Forecasting the 2010 UK election

Resolver Systems' co-founder Robert Smithson is a keen observer of the UK political scene, and in his spare time he looks after the technical side of the immensely popular PoliticalBetting.com. He thinks the way we forecast UK election results is no longer effective, and using our programmable spreadsheet, Resolver One, he's built a model based on Voting Intentions Predictive Analysis (VIPA) — which gets surprising results.

While almost all commentators are expecting the Conservative Party to win from the incumbent Labour Party, Smithson's VIPA model predicts something more like a landslide — a change comparable to Labour's own 1997 win.

VIPA has been picking up interest in the press, with articles in the Financial Times' Westminster blog and in The Register. To see what the fuss is about...

The problem with the Swing

Most existing election forecasts in the UK use techniques based on a Unified National Swing. This uses poll data to determine how each party's share of the vote has changed since the last election, and then applies that "swing" to each constituency in the country to see who will win it this time around.

For example, using recent poll data we get:

So, let's take a typical constituency, Bedford. Leaving out Plaid Cymru and the Scottish National Party, neither of which fielded candidates in Bedford, we can apply the swing to 2005's results to get a quick and easy forecast of how people would vote there if an election was held now:

So this model would suggest that the Conservatives will win Bedford from Labour. Repeat that process for every constituency in the country, and you can work out how many seats you expect each party to win.

The problem with forecasting like this is that it doesn't take account of where the votes "come from" — for example, if the Conservatives get more voters, were they previously Liberal Democrats, or Labour, or non-voters? This matters because people are far more likely to move in certain directions; for example, recent data from ICM suggests that people who voted Liberal Democrat in 2005 and decide to vote differently now are much more likely to vote for the Conservatives than they are to not vote at all. However, people who voted for the Scottish National Party are much more likely to stay home than they are to vote for the Conservatives.

So, if you imagine a marginal constituency where the Conservatives and Labour are the two largest parties, if there are lots of Liberal Democrat voters there who decide to vote differently it will benefit the Conservatives much more than it would if the third party in that seat was the Scottish National Party. Election forecasters know of this effect, and it is often discussed in election night broadcasts, but it rarely gets into the models used to predict the results.

The problem is that the Unified National Swing does not work very well when there are more than two parties. It was developed in the 1950s, when Labour and the Conservatives took 95% of the vote, but now that there are effective third parties in different parts of the country, it has become considerably less effective because it fails to take account of these local differences.

The solution: VIPA

Smithson's solution to the problem is simple: the ICM data, suitably adjusted, tells you how likely people are likely to move from party to party. Given this detailed voting intentions information, a list of the election results from 2005, a spreadsheet that's easy to program, and a lot of hard work, he was able to build a model that works out on a per-constituency basis how people were likely to vote, allowing for third parties and regional differences. This is Voting Intentions Predictive Analysis.

...and the forecast:

Most recent forecasts suggest that the Conservatives will win a small majority of seats in the 2010 election; Labour will lose many seats, and the Liberal Democrats will lose one or two

VIPA disagrees; once again, it predicts:

A landslide to the Conservatives, and in addition a very suprising fall in the number of Liberal Democrat seats.

Try it out for yourself!

If you're using a PC with Microsoft Windows XP or Vista, you can try out the full election forecasting model yourself:

1. Click here to download Resolver One, the spreadsheet program that the model uses.
   • If you're using Internet Explorer, when asked if you want to run or save the file, select "Run"
   • If you're using Firefox, wait until the file has downloaded and then double-click on it in the "Downloads" window.
2. Once it is installing, it will ask you whether you want to use the "Default" or "Custom" options. Choose "Default".
3. Click here to download the model itself.
4. Double-click on the model file to start it, or start Resolver One and use the "Open" option on the "File" menu. You will see the main page of the model, which looks like this:
   
5. At the top of the spreadsheet, there is a table showing voting intentions. We'll explain this in detail later.
6. Further down is a table showing the 2005 Results, which should be self-explanatory.
7. Further down still (you'll have to scroll a little) there's the 2010 Forecasts section. Initially this is empty, so...
8. Click the "Execute forecasts" button to see the number of seats each party will get.
   
   ...and the results will appear:
9. Perhaps a chart would help? If you look beneath the sheet, you will see a number of tabs, each of which is for a different sheet (just as in other spreadsheet programs). Click on the one called "Pie Charts (RESULT)"...
   
   ...and you will see pie charts showing the 2005 and the forecast results (you may need to make the window larger to see it all):
   
10. Let's go back and change some of the inputs to the model to see how that changes the output. Click on the first tab (called "Summary"), and scroll up to to the top of the sheet, back to the voting intentions grid. This grid has how people voted in 2005 from left to right, and how they say they will vote from top to bottom. So, for example, to see how likely a voter who chose the Liberal Democrats in 2005 is to choose the Labour Party next time, find the column headed "LD", then go down to the row with header "LAB":
    
    The figures that we have are an adjusted version of the ICM data (we've stripped out the people who didn't give an answer, scaled everyone else up to compensate, and made some adjustments to allow for missing data for the SNP and Plaid Cymru). Let's try changing them temporarily to show that happens (perhaps you may want to change them permanently later, if for example you find new data). For now, change the figure we found above to 8.5%, and then go down and hit the "Execute forecasts" button once again. You can see that it gives noticably different results — the Conservative lead is reduced:
    

...and that's it for now! You now know how VIPA analysis predicts the election results for the UK 2010 General Election.

Next steps

Monte Carlo analysis

This is useful for people who want to make spread bets on the election, as it gives you the VIPA model's predictions of the probabilities of a range of outcomes rather than a single set of numbers. You can read more about the Monte Carlo analysis feature here.

Advanced parameters

The VIPA model takes account of incumbent bias and the different political make-up of the different countries of the UK; you can configure these options on the "Advanced" sheet.

Incumbent bias: Historically, current MPs have been more likely to win elections than the polling numbers would predict. And many Liberal Democrats claim that MPs from their party benefit from this effect even more than members from other parties. On the other hand, the recent expenses scandal might have biased voters against incumbents. The VIPA model allows you to specify percentage amounts for the incumbency bias for MPs in general, and a separate one for the Liberal Democrats, at the bottom of the "Advanced" tab.

Per-country voting intentions. The voting intentions grid you enter on the "Summary" sheet is for the UK as a whole. However, it masks obvious differences between England, Scotland and Wales. For example, if the national intentions grid suggesteds a shift from, say, Labour to the SNP, then this is an average number and cannot just be applied nationwide, because this would lead to both smaller shifts to the SNP than are reasonable in Scotland (because the shift is an average across the whole of the UK) and to small shifts to the SNP in England and Wales, which is not realistic. In addition, English, Scottish and Welsh voters switch parties in very different ways; in particular, national parties in Scotland and Wales are attractive protest votes. Different voting intentions grids are needed to express this. (Of course, one could go futher and require even more detailed local voting intention analysis, on the basis that Tyneside is rather different to Henley-on-Thames. Unfortunately we don't have the data for this.)

The way the VIPA fixes this country problem is twofold:

1. On every analysis you run, it starts off by taking the nationwide grid, and inferring from it separate grids for England, Scotland and Wales, making various changes — for example, removing the lines for the SNP from the grids for England and Wales. This helps address both problems, and while the values it infers are never entirely perfect they are much better than simply using the same national figures for every constituency would be.
2. In addition, you can override these per-country grids yourself in the "Advanced" worksheet. So if you have better data than the data that VIPA has inferred, you can enter it there.

Finally, if you just want a simple prediction based on the national numbers with no incumbency bonus, you can switch off the advanced parameters by setting the "Use Advanced Settings" at the bottom of the "Summary" sheet to "No". We'll write more about VIPA soon, so if you want to hear more, keep an eye on our news blog.

If in the meantime you'd like to ask questions about the model, or tell us what it does wrong, you can post a comment in our forums.