Poll of Polls
How can you aggregate all the polls being published? There are two elements to the equation.
First, you need to decide whether and how to weight each polling firm. These ‘house’ measures would take into account everything from the way a firm asks certain questions (does it prompt for UKIP?), weights its respondents by certain demographics (how does it account for who people voted for at the last election?), or uses different sample sizes. It could also include their track record over past elections.
We think these judgements are largely subjective and arbitrary. We have given all of our pollsters the same house weighting. Eight of the nine we use are members of the British Polling Council, and the ninth, Lord Ashcroft, is producing well-respected polls of the election’s key marginals.
The other eight pollsters in our sample are ComRes, ICM, Ipsos MORI, Opininum, Populus, Survation, TNS-BMRB and YouGov. They are all Council members and the media report their polls with equal regard. House weightings could be adjusted on demand.
The more important challenge is deciding how to weight the polls these firms produce at different times. How much weight should YouGov’s daily polls get in our model versus the monthly surveys published by other firms? This is something you can decide. You can choose to pick up the ‘noisy’ day-to-day polls or average results over time – across 5 days, or 2, 4 and 8 weeks (our default average is 5 days). Smoothed versions of the data iron out the effect of individual pollsters who publish regularly, like YouGov. But if you take them out of a smoothed chart the data displays very little difference.
As you head back to 1970 you can also choose between picking up polls each day or just see those published week-by-week or month-by-month.
You can read a full explanation of the challenges when it comes to weighing pollsters on Anthony Well’s comprehensive UK Polling Report website – the basis for our data.
For a full explanation of our election-forecasting machine, click-through to our seat calculator page.
Our Seat Calculator gives you the option of generating seat projections in two different ways: the widely used Uniform National Swing model, or our preferred Strong Transition Model, developed by Electoral Calculus.
The Uniform National Swing (UNS) model is very simple, but does produce reasonably accurate seat projections. If a party increases or falls by X% in the opinion polls, the model projects that votes will change by X% in every single seat. This model can clearly lead to some illogical and impossible outcomes. A party could win over 100% or less than 0% in a certain seat with a large swing.
Nevertheless, the model has proven as accurate as almost any over recent decades. It was the preferred model of the BBC’s legendary David Butler, who first took the nation through election night results in 1950, and appeared by David Dimbleby’s side as recently as 2010.
But under a four-party political system, we prefer Electoral Calculus’s Strong Transition Model (STM). This attempts to probabilistically model the transition of votes. It improves on UNS by preventing illogical outcomes, and accommodates the effect of incumbency in a way UNS cannot.
If a party’s national share of the vote increases, the party gains voters in each seat from the declining parties in proportion to its relative share of the overall increase from other increasing parties.
On a second level, the model divides voters into strong and weak supporters in each seat. Weak supporters will defect to a different party before strong supporters do if the party’s vote declines in that seat. Weak supporters are modelled as simply those who vote for a party up to a limit of 20 per cent of the turnout; strong voters are any voters beyond that. Thus, parties who currently hold a seat are more likely to have strong supporters, and will lose fewer votes if their national polls fall.
Electoral Calculus has, on average, estimated the parties to within 15 (Lib Dems), 20 (Tories) and 23 (Labour) seats of their actual results in the four elections since 1997.
You can toggle between this model and uniform swing on May2015. They give very similar results. But to try and improve on both, we have integrated Lord Ashcroft’s marginal polls into May2015’s new election-forecasting machine.
Final vote shares are formed for each seat, and then the projected winner is calculated. Our system updates a live table of all 650 seats in the UK, categorised by the winning party, and sorted by their majority.
You can read a full explanation of the model on the Electoral Calculus website.