To see the election forecast as it appeared on election day (7 May 2015), follow this link.
To see our report evaluating our forecast, follow this link.
Before the election, we wrote several posts and articles about our forecasting methods, in addition to what was available on this site .
25 March 2015 How our model at electionforecast.co.uk works Significance Magazine
30 March 2015 How Our U.K. Election Forecasting Model Works FiveThirtyEight.com
6 May 2015 Are we right about the UK General Election? FiveThirtyEight.com
7 May 2015 Combining national and constituency polling for forecasting A 2500 word article for a forthcoming special issue of the journal Electoral Studies outlining our forecasting method.
We posted continuously updating election-night projections, starting from our pre-election forecast. These only include Great Britain, as the limited number of seats in Northern Ireland made this sort of projection model difficult to implement there. Because our pre-election forecast had poor coverage of the ultimate result, it took a while for the model to catch up to the final result as the night progressed and the model was over-confident until late in the night. This is what we would expect to see given Bayesian updating from bad priors. Here are links to snapshots of what the projection looked like at various points in the evening:
Many sites report the final Conservative total as 331, because they include the speaker of the House of Commons as a Conservative, whereas we include him as Other.
We have written several posts and articles since the election about what went wrong (and what went right) with our forecasts.
8 May 2015 What we got wrong in our UK General Election Model Initial assessment of what went wrong at FiveThirtyEight.com
11 May 2015 What else we got wrong in our UK General Election Model Update to initial assessment of what went wrong at FiveThirtyEight.com
18 May 2015 If not polls, then betting markets? Comparison to betting market predictions at LSE General Election Blog
21 May 2015 Evaluating our 2015 General Election Forecast
We have posted a detailed report explaining the issues with our forecasting model and some of the lessons we have learned from them.
We have posted a Github Repository with an amended version of the source code used to generate the forecasts for electionforecast.co.uk
The code has been amended in three respects:
In the first and second case, we have omitted code because we are unable to redistribute the underlying data (YouGov campaign data for the constituency polling; shapefiles for the mapping).
Although the project contains data on YouGov constituency subsamples (included alongside information from Ashcroft polls, and found in working/constituency), these have been reweighted by us to match Census demographics, and cannot be considered "raw data".
We thank YouGov for giving us access to their data, and would also like to thank Anthony Wells for permission to include his data on historical vote intention, which can be found at ukpollingreport.co.uk