UK Election Forecasting - A detailed explanation of the techniques used by UK-Elect
Please note that UK-Elect is highly configurable and can take account of many factors, e.g. by-elections, local constituency opinion polls, tactical voting, the incumbency effect etc. the ability to adjust predicted percentages depending on the time until an election, and other factors not mentioned below. This article will be expanded to include them soon.
For theoretical information about, and discussion on, the advantages and disadvantages of UNS, Proportional and other election forecasting methods see our Election Forecasting Theory and Controversial Discussions page.
For the latest UK-Elect manual, which may be more up to date than this page and which includes screen shots showing the configuration of the forecasting options, see the UK-Elect Manual
How UK-Elect works
The target party support levels are usually input as regional (Scotland, Wales, London etc.) or GB/UK opinion poll percentages. (They can however be calculated based on by-election results, constituency opinion polls, or early election night results, or on a specified past election, or some combination of these. The way in which new results can be converted into national or regional party support levels is also highly configurable.)
The difference between the target support levels for the political parties and the support at the previous election is applied to each constituency in turn. The way in which the change in percentage is applied is highly configurable, e.g. it can be applied in a uniform way or a proportional way, or a mix. In addition, behaviour can be tailored to party size, better reflecting differences in behaviour whether the parties votes are more concentrated or evenly spread across the country.
The strength of public feeling about Brexit can be used as a factor in calculations, adjusting results based on the 2018 European Referendum results (estimated on a constituency basis) and how strongly parties are perceived to be associated with "Leave" and "Remain".
Tactical voting and incumbency can be considered if desired and are highly configurable (see the forecasting election results section in the manual). As an example, the incumbency settings can configured on a party or candidate basis, taking account of whether the incumbent was previously a first time winner (as some academic studies have suggested that the greatest incumbency advantage is associated with people who won their seats at the preceding election). The tactical voting settings can optionally be configured to use defined party vote-transfer percentages - e.g to take account of the preferred "2nd choice" parties of voters.
The forecast results can be blended with previous by-election results and constituency opinion polls, which can be adjusted to reduce weighting according to the age of the polls or the change in national or regional levels of party support since the date of the by-election or opinion poll fieldwork.
Each forecast can be done on a regional (up to 6 regions) or national basis, or these can be combined. Where one forecasting area forms part of another (e.g. London forms part of England, England and Wales, GB, and the UK) the smaller area is forecast first, and the forecast results then incorporated into the forecast for the larger area. (Percentages are checked against the targets at every stage, so, for example, the previously forecast London results would affect the calculations for England, and the English results would in turn affect the overall GB calculations.)
The forecasting methods supported in UK-Elect are:
This method assumes the same uniform swing between parties takes place in every constituency - e.g. in every constituency the Tories lose 5% of the total vote and Labour gain 5%.
Quick. Easiest method to understand - a 5% swing to Labour from the Conservatives would gain the party all those seats where it is less than 10% behind the Tories (the Tory vote would be assumed to fall by 5% and the Labour vote rise by 5%). It is commonly used by the media and has a generally good reputation among most psephologists and political scientists because analysis of past elections has shown that the national swing often is surprisingly uniform across a range of different constituency types and almost irrespective of the starting percentages. In addition it can be relatively easily tweaked to improve accuracy provided that a good understanding exists of voter behaviour in particular types of constituencies, although it may be noted that the evidence of past elections and their forecasts is that such tweaks can sometimes be counterproductive. They are usually based on behaviour at the preceding elections - which can change.
If used in its pure "uniform" form it is likely to be less accurate than any method that correctly makes allowances for human behaviour. Swings are rarely completely uniform even within a region. In addition, it is extremely poor at forecasting the votes in cases where a party vote is very low or high (if a party has only 4% of the vote in a constituency, it can't lose 5%). It is extremely difficult to explain a forecast where one or more of the parties are predicted to have a negative number of votes in some constituencies!
Uniform Swing is a reasonable choice, which is rarely very inaccurate, and is traditionally a good safe choice whenever it is important that predictions agree with those made by most of the media (although the proportional-loss based combined method was also commonly used for the 2010 election - see below). As with other methods it can be modified to try to try to better take account of voter behaviour - although not without risk. It is still probably the most commonly used technique, both for pre-election forecasting and also as a component of election-night exit-poll based forecasting.
A common alternative to Uniform National Swing that exists in several variations, this method is just as easy to calculate. It simply uses the ratio between a party's current opinion poll percentages and its percentages at the previous election.
For example, if PartyX got 20% last time, and is currently at 50% in the polls, then the multiplier will be 50/20 = 2.5. This multiplier is then applied evenly to PartyX's votes in all constituencies. If the resulting vote total is too large or too small then a further multiplication is applied to all the party’s votes in the constituency to achieve the correct total number of votes. If, as a result of this first pass the national vote totals do not match the target vote percentages then the ratios are adjusted and a further forecast made. This process is repeated until the vote target is achieved (or the maximum number of iterations reached). It is also possible to modify this system to take account of different characteristics of seats, and use different ratios for each of these characteristics.
Proved to be more accurate than UNS when tried on some past elections. Doesn't produce negative vote totals. Lacks some of the disadvantages of the other methods. As with other proportional systems, it can be modified to try to protect against some of the disadvantages discussed below – e.g. to prevent wholesale predicted loss of Liberal Democrat seats whenever their vote declines significantly nationally! This is usually done by introducing some protective mechanism that applies to seats where their vote is already strong - e.g. some sort of threshold above which their votes are protected, so that most of the predicted vote losses come from seats where they are weak. (This, in turn, sometimes leads however to extremely low – almost zero - vote totals being forecast in their weaker seats, something that in reality more commonly happens at by-elections than General Elections).
It is likely to be less accurate than any method that correctly makes allowances for human behaviour. Swings are rarely in the same ratio in every seat even within a region. When wrong, it can be far more inaccurate than the Uniform National Swing method. Many methods based on this system proved to be significantly less accurate than the UNS method at forecasting the 2010 General Election. As with other proportional systems, it also tends to exaggerate losses where a party declines significantly nationwide – for example, when the Liberal Democrat vote declines significantly it quite often forecasts zero seats for them – even though past experience of the Liberal Democrat and Liberal voting history shows that many Liberal / Liberal Democrat held seats are very resilient to even very sharp national vote declines.
A worthwhile method - it quite often produces slightly better results than UNS, but don't take it too seriously - when wrong, it can be very wrong - and generally it was wrong in 2010.
An alternative to UNS that gained noticeably in popularity before the 2010 General Election, this proportional-loss based method combines aspects of the other two. In each constituency it simply uses the ratio between a party's current opinion poll percentages and its percentages at the previous election to calculate votes lost and then re-adds those lost votes to the parties which gain votes nationally in proportion to their uniform national swing percentages.
For example, if PartyX got 50% last time, and is currently at 25% in the polls, then the multiplier will be 25/50 = 0.5. This multiplier is then applied evenly to PartyX's votes in all constituencies. In each constituency, the votes lost are then added to those parties which gained votes nationally in proportion to their national percentage gain. The result is that the total number of votes stays the same as at the previous election in every constituency, and that unlike uniform swing, negative vote totals are not possible.
Proved to be slightly more accurate than UNS when tried on some past elections. Doesn't produce negative vote totals. As with other proportional system, it can be modified to try to protect against some of the disadvantages discussed below and in the "Proportional" section.
It is likely to be less accurate than any method that correctly makes allowances for human behaviour. Assumes parties can only gain votes in any seat if others lose them nationally - as an example, if there were only 3 parties and PartyX gained 20% nationally, PartyY stayed the same, and PartyZ lost 20%, then the forecast votes for PartyX and PartyY would remain unchanged in every seat where PartyZ didn't stand (and would remain substantially the same wherever it had a low total). In other words, it doesn't cope well with direct swings between the gaining parties. As with other proportional systems, in its pure form it also tends to exaggerate losses where a party declines significantly nationwide - for example, when the Liberal Democrat vote declines it quite often forecasts zero seats for them
A worthwhile method - it can sometimes produce slightly better results than Uniform National Swing, although it may also be a lot worse. May also be useful to replicate some media and online forecasts. Be aware that in its un-modified form it may produce unrealistic results in situations where there is a major change in votes for a nationally significant party.
This variation of Proportional Loss is also called the "Strong Transition Model" and has been used by some online forecasting sites such as Martin Baxter's electoralcalculus.co.uk site. It differs only in that a threshold can be set (e.g. 20%) which (in Martin's terminology) divides voters into "strong" and "weak" categories. Below the threshold voters are assumed to be "weak" voters, more likely to switch parties. Accordingly, when a parties vote declines it is assumed that the first losses will come from this portion of its votes (i.e. more from voters in seats where it is weak), and "weak" and "strong" ratios are calculated separately (with the "strong" ratio usually being 1.0 unless all the "weak" ratio is calculated as 0.0 - i.e. all the weak voters are assumed to have defected.
Compared to Proportional Loss it much better reflects the situation when the votes for a party such as the Liberal Democrats (and the Liberals in the past) decline - despite the reduction in votes they tend to retain many of their seats. It was one of the better performing methods in 2010.
Although better than Proportional Loss at coping with significant vote declines, it otherwise shares the same disadvantages as other proportional methods. The correct threshold to use is also not always obvious,and may vary from election to election.
A worthwhile method that does seem to be an improvement on "pure" Proportional Loss - it can sometimes produce slightly better results than Uniform National Swing, although it may also be a lot worse. May also be useful to replicate some media and online forecasts..
Note: Different versions of UK-Elect normally have different methods associated with them, adapted to the then current electoral circumstances - e.g. the "UK-Elect 12.3 Method" if and when it exists, is likely to be very different from the method described here.
The UK-Elect v11.3 method is based on the Proportional Loss with Threshold method, combined with a variation of UNS. The threshold used is varied depending on the specific configuration (e.g. on the Brexit strength of feeling, a high setting for the Brexit strength of feeling raises the threshold, thus protecting fewer incumbent MPs even where they have a very large majority). It is highly recommended that this method be used with separate regional percentages (at least for Scotland) in addition to national percentages, rather than as a simple national percentage based forecast.
Mostly the same as a blend of the Proportional Loss With Threshold and UNS combined with the preferred UK-Elect defaults. It also contains many adjustments designed to cope better with the political situation from 2019 onwards which is heavily influenced by Brexit issues.
Mostly the same disadvantages as the Proportional Loss System, but with some edges softened. Also, if the importance of Brexit declines, or the party stances on the issue change suddenly, then the Brexit adjustments may prove counter-productive.
A method very much designed for the current (2019) political situation, but with it's adjustemtns to cope better with Brexit-based political volatility untested.
This isn't a single method - it is actually all of them! It allows access to the detailed settings that can be used to define the other forecasting methods, or to create a new one. The configuration dialog for this method provides buttons with the names of the main methods on them, which when pressed set the advanced options to the appropriate state for those methods. Making a custom forecast with the advanced options set appropriately is identical to forecasting by selecting that method from the main forecasting configuration dialog. In addition to the previously mentioned methods there is a UK-Elect "Experimental" method that can be set on the Custom dialog - it is discussed below.
Fully configurable - create your own method and test it against past elections by setting the forecasting base election and forecasting target percentages source options (found on the main forecasting configuration dialog) appropriately.
It is possible to create some very dubious methods, and achieve some very strange results!
Allows users to see how the other methods are configured, and to create new ones - one of which might be the best method for forecasting the next election!
This is a set of Custom method settings, accessed from the Custom forecast method dialog, that is basically the method the developers of UK-Elect are currently experimenting with. It may evolve to become a new method in future, or it may be replaced.
It is used to try out possible improvements, and they may turn out to be real improvements that do improve the accuracy of forecasts.
Because it is "experimental" confidence in any forecasts produced by it is lower than usual.
By definition an experimental method, with the advantages and disadvantages that implies
For further details on UK-Elect forecasting, including screenshots showing the configuration of the advanced custom forecasting options, see the UK-Elect Manual
Subscribe to our forecast email updates here
|©Copyright 2021 UK-Elect|