Jason Duke, Founder, Kronaxis
Tag: Research
Reform UK is about to sweep English local government.
That is the headline prediction from our synthetic persona panel, and this post explains how we got there, what we got right in validation, what we got wrong, and why you should take it seriously. We are publishing predictions for the first 20 councils contesting the 7 May 2026 local elections, with the full 136-council prediction set coming on 1 May. After the results, we will publish a complete accuracy report.
The Experiment
We built 65,000 synthetic UK voters. One hundred per parliamentary constituency, each generated from Census 2021 data and weighted to match the real demographic profile of their area. Every persona has a full identity: age, gender, ethnicity, education, occupation, income, housing type, household composition. But the identity is not the interesting part. What makes these personas useful for prediction is that each one also has a psychological profile, a political history, a financial situation, and a set of beliefs that interact with each other.
The psychological profile uses DYNAMICS-8, a personality model we developed specifically for synthetic population simulation. It measures eight dimensions on a 0 to 1 scale: Discipline, Yielding, Novelty, Acuity, Mercuriality, Impulsivity, Candour, and Sociability. These are not abstract constructs. A high-Discipline persona is organised, habitual, and follows through on commitments. They are also more likely to actually vote in a low turnout council election. A high-Novelty persona is open to new experiences and ideas. They are also more likely to switch parties. A high-Yielding persona is sensitive to social pressure. They are more susceptible to protest vote momentum. Each dimension interacts with demographics, economic circumstances, and political context to shape how a persona responds when you ask them a question.
The political history covers who they voted for in 2019 and 2024, and why. It also captures how their political views have shifted since the 2024 general election, filtered through their personal circumstances: the cost of living, housing costs, immigration, the NHS, Partygate, the Truss mini-budget, the Southport riots. A 62-year-old retired steelworker in Scunthorpe who voted Labour his entire life remembers the winter fuel payment cut differently from a 28-year-old software developer in Bristol who is angry about Gaza.
We asked each persona how they would vote in their local council election on 7 May 2026. We gave them their ward context, the parties standing, and current national political conditions. They responded with a vote, a government satisfaction rating, a confidence level, and a one-sentence explanation of their reasoning. We then aggregated the raw responses through a multilayer correction pipeline that adjusts for turnout, incumbency, protest dynamics, and known model biases.
Validation Against Real By-Elections
Before predicting the future, we tested against the past. Ten real council by-elections held across England in March 2026. Nobody polls these contests. Turnout is 25 to 30 per cent. The results are driven by local factors, protest sentiment, and candidate quality that no national model can observe. They are, in short, a genuinely difficult prediction target.
Our first attempt was bad. Version 1 of the pipeline took 30 random personas from the region, asked them how they would vote, and aggregated the answers. It predicted the winner correctly in 1 out of 10 wards. Mean absolute error: 23.7 percentage points per party. The model defaulted to Labour or Conservative everywhere because it had no ward level context and no personality driven voting behaviour.
Nine iterations later, V9 predicted 6 of 8 testable by-election winners (75%) with 7.0 percentage point mean absolute error. Two wards were excluded for parse failures. That is approaching the accuracy that MRP models achieve at constituency level in general elections, in a domain where no polling data exists.
The best predictions were encouraging. In Brumby (North Lincolnshire), we called Reform UK as the winner with a 6.1pp error. In Aigburth (Liverpool), we called Green with a 5.3pp error. In Gorton South (Manchester), we called Labour with 5.2pp error. In Sleaford (North Kesteven), we correctly identified Reform UK as the winner after earlier versions had wrongly called it for Conservative.
The failures are equally instructive. In Zetland (Redcar and Cleveland), we predicted Reform UK but the Liberal Democrats won on 48.2% of the vote. That was a strong local campaign by a well known candidate in a ward where the Lib Dems had no historical structural advantage. No demographic model can see that. In Stanford (Vale of White Horse), we predicted Liberal Democrat in an affluent Southern ward, but it is rural and agricultural, and traditional Conservative loyalty held. Our model reads "affluent South" as Lib Dem territory because that is the dominant national pattern. Stanford is the exception.
The most important discovery during validation was systematic: the model consistently over-predicts Reform UK by approximately 10 percentage points and under-predicts the Liberal Democrats by approximately 7 percentage points. Applying a simple linear correction based on backtesting improved winner accuracy from 50% to 75%. That calibration correction is baked into the predictions below.
We are being transparent about the limitation: the calibration was derived from the same by-elections used for evaluation. We do not yet know whether it generalises. The 7 May results will answer that question.
The Prediction for 7 May
We ran the full V9 pipeline against the first 20 councils contesting the 7 May local elections. Every council in this batch is currently Labour-held. The results paint a stark picture.
National Headline (Across 20 Councils)
| Party | Predicted Vote Share |
|---|---|
| Reform UK | 39.6% |
| Labour | 23.2% |
| Liberal Democrat | 16.9% |
| Green | 12.7% |
| Conservative | 7.7% |
Reform UK is predicted to win 18 of the 20 councils. Labour holds Manchester. Bristol goes Green. The Conservatives do not win a single council.
Full 20-Council Prediction Table
| Council | Region | Winner | Reform | Lab | LD | Grn | Con | Margin | Conf |
|---|---|---|---|---|---|---|---|---|---|
| Barnsley | Yorkshire & Humber | Reform UK | 37.2% | 20.9% | 15.7% | 18.3% | 7.9% | 16.4pp | High |
| Birmingham | West Midlands | Reform UK | 44.7% | 21.8% | 14.5% | 9.4% | 9.6% | 22.9pp | High |
| Bradford | Yorkshire & Humber | Reform UK | 41.5% | 28.6% | 15.2% | 10.8% | 3.8% | 12.9pp | Medium |
| Bristol | South West | Green | 22.3% | 11.7% | 26.6% | 29.5% | 10.0% | 3.0pp | Low |
| Coventry | West Midlands | Reform UK | 46.6% | 23.8% | 12.3% | 9.8% | 7.5% | 22.7pp | High |
| Derby | East Midlands | Reform UK | 40.0% | 21.2% | 19.8% | 6.9% | 12.1% | 18.7pp | High |
| Doncaster | Yorkshire & Humber | Reform UK | 54.3% | 22.1% | 11.3% | 5.7% | 6.7% | 32.2pp | High |
| Leeds | Yorkshire & Humber | Reform UK | 30.0% | 23.3% | 21.7% | 15.0% | 10.1% | 6.7pp | Low |
| Leicester | East Midlands | Reform UK | 37.0% | 18.1% | 17.3% | 15.5% | 12.1% | 18.9pp | High |
| Liverpool | North West | Reform UK | 44.3% | 31.8% | 8.5% | 12.2% | 3.2% | 12.5pp | Medium |
| Manchester | North West | Labour | 27.3% | 29.2% | 15.4% | 23.5% | 4.5% | 1.8pp | Low |
| Newcastle | North East | Reform UK | 42.2% | 33.5% | 8.3% | 12.7% | 3.3% | 8.6pp | Medium |
| Nottingham | East Midlands | Reform UK | 38.1% | 23.0% | 19.1% | 12.5% | 7.3% | 15.1pp | High |
| Plymouth | South West | Reform UK | 32.4% | 13.9% | 22.9% | 13.6% | 17.2% | 9.5pp | Medium |
| Sheffield | Yorkshire & Humber | Reform UK | 50.1% | 22.1% | 16.0% | 7.5% | 4.4% | 28.0pp | High |
| Southampton | South East | Reform UK | 32.5% | 20.3% | 28.6% | 10.2% | 8.4% | 3.9pp | Low |
| Stockport | North West | Reform UK | 36.7% | 23.4% | 21.0% | 11.7% | 7.3% | 13.4pp | Medium |
| Sunderland | North East | Reform UK | 52.2% | 22.3% | 11.0% | 8.7% | 5.8% | 29.9pp | High |
| Wigan | North West | Reform UK | 44.8% | 27.3% | 14.0% | 10.8% | 3.1% | 17.4pp | High |
| Wolverhampton | West Midlands | Reform UK | 38.8% | 25.9% | 17.1% | 8.8% | 9.4% | 12.8pp | Medium |
A few specific calls worth flagging:
Doncaster and Sunderland go Reform 50%+. These are post industrial, heavily Leave-voting areas with deep anti-government sentiment. The model predicts Reform UK will win outright majorities, with margins above 30 points over the nearest challenger. If those numbers are even directionally correct, it represents a fundamental realignment in English local politics.
Bristol goes Green. The only non-Reform winner outside Manchester. Bristol has a young, educated, progressive population, and the model puts the Greens narrowly ahead of the Lib Dems (29.5% vs 26.6%) with Reform UK a distant third at 22.3%. This is a low-confidence call with a 3-point margin.
Manchester holds for Labour, but barely. Labour leads Reform UK by just 1.8 points (29.2% vs 27.3%), with the Greens on 23.5%. This is a three way marginal. If the model is even slightly wrong on Reform UK calibration, Manchester could flip.
Leeds is a three way contest. Reform UK leads on 30.0%, but Labour (23.3%), the Lib Dems (21.7%), and the Greens (15.0%) are all within striking distance. This is a city where demographics vary enormously by ward, and aggregate council level numbers may hide significant variation.
Birmingham, Coventry, and Sheffield look like Reform blowouts. All three show Reform UK above 44%, with margins above 22 points. If these predictions hold even approximately, they would represent the most dramatic council results in modern English political history.
A Note on Calibration
These numbers include our calibration adjustment, which reduces raw Reform UK predictions by approximately 10 percentage points based on our by-election backtesting. Without calibration, the raw model output puts Reform UK even higher. We believe the calibrated numbers are more realistic, but we are stating plainly that even after calibration, Reform UK may be over-predicted. Our by-election validation showed the calibration improved accuracy substantially, but we only have 8 testable data points. The 7 May results across hundreds of wards will tell us whether the calibration generalises.
What This Means
Four findings stand out from the prediction data.
Reform UK is not a protest vote that will evaporate. The synthetic personas show consistent Reform preference across multiple demographics and personality types. The strongest predictors are personality driven: low Novelty (traditional, routine-oriented), low Yielding (resistant to social pressure, unlikely to be talked out of their position), and high Discipline (organised, likely to actually turn out and vote). These are not people who are casually angry. They have made a considered decision to switch allegiance, often from a lifetime of Labour or Conservative voting, and the model shows no mechanism by which they switch back before 7 May. The economic drivers are reinforcing: cost of living pressure, housing costs, stagnant wages. Reform UK is offering simple answers to material problems, and the personas who are drawn to those answers are the ones who vote in local elections.
The Conservative collapse is real. Predicted at 7.7% across these 20 councils. Not a single council win. Not a single second-place finish. The Conservative vote has been squeezed from both directions: Reform UK takes their right flank, the Liberal Democrats take their affluent suburban voters, and the Greens take their younger, environmentally conscious supporters. What remains is a core of older, wealthy, high-Discipline voters in rural areas that are not well represented in this batch of metropolitan and unitary councils. The county council elections on 7 May may tell a different Conservative story, but in England's cities and towns, the party is heading for single digits.
Green is the anti-Labour protest. In progressive urban areas, the Green Party is doing to Labour what Reform UK is doing to the Conservatives. Bristol is the clearest example: the model predicts Green as the largest party, drawing support from younger, higher-educated, high-Novelty personas who voted Labour in 2024 but are disillusioned over Gaza, climate policy, and perceived centrism. The Greens are strongest where Reform is weakest, and vice versa. England's political map is splitting along a personality axis: traditional, security-oriented voters moving to Reform; progressive, change-oriented voters moving to Green. Labour and the Conservatives are being hollowed out from opposite ends.
The Liberal Democrats are strong in affluent areas but invisible in northern cities. The model predicts Lib Dem vote shares above 20% in Southampton, Leeds, Plymouth, Stockport, and Bristol, but below 10% in Liverpool, Newcastle, and Sunderland. This is consistent with the party's real electoral pattern: strong local campaigns in places where they have an established base, but unable to break through in areas dominated by Labour-Reform dynamics. The Lib Dem story on 7 May will be written in the county councils and district councils that are not in this batch, not in the metropolitan boroughs.
Methodology in Plain English
For readers who have not seen our earlier work, here is how the prediction system works.
65,000 personas. Generated from Census 2021 constituency level data. One hundred per parliamentary constituency. Each has demographics matching their local area, a DYNAMICS-8 personality profile, a political history, financial circumstances, and beliefs. The DYNAMICS-8 model captures eight personality dimensions that drive how people respond to situations: whether they follow rules or question authority, whether they seek novelty or prefer routine, whether they are influenced by social pressure or resist it. These traits predict behaviour that demographics alone cannot explain, such as why two people with identical income, age, and education vote differently.
Ward-level matching. For each council, we build a panel of personas whose demographics match the council's profile. A panel for Sunderland draws heavily from older, working class, post industrial personas. A panel for Bristol draws from younger, educated, urban professionals. The matching is automatic and uses the same function validated against our by-election results.
Multi-layer correction. Raw vote shares from persona responses go through seven correction layers: turnout modelling (personality driven, because organised people vote more in low-salience elections), government satisfaction weighting (angry voters are over-represented in local elections), incumbency adjustment (sitting parties have name recognition), protest vote correction (local elections punish the governing party), statistical ensemble blending, shy voter adjustment, and historical precedent blending. Each layer was individually validated against the by-election results.
Calibration from backtesting. The raw model over-predicts Reform UK by approximately 10 percentage points and under-predicts the Liberal Democrats by approximately 7 points. We apply a linear correction derived from our by-election validation. This is the single biggest source of accuracy improvement in the pipeline.
Inference via Kronaxis Imprint. All persona responses are generated by Kronaxis Imprint, our sovereign 27-billion parameter language model, fine tuned with a persona response adapter. Each council prediction requires approximately 30 inference calls and takes around 15 minutes.
The full methodology, including the mathematical specification of each correction layer and the complete by-election validation results, is published in our research paper.
What Comes Next
1 May 2026: Full prediction for all 136 councils contesting the 7 May elections, published at kronaxis.co.uk. This will include county councils and district councils, which may show different patterns (stronger Conservative and Lib Dem performance) than the metropolitan boroughs in this batch.
Ward-level predictions: Where ward level demographic data is available, we will publish ward level predictions alongside the council aggregates.
8 May 2026 onwards: As results come in, we will publish a complete accuracy report comparing our predictions against actual outcomes. Every correct call and every wrong call, with analysis of where the model succeeded and where it failed. No cherrypicking.
Ongoing: The same technology that produces these predictions is available through Kronaxis Panel Studio for market research, product testing, and policy analysis. If synthetic personas can predict elections, the hardest possible test, they can predict consumer behaviour, policy responses, and market dynamics. Sign up for a free API key at kronaxis.co.uk/register and build your own panel.
The DYNAMICS-8 personality model specification is published under CC BY 4.0 at kronaxis.co.uk/research. The by-election validation data and prediction methodology are fully open.
Previous post: Can Synthetic Personas Predict Elections? We Tested Against Real By-Elections
Research paper: Predicting UK Council By-Elections Using Synthetic Persona Panels
Panel Studio: panel.kronaxis.co.uk