This year, we divided prediction errors into four different types of effect:
In summary, 21 seats were incorrect due to opinion poll bias, 14 seats due to regional effects, 7 due to tactical voting and 32 seats due to local factors and residual model error. But in total only 52 seats were incorrect because in some seats two or more errors cancelled each other out.
The causes of opinion poll error include:
The opinion poll error of 2.9% caused 21 seats to be incorrectly predicted. This number is calculated by comparing the prediction with two different sets of inputs. The first uses the average of the final opinion polls of the campaign. Those were: Con 31.7%, Lab 37.5%, Lib 22.8%. The second uses the actual support levels in the election of Con 33.2%, Lab 36.2%, Lib 22.6%. Seats which have different predictions under the two sets of inputs are seats predicted incorrectly due to opinion poll error.
In total, opinion poll error understated the number of Conservative seats by 17, overstated Labour by 15, and overstated the nationalist parties (SNP and PC) by 2.
|Bexleyheath and Crayford||LAB||0.2%||CON||2.9%|
|Forest of Dean||LAB||1.1%||CON||2.0%|
|Hammersmith and Fulham||LAB||1.0%||CON||2.2%|
|Inverness Nairn Badenoch and Strathspey||LAB||0.7%||LIB||3.5%|
|Milton Keynes North East||LAB||0.6%||CON||2.4%|
|Rugby and Kenilworth||LAB||1.7%||CON||1.4%|
|Surrey South West||LIB||1.4%||CON||1.1%|
Since the Conservatives were underestimated by the opinion polls, these seats are mostly those the Conservatives should have been predicted to win, but were not.
We can calculate how many seats were incorrectly predicted due to regional effects. We do this by taking the vote-based prediction above, and applying the actual regional swings to each seat, depending on its region. The seats which change hands are those for which regional swing was crucial.
|Croydon Central||LAB||1.6%||CON||0.4%||South London|
|Finchley and Golders Green||LAB||1.5%||CON||3.5%||North London|
|Shrewsbury and Atcham||LAB||0.4%||CON||0.0%||Severn|
The Conservatives polled relatively better than Labour, allowing for national trends, in London and southern England, but worse in the midlands and northern England. This explains the Conservative victories around London, and their losses in Yorkshire.
The Liberal Democrats gained comparatively in parts of northern England, but lost ground in southern England, particularly the South West and South regions. This is probably due to their realignment as a more left-wing party, and it cost them a few seats in the south and west.
Two distinct trends are visible here. Firstly the Conservative vote is polarising. It is becoming even stronger in the places where it is already strong, and even weaker where it is already weak. Secondly, the LibDem vote is shifting. It is moving away from its previous heartland in the "celtic fringes" of the south west and growing in the "old Labour" north.
We see this by comparing the prediction above (which makes allowance for poll error and regional swing), and applying the tactical voting parameters. The parameters are estimated using our standard analysis (details), and are:
Making the comparison with the section 2 results, we see that seven seats are predicted to change hands due to tactical voting. They are:
Note that these are still predictions and not the actual result. As it happens, Dartford did not change hands, but all the other six predictions were correct.
The table below shows the 32 seats, with the predicted winner and majority against the actual results. Important local or other factors are included as a comment, if applicable.
|Bethnal Green and Bow||George Galloway||LAB||13.5%||MIN||1.9%||Strong ex-Labour candidate|
|Blaenau Gwent||Peter Law||LAB||44.8%||MIN||25.9%||Labour in-fighting|
|Brent East||Sarah Teather||LAB||30.4%||LIB||8.7%||By-election hold|
|Cambridge||David Howarth||LAB||5.4%||LIB||10.0%||University seat|
|Dartford||Howard Stoate||CON||0.4%||LAB||1.5%||Very marginal|
|Devon West and Torridge||Geoffrey Cox||LIB||1.0%||CON||5.5%|
|Dorset South||Jim Knight||CON||1.9%||LAB||3.7%|
|Dumfries and Galloway||Russell Brown||CON||1.7%||LAB||5.7%|
|Dumfriesshire, Clydesdale and Tweeddale||David Mundell||LAB||4.8%||CON||3.9%|
|Dunbartonshire East||Jo Swinson||LAB||2.2%||LIB||8.7%|
|Dundee East||Stewart Hosie||LAB||1.4%||NAT||1.0%||Very marginal|
|Enfield North||Joan Ryan||CON||6.6%||LAB||4.7%|
|Falmouth and Camborne||Julia Goldsworthy||LAB||7.8%||LIB||3.9%|
|Finchley and Golders Green||Rudi Vis||CON||4.0%||LAB||1.7%|
|Gillingham||Paul Clark||CON||1.9%||LAB||0.6%||Very marginal|
|Hornsey and Wood Green||Lynne Featherstone||LAB||3.9%||LIB||5.1%|
|Leeds North West||Greg Mulholland||LAB||4.7%||LIB||4.2%||Anti-war hotspot|
|Manchester Withington||John Leech||LAB||18.7%||LIB||1.8%||Anti-war hotspot|
|Na h-Eileanan An Iar (Western Isles)||Angus MacNeil||LAB||7.3%||NAT||10.4%|
|Oldham East and Saddleworth||Phil Woolas||LIB||7.0%||LAB||8.3%|
|Preseli Pembrokeshire||Stephen Crabb||LAB||1.2%||CON||1.6%||Very marginal|
|Reading East||Rob Wilson||LAB||4.6%||CON||1.1%|
|Scarborough and Whitby||Robert Goodwill||LAB||3.5%||CON||2.7%|
|Solihull||Lorely Burt||CON||14.8%||LIB||0.5%||Big swing|
|Taunton||Jeremy Browne||CON||1.2%||LIB||1.0%||Very marginal|
|Thanet South||Stephen Ladyman||CON||2.9%||LAB||1.6%||Marginal|
|Westmorland and Lonsdale||Tim Farron||CON||1.0%||LIB||0.5%||Very marginal, decapitation|
|Wrekin, The||Mark Pritchard||LAB||0.6%||CON||2.1%||Marginal|
We can categorise these seats into four types:
Because eleven seats suffered two mis-predictions in opposite directions, the total number of seats mis-predicted (52) is less than the sum of all the seat errors (74).
In summary, the opinipon poll error was less than previous years but still contributed significantly to the total error in Labour's majority. Model error was a more important factor this year, with regional effects, tactical voting, and local factors all working in the same direction as poll error.
As it happens, the most accurate prediction of the result beforehand came from Electoral Calculus users who predicted a Labour majority of 66.
It should be assumed that about 50 seats will be mis-predicted at the next election.