, SDP 1(predicted 77/82/4/0)

Wales: Con 14, Lab 20,Lib 2,Plaid Cymru 2(predicted 18/15/2,SDP 0,Plaid 3)

Scotland: Con 21,Lab 41,Lib 5,SDP 3, SNP 2(predicted 21/43/3/2/3)

[nodealbrexiteer]

error in BBC/NOP by election exit polls 1987-1992(differences from result for Con/Lab/Lib dem respectively)- second set of 4 by elections under revised methodology:

Kensington:-2,+2,0
Govan:-2,+1,-2
Glamorgan:0,+2,-1
Mid Staffordshire:-1,+2,-1

Eastbourne:-1,0,-1
Bradford:0,-1,+1
Paisley North:-2,+1,0
Ribble Valley:0,0,0.

[nodealbrexiteer]

BBC exit poll 1997 forecast vote share changes(real changes in brackets):

Scotland 72 seats: Con -9.7(-8.2),Lab 8.5(6.9),Lib Dem 0.2(-0.3),SNP 1.1(0.5)
Con-Lab 178 seats: Con -15.0(-12.7),Lab 15.7(13.0),Lib Dem -3.1(-3.1)
Con-Dem 154 seats: Con -13.8(-12.3),Lab 10.0(8.3),Lib Dem 0.6(0.2)
Rest of England and Wales 255 seats: Con -12.4(-11.4),Lab 10.7(9.6),Lib Dem -1.2(-1.3)

[nodealbrexiteer]

Apr 14, 2021 21:47:13 GMT nodealbrexiteer said:

BBC exit poll 1997 forecast vote share changes(real changes in brackets):

Scotland 72 seats: Con -9.7(-8.2),Lab 8.5(6.9),Lib Dem 0.2(-0.3),SNP 1.1(0.5)
Con-Lab 178 seats: Con -15.0(-12.7),Lab 15.7(13.0),Lib Dem -3.1(-3.1)
Con-Dem 154 seats: Con -13.8(-12.3),Lab 10.0(8.3),Lib Dem 0.6(0.2)
Rest of England and Wales 255 seats: Con -12.4(-11.4),Lab 10.7(9.6),Lib Dem -1.2(-1.3)

Source 1999 JSTOR article on BBC's 1997 election night forecast

[nodealbrexiteer]

1992 exit polls forecast changes in the marginals

real changes in brackets.

Con/Lab marginals:

BBC/NOP: Con -3.1(-1.2), Lab 8.4(6.9), Lib Dem -6.5(-6.6)

ITN/Harris: Con -2.4(-0.7), Lab 8.1(6.4), Lib Dem -6.6(-6.6)

Con/Lib Dem marginals:

BBC/NOP: Con -4.2(-1.7), Lab 4.9(2.3), Lib Dem -1.5(-1.4)

ITN/Harris: Con -0.5(-0.8), Lab 0.5(2.7), Lib Dem -1.2(-3.1)

[Andrew_S]

Apr 16, 2021 21:37:51 GMT nodealbrexiteer said:

1992 exit polls forecast changes in the marginals

real changes in brackets.

Con/Lab marginals:

BBC/NOP: Con -3.1(-1.2), Lab 8.4(6.9), Lib Dem -6.5(-6.6)

ITN/Harris: Con -2.4(-0.7), Lab 8.1(6.4), Lib Dem -6.6(-6.6)

Con/Lib Dem marginals:

BBC/NOP: Con -4.2(-1.7), Lab 4.9(2.3), Lib Dem -1.5(-1.4)

ITN/Harris: Con -0.5(-0.8), Lab 0.5(2.7), Lib Dem -1.2(-3.1)

Interesting. Obviously BBC and ITN were using a different collection of marginals because the real changes are different. Where did you get the figures from by the way?

[nodealbrexiteer]

Apr 16, 2021 21:40:21 GMT Andrew_S said:

Apr 16, 2021 21:37:51 GMT nodealbrexiteer said:

1992 exit polls forecast changes in the marginals

real changes in brackets.

Con/Lab marginals:

BBC/NOP: Con -3.1(-1.2), Lab 8.4(6.9), Lib Dem -6.5(-6.6)

ITN/Harris: Con -2.4(-0.7), Lab 8.1(6.4), Lib Dem -6.6(-6.6)

Con/Lib Dem marginals:

BBC/NOP: Con -4.2(-1.7), Lab 4.9(2.3), Lib Dem -1.5(-1.4)

ITN/Harris: Con -0.5(-0.8), Lab 0.5(2.7), Lib Dem -1.2(-3.1)

Interesting. Obviously BBC and ITN were using a different collection of marginals because the real changes are different. Where did you get the figures from by the way?

Yes different collections of marginals presumably with different maximum percentage gaps between the 2 parties-source below:


www.ncrm.ac.uk/polling/documents/The%20Opinion%20Polls%20and%20the%201992%20General%20Election.pdf

[nodealbrexiteer]

The article below on page 195(table 2) for the Con/Lab marginals also gives slightly different figures of BBC/NOP Con -3.1(-0.9),Lab 8.4(6.9),Lib Dem -6.5(-6.9).

ITN/Harris Con -2.4(-0.7), Lab 8.1(6.4), Lib Dem -6.6(-6.6)

Table 3-seat projections based on accurate polling in the exit poll marginals

BBC-Con 338,Lab 265,Lib Dem 24,Other 24-Con maj of 25
ITN-Con 329,Lab 270,Lib Dem 25 Other 27-Con maj of 7


www.electionscentre.co.uk/wp-content/uploads/2015/07/20150706093835OpinionPollingandtheAftermath.pdf

[nodealbrexiteer]

bbc/nop 1992 exit poll estimated mean changes since 1987 in different categories of marginals in figures that formed basis of 10pm forecast(from article in Political Communication book for the 1992 election):

figures in brackets are the actual outcome in all seats in the relevant category.

Con/Lib Dem-South:Con -7.1(-2.1), Lab 7.5(2.8),Lib Dem -1.3(-1.4)-swing 7.3(2.5)
Con/Lib Dem-North:Con -3.1(2.2), Lab 4.0(4.9),Lib Dem -1.0(-7.6)-swing 3.5(3.6)
Con/Lib Dem-Wales and Scotland:Con -7.5(-2.6), Lab 3.2(-0.4), Lib Dem 2.5(0.3)-swing 5.3(1.1)
Con/Lab & Con/Lib Dem-London:Con -5.0(-1.0), Lab 9.3(6.4), Lib Dem -5.0(-5.9)-swing 7.2(3.7)
Con/Lab-South:Con -4.6(-1.4), Lab 7.7(7.3), Lib Dem -4.4(-7.0)-swing 6.1(4.4)
Con/Lab and Con/Lib Dem-Midlands:Con -3.8(-1.3), Lab 10.4(7.7), Lib Dem -7.3(-7.1)-swing 7.1(4.5)
Con/Lab-North:Con 0.1(-0.2), Lab 6.3(6.0), Lib Dem -6.7(-6.3)-swing-3.1(3.1)
Con/Lab-Wales and Scotland:Con -3.4(1.0), Lab 10.4(5.9), Lib Dem -11.0(-9.4)-swing 6.9(3.5)

at 10pm the mean swing across all the 100 marginals polled was 6.0%(When the forecast was Con 301,Lab 298) and later the data for the revised 11pm forecast showed a lower swing of 5.8%(Con 305,Lab 296). When the final data was available after midnight the final swing figure was 5.5%(would have upped the Con figure to 306). At 9pm the figure was 7.1%(when as I said elsewhere Lab were being forecast as the largest party on 311 seats). Below are the mean changes in the same categories as above but at 9pm:

Con/Lib Dem-South:Con -6.7, Lab 7.3, Lib Dem -1.6-swing 7.0
Con/Lib Dem-North:Con -0.3, Lab 2.6 Lib Dem -2.3-swing 1.4
Con/Lib Dem-Wales and Scotland:Con -4.5 Lab 0.9, Lib Dem 0.2-swing 2.7
Con/Lab & Con/Lib Dem-London:Con -8.0, Lab 12.6, Lib Dem -5.1-swing 10.3
Con/Lab-South:Con -6.0, Lab 9.3, Lib Dem -4.4-swing-7.7
Con/Lab and Con/Lib Dem-Midlands:Con -5.4, Lab 11.8, Lib Dem -7.0-swing 8.6
Con/Lab-North:Con 0.2, Lab 6.9, Lib Dem -7.4-swing-3.3
Con/Lab-Wales and Scotland:Con -4.5, Lab 12.2, Lib Dem -11.4-swing 8.4

In this articles table of looking at the accuracy of the marginals poll it downweights seats in Scotland and Wales and excludes 'special seats'(those not forecast by the exit poll due to local factors) from the calculation(real changes in brackets:

Con/Lab: Con -3.1(-0.9), Lab 8.4(6.9), Lib Dem -6.5(-6.9)-swing 5.8(3.9)
Con/Lib Dem: Con -4.2(-0.8), Lab 4.9(3.2), Lib Dem -1.5(-3.2)

They also did a table comparing the actual mean changes in the sampled constituencies and the change in all constituencies that lay within the definition of marginal seats used to select the sample(excluding special seats):

Con/Lab:Con -1.2,Lab 6.9,Lib Dem -6.6 in sampled seats, changes in all marginals Con -0.9, Lab 6.9, Lib Dem -6.9-so a 4.1% swing in sampled seats compared to the actual 3.9%

Con/Lib Dem: Con-1.7, Lab 2.3, Lib Dem -1.4 in sampled seats, changes in all marginals Con -0.8, Lab 3.2, Lib Dem -3.2

[nodealbrexiteer]

From the same Political Comms book this is what they forecast for the 45 special seats(actual result in brackets):

Con 7.0(16)
Lab 12.1(9)
Lib Dem 18.4(16)
Other 7.5(4)

They are expressed as decimals as remember they are the sum of the probabilities of each party winning each seat

[Davıd Boothroyd]

That explains why David Dimbleby remarked (in the documentary Swing Time) about the fact that a big picture of Neil Kinnock was being displayed as the forecast winner of the election, until a few minutes before going on air when the 10pm numbers showed the Conservatives narrowly ahead.

[nodealbrexiteer]

And here is a breakdown of how those 45 special seats were expected to go(again real results in brackets):

Rebel MP standing 4 seats: Con 0.6(1), Lab 2.7(3), Lib Dem 0.7(0)
SDP seats/Plymouth 5 seats: Con 1.8(2), Lab 2.6(3),Other 0.6(0)
SNP target 5 seats: Con 1.0(3), Lab 1.2(2), Other 2.8(0)
Lib Dem defending 18 seats: Con 0.4(1), Lab 2.1(0), Lib Dem 15.1(16), Other 0.4(0)
Plaid Cymru defending 3 seats: Other 3.0(3)
By-election change 1987-1992 7 seats: Con 1.9(6), Lab 2.7(1), Lib Dem 1.7(0), Other 0.7(0)
Alliance loss 1987 3 seats: Con 1.3(3), Lab 0.8(0), Lib Dem 0.9(0)

[Andrew_S]

Not quite the same thing, but just after 10pm in 1992 Peter Snow went through his battleground with the forecasts for the 94 target seats, and you can seat what was expected to happen in the seats displayed on the big screen, such as Edinburgh West turning gold for the LDs (despite being on the Labour target list). But that was of course after the adjustment was made putting the Tories narrowly ahead by 301 to 298.

Incidentally, I have Sky's 1992 election show on video, but Sky don't allow their election shows to be uploaded to YouTube as far as I'm aware.