Election Campaign 2010: Campaign model improvements

This page first posted 24 April 2010

During the election campaign we can introduce special model features to increase the accuracy of our prediction. These features are possible because of the increased amount of available information during the campaign. There are two main features

During the campaign, the headline Electoral Calculus prediction is calculated using the Regional Predictor, and you should use that if you want to use the same regional model. See the next section for details.

1. Regional Swing

Normally we use models based on a uniform national swing (UNS) assumption, because there is not enough data to support more detailed geographical modelling. But during the campaign there are one or two large polls which provide some detail about the individual swings in each of the regions. We can use these swings to create a "uniform regional swing" model in each region. This should give more accurate predictions than the relatively cruder UNS assumption.

The methodology we use is described fully in the analysis article on Regional Swing

So far in the campaign, there has been one large regional poll conducted by YouGov between 4-11 April 2010, with a total sample size of 28,154. As other regional-scale polls are published in the campaign, we will use them to update the regional swing estimates.

The values of the YouGov poll, expressed using the same Regional Swing Analysis notation are shown in the tables below. These show

  1. Regional Swing : the change in party support between 2005 and 2010 by region
  2. Regional Swing Differentials : the change between the region's own swing and the average national swing

Note that the Regional Swing Differentials do not show the absolute swing between parties, nor the absolute difference between regions of a party's support. Instead they show the relative change in support of a party across the regions. Note that the sum of Regional Swing Differentials across all the parties in any one region is zero. This is because the total support in a region must be 100%, so the sum of changes must be zero. Additionally, the (weighted) sum over all regions of a single party's Regional Swing Differentials is also zero. This is because the differentials are relative measures of a party's success in regions, not an absolute measure. If a party does relatively well in some regions, it must do relatively worse in others.

  (3) Regional Swing
2005 - 2010
    (4) Regional Swing Differentials
2005 - 2010
Scotland 1-1-760835-5.04.5-2.86.2-2.9Strong SNP and Lab gain over the rest
The North7-10-363770.8- gain over Lab
North West9-6-639772.8-0.9-2.10.1Con gain over LibDem
Yorks/Humber 5-8-13883-0.9- gain over Lab
Wales 9-5-4-225102.80.1-0.1-2.0-0.9Con gain over PC
West Midlands9-9-217493.2-3.52.1-1.8Con and LibDem gain over Lab
East Midlands8-7-226631.4-2.21.8-1.0LibDem gain over Lab
Anglia 8-7-438591.8-1.9-0.10.1Con gain over Lab
South West0-1-22860- Lab and LibDem over Con
London 6-4-641223-0.21.1-2.11.1LibDem lose ground
South East7-3-6213310.92.1-2.1-0.9Lab gain over LibDem
Great Britain6.2-5.1-,1540.

We see the following trends:

 Party Relative GainRelative Loss
CONNorth West, Wales,
West Midlands
Scotland, South West
LABScotland, South West,
South East
North, Yorks/Humber,
West Midlands, East Midlands
LIBYorks/Humber, West Midlands,
South West
North West, London,
South East

Overall there continues to be some "depolarisation" as parties lose ground in their heartlands and make relative gains elsewhere.

You can make predictions based on these Regional Swing figures by going to the Regional Predictor, entering national support in the boxes marked "National Support", and pressing the buttons marked "Use YouGov Swing", and then "Predict Election". This is the same method that is used for the headline website prediction during the campaign.

2. Spread Betting Markets

Spread betting markets are another useful source of prediction information. Although opinion polls are valuable, they have shown a consistent bias in their estimates of party support. At each of the last four general elections (from 1992 to 2005), the final opinion polls of the campaign overestimated Labour's support and underestimated the Conservatives' support. The error in Labour's lead at the previous election was 2.9%, which is not insignificant. See Track Record for details.

Betting markets have the advantage that people are putting their own money at risk, which forces them to think clearly and make use of all possible information available to them. The markets also provide a useful function of taking an average over all (active) opinion on the subject to get a consensus view. The market is also forward-looking, compared with the backward-looking opinion polls.

To use betting markets to help make our prediction, we perform the following steps

For instance, the Sporting Index betting market showed the following prices as at 24 April 2010.

PartyBid SeatsOffer Seats

This has mid-prices of 311, 223 and 83 seats respectively (rounding up in all cases). We can then repeatedly experiment with the Regional Predictor to find the solution point (Con 36.7%, Lab 27.4%, Lib 25.6%):

Party2005 Votes2005 SeatsPred VotesGainsLossesPred Seats

This is a close match to the mid-market seats, so these support percentages are the market-implied figures.

During the campaign we regularly calculate the market-implied support figures. These usually show the Conservatives a little higher than the opinion polls, and Labour a little lower. This is consistent with the long-term pro-Labour bias in the opinion polls. To partially correct for this bias, we add the market-implied figures into our poll-of-polls. This "market poll" is given a sample size of 10,000 which means it will be weighted around 50:50 with the regular opinion polls.

Useful links