How Deltapoll Measures Voting Intention 

May 2026

Deltapoll has opted not to publish voting intention polls since the last General Election in 2024, not so much because of the level of error obtained in our final poll at that election, which was to our mind acceptable, but because we wanted to take time out to properly review and test potential new ideas and methods safe in the knowledge that with a large majority for the government, time would be on our side.

Different opinion polling companies can and should utilise different methodological interventions in producing headline polling figures for a number of reasons. Firstly, outlier prediction polls have been the most accurate final polls on a number of occasions over the last couple of decades. Reputational safety implied by actual or accidental herding is not something we could ever subscribe to – our only aspiration is to produce the most accurate prediction poll we possibly can.

Secondly, new techniques and new ideas encourage plurality in methodological approach. As pollsters who have arguably innovated for good and bad down the years, we strongly believe that the best pollsters go out of their way to challenge assumptions and include new processes in the pursuit of accuracy, if they think them worthwhile.

We have found two ideas that we think worthwhile, and we are delighted to relaunch our headline voting intentions from now onward with these innovations at the heart of our approach.

The first of these methodological interventions involves exponentially weighting moving estimates. This method involves combining multiple waves of polling rather than only using the most recent. This is because polling with a sample size of around 2,000 inevitably leaves gaps or small samples for a number of major demographic groups. Current methods allow those demographic groups to suffer from missingness, creating distorted or ‘noisy’ estimates wave-on-wave. With our view being that old data is better than missing data, we have resolved to supplement deficient demographics by drawing from previous waves.

We do this by combining data from multiple survey waves and then applying an exponential weight towards more recent waves. When applying rake weights, this allows the weight scheme to upweight a handful of respondents from previous waves who meet demographic deficits, without the overall representation of old data within the sample exceeding a certain (low) threshold.
The overall effect of this is that we can weight to higher-resolution targets, including complex interactions of demographic variables, ensuring a balanced and representative sample even within specific regions, income brackets, age brackets, etc.

The second intervention introduces a new method for modelling turnout. This involves accounting for varying reliability and meaning when respondents report their likelihood to vote.

Respondents from backgrounds and social circles which are highly politically engaged are likely to underestimate their likelihood to vote, simply because they think they are probably less likely to vote than their peers. Similarly, others may overestimate their likelihood because they are more likely to vote than their peers. Personality factors to do with conscientiousness or agreeableness may also be correlated with variations in accuracy when estimating one’s own likelihood to vote.

As such, we take British Election Study data from before and after the 2024 election, and use logistic regression to model the effect of different demographics on whether they voted (declared after the election). We then predict the voting probability for our demographic target groups, down-weighting based on vote likelihood. Our demographic targets are therefore representative of the expected population of voters, rather than the population as a whole. We intend to develop these turnout methods further so that we can incorporate political variables that may be correlated with vote likelihood, so that our modelling of turnout is more responsive to changes.

We then use this model to predict our respondents’ true likelihood to vote based on a combination of their demographics and their self-reported likelihood to vote, and re-weight respondents’ voting intention based on this.

We believe these two new methods will improve the reliability and accuracy of our vote intention figures. The effect of these measures often does not have material impact on headlines numbers in comparison to our old method, but can when, or if, heavier lifting is required to deal with sample imbalances.

Over recent waves, the general, but not consistent impact of these two new measures has been slightly higher party support for the Conservatives, Labour, and the Liberal Democrats, and slightly lower party support for the Green Party. Overall, the main effect is a large reduction in noise between estimates without artificially smoothing or reducing the responsiveness of our estimates to recent changes in public opinion. We are comfortable with these outcomes, and hope they will continue to serve Deltapoll’s reputation for high quality, trustworthy, and reputable polling.