Predicting UK Council By-Elections Using Synthetic Persona Panels: A DYNAMICS-8 Approach

Jason Duke, Kronaxis Limited

Preprint, 31 March 2026

Abstract

We present a method for predicting UK council by-election outcomes using synthetic persona panels driven by the DYNAMICS-8 personality model. Each persona is a census weighted synthetic individual with 187 demographic, psychographic, and political fields, including an eight-dimension personality profile that modulates voting behaviour. We tested the method against 10 real council by-elections held in March 2026 across England, iterating through nine versions of the prediction pipeline. Our naive baseline (V1) predicted 1 of 10 winners with a mean absolute error (MAE) of 23.7 percentage points per party. After introducing ward level demographic matching, DYNAMICS-8 turnout modelling, incumbency adjustment, protest vote correction, satisfaction weighting, and chain-of-thought prompting, V3 predicted 5 of 10 winners with 13.9pp MAE. Our final version (V9) predicted 6 of 8 testable by-election winners (75%) with 7.0pp mean absolute error, following the discovery that systematic calibration corrections for Reform UK over-prediction and Liberal Democrat under-prediction could be derived from backtesting. We then applied the validated pipeline to predict the 7 May 2026 English local elections across 20 councils using a 65,000-persona constituency level dataset, predicting Reform UK as the dominant force in English local government with an estimated 30% national vote share after calibration. We pre-register the full methodology for application to all 136 councils contesting the 7 May elections.

1. Introduction

Election prediction in the United Kingdom relies on a well-established hierarchy of methods. National opinion polls, conducted by firms such as YouGov, Ipsos, and Savanta, sample 1,000 to 2,000 respondents and produce headline voting intention figures accurate to within 2 to 4 percentage points at the national level. Multi-level regression and post-stratification (MRP) models extend these polls to constituency level estimates by weighting responses against census demographics, achieving useful predictions in general elections.

Council by-elections sit outside this infrastructure entirely. There is no polling for individual ward contests. The wards are too small, the electorate too few, and the commercial interest too low to justify the cost of a traditional survey. Turnout is typically 25 to 30 percent. The results are driven by local factors, protest dynamics, and incumbency effects that national models cannot capture. Yet these by-elections are the most frequent direct measure of public opinion between general elections. In 2025 and 2026, Reform UK's emergence as a force in English local government was visible in by-election results months before it appeared in national polling. Researchers, journalists, and political strategists who want to understand what is happening in English politics before 7 May 2026 have limited tools available.

We propose an alternative approach: synthetic persona panels. Rather than surveying real people, we construct artificial respondents whose demographics, personality profiles, political histories, and memory of recent events are calibrated to census and survey data. We then prompt these personas through a large language model to produce first person voting intention responses, which we aggregate and correct using a multilayer prediction pipeline.

This paper reports results from testing the method against 10 real council by-elections across nine pipeline iterations. The final version (V9) predicted 6 of 8 testable winners with 7.0pp mean absolute error. We then applied the validated pipeline to predict outcomes in 20 councils contesting the 7 May 2026 English local elections. The method works well in certain ward types and fails systematically in others. We report both successes and failures in full.

2. Methodology

2.1 Persona Generation

For the by-election validation, we used a 5,500-persona national dataset. For the 7 May predictions, we scaled to 65,000 personas: 100 per parliamentary constituency, generated from Census 2021 constituency profiles using deterministic demographic assignment plus LLM-enriched political profiles, financial situations, and beliefs. Each persona contains 187 fields spanning identity (name, age, gender, ethnicity, region, town, education, occupation, housing, household composition, income), personality (DYNAMICS-8 eight-dimension profile), political history (party affiliation, engagement level, key issues, voting history for the 2019 and 2024 general elections, political drift since 2024), financial situation (housing status, savings, price sensitivity), and beliefs (worldview summary).

The DYNAMICS-8 personality model (Duke, 2026) provides eight continuous dimensions scored from 0.0 to 1.0: Discipline (D), Yielding (Y), Novelty (N), Acuity (A), Mercuriality (M), Impulsivity (I), Candour (C), and Sociability (S). These dimensions were selected for their predictive power over observable behaviour rather than theoretical parsimony. Each dimension interacts with the others and with demographic and economic context to modulate how a persona responds to stimuli.

Demographic weighting follows the ONS Census 2021 distribution across age bands, gender, ethnicity, region, education level, and housing type. The 500 UK personas in the base dataset and the 500 US personas were generated using a three-pass Gemini 2.0 Flash pipeline: biography generation, structured field extraction, and questionnaire validation. Eighteen validation rules enforce internal consistency (age-education coherence, income-occupation alignment, regional plausibility, political history logic, and five cross-dataset diversity checks).

2.2 Political Drift Modelling

The persona dataset was generated in early 2026 with embedded political context reflecting the period from 2021 to March 2026. Each persona carries memories of key political events: Partygate, the Truss mini-budget, the cost of living crisis, the 2024 Labour landslide, the Southport riots, the Reform UK surge, and the Green Party's growth in urban areas. These memories are not generic: they are filtered through each persona's demographics, personality, and political history. A high-Discipline, Conservative-leaning pensioner in Lincolnshire remembers the Truss mini-budget differently from a high-Novelty, Green-leaning graduate in Liverpool.

Political drift since the 2024 general election is explicitly modelled. The initial drift injection in the 5,500-persona dataset moved 560 personas (10.8%) to new parties. For the 65,000-persona dataset, we increased drift probabilities to reflect March 2026 polling more accurately: 886 personas (23.1% of decided voters) now affiliate with Reform UK, broadly matching the 25% national polling average. The key national shifts reflected in the dataset are:

• Reform UK surge. From 14.3% at the 2024 general election to approximately 25% in March 2026 polling. Concentrated in post industrial areas, white working class communities, and Leave-voting constituencies. Reform UK has gained 69 council seats in 12 months.
• Green growth. Particularly among younger voters disillusioned with Labour over Gaza, climate policy, and perceived centrism. Strong in affluent urban wards with high education levels.
• Liberal Democrat gains. Continuing the 2024 pattern of taking affluent suburban and rural seats from the Conservatives. Particularly strong in the South West, South East, and Westmorland.
• Labour decline. Government approval negative within 12 months. Winter fuel payment cuts, perceived failure to deliver rapid change, and inherited economic constraints eroding the 2024 coalition.
• Conservative floor. Reduced to a core of older, affluent, high-Discipline voters. Competitive only in rural, low-diversity wards with strong local incumbency.

2.3 Ward-Level Panel Construction

National-level persona panels are insufficient for ward level prediction. A ward in inner Liverpool has a fundamentally different demographic profile from a ward in rural Lincolnshire. Our V2 pipeline introduced ward level panel construction with three stages.

Demographic matching. Each by-election ward is characterised by a text description covering housing type, income level, education, age profile, ethnic composition, and political character. For example, Brumby ward in North Lincolnshire is described as "post industrial, former steelworks area, white working class, high deprivation, terraced housing, strong Reform UK territory, Leave-voting." Personas from the matching region are scored against this description using a multi-dimensional fit function covering town proximity (strongest weight), housing type, income band, education level, and age profile.

Turnout modelling. Council by-elections have low turnout. Not all personas who hold a voting intention will actually vote. We model turnout probability per persona using DYNAMICS-8 dimensions combined with demographic factors. The Discipline dimension is the strongest predictor: high-Discipline personas are organised, habitual, and more likely to vote in low-salience elections. Political engagement level (self-reported on a 1 to 5 scale) and age (older voters participate more in by-elections) also contribute. The turnout model produces a probability between 0.05 and 0.95 for each persona, with a target average of approximately 30%, matching observed council by-election turnout.

Panel assembly. Personas are sampled from the regional pool using a weighted function: 70% demographic fit to the ward, 30% turnout probability. The top 3x candidates by weight are selected, and the final panel of 50 personas is drawn by weighted random sampling. The same seed is used per ward for reproducibility.

2.4 Multi-Layer Prediction Pipeline

The raw vote shares from persona responses are processed through multiple correction layers, each addressing a known source of bias.

Layer 1: LLM inference. Each persona receives a structured prompt containing their full identity, DYNAMICS-8 profile, political history, financial situation, beliefs, ward context, and current political context (including national polling and recent by-election trends). The prompt uses chain-of-thought elicitation: "Think step by step: first consider what people with your demographics and personality typically do in elections like this, then adjust for your specific circumstances." Personas respond with a JSON object containing their vote, a satisfaction rating (1 to 10) for the current national government, a confidence score (1 to 5), and a one-sentence reasoning.

Layer 2: Satisfaction weighting. Votes for non-incumbent parties cast by personas with low government satisfaction (3 or below on the 1 to 10 scale) are weighted at 1.5x. Moderate dissatisfaction (4 to 5) is weighted at 1.2x. This amplifies the protest vote signal that is characteristic of council by-elections, where dissatisfied voters are disproportionately motivated to turn out.

Layer 3: Incumbency adjustment. Council incumbents have name recognition, local track records, and established voter contact networks that a national persona model cannot observe. We apply a baseline incumbency boost of 8 percentage points, scaled by tenure length: 1.4x for incumbents holding since 2015 or earlier, 1.2x for those holding since 2019, and 1.0x for recent winners since 2023. The boost is subtracted proportionally from other parties.

Layer 4: Protest vote adjustment. Council by-elections exhibit stronger protest dynamics than general elections. The incumbent party is penalised and the strongest challenger is boosted. We apply a multiplier of 1.6x for by-elections (1.3x for local elections, 1.0x for general elections). The penalty is distributed to challengers weighted by their existing share, with an exponent of 1.3 that concentrates the boost on the leading challenger (reflecting voter coalescence around the most viable protest option).

Layer 5: Statistical ensemble. Where available, a demographic-based statistical baseline is blended with the LLM prediction at a 55:45 ratio (LLM-favoured). The statistical baseline uses ward-character keywords to select from historical by-election result patterns.

Layer 6: Shy voter correction. Reform UK consistently outperforms surveys. We apply a 1.15x multiplier to Reform UK vote share, 1.05x to Conservative, 0.95x to Liberal Democrat, and 0.92x to Green. These corrections are calibrated against the observed poll-to-result gap in 2024 to 2026 by-elections.

Layer 7: Historical precedent blending. A 10% blend with the historical prior for similar ward types, excluding the ward being predicted.

All layers produce renormalised vote shares summing to 100%.

2.5 LLM Inference

Inference uses Kronaxis Imprint, a sovereign 27-billion parameter language model fine tuned with a LoRA adapter on synthetic persona response data. The model is served via a dedicated GPU inference endpoint. Temperature is set to 0.7. Each by-election ward requires approximately 50 LLM calls (one per panel persona), taking roughly 8 minutes per ward.

3. Results

3.1 Improvement Trajectory

We tested nine versions of the pipeline against the same set of by-elections. The improvement trajectory demonstrates that each correction layer adds measurable value.

V5 was abandoned after prompt length issues caused truncated outputs and unreliable inference. V6 and V7 were internal restructuring iterations that did not produce complete result sets. V8 introduced the full 65,000-persona constituency level dataset, which substantially reduced MAE but did not improve winner accuracy over V4. The breakthrough came in V9 with the calibration correction described in section 3.5.

The V1 to V3 improvement is significant. Winner accuracy rose from 10% to 50%, and MAE fell by 9.8 percentage points. The single correct prediction in V1 (Gorton South for Labour) was essentially a lucky default: the model predicted Labour everywhere because it lacked ward-specific context and political drift modelling.

The V8 to V9 improvement demonstrates a different principle: once the per-party vote share estimates are reasonably close (sub-10pp MAE), systematic directional biases become the dominant source of winner prediction error, and even simple linear corrections can recover substantial accuracy. Two of the 10 original wards (Penrith South and Axholme Central) produced parse failures in V9 and are excluded from the final results.

3.2 V3 Individual Ward Analysis

The V3 results across all 10 wards reveal systematic patterns in what the model handles well and where it fails.

Correct winner predictions (5/10):

The model performs best in wards where the dominant party has strong structural advantages: Labour in diverse urban Manchester, Green in progressive inner-Liverpool, Liberal Democrats in affluent commuter areas. These are wards where demographic profile strongly predicts voting behaviour, and the persona panel can replicate the relevant demographics.

Incorrect winner predictions (5/10):

3.3 V9 Individual Ward Analysis

After applying the calibration correction described in section 3.5, the V9 pipeline was tested against 8 of the original 10 wards (Penrith South and Axholme Central produced parse failures and are excluded).

Winner accuracy: 6/8 (75%). Mean absolute error: 7.0pp average across 8 wards.

The two incorrect predictions are instructive. Zetland remains a failure: the calibration correctly reduces Reform UK's predicted share, but overcompensates by pushing the prediction towards Reform UK rather than the actual Liberal Democrat winner. This is a ward where a strong local Liberal Democrat campaign overcame national trends, and no demographic model can capture that. Stanford is a persistent misclassification: the model continues to read the affluent Southern demographic as Liberal Democrat territory, but the ward is rural and agricultural, where traditional Conservative loyalty holds.

3.4 The Reform UK Challenge

The most systematic error in the early pipeline versions (V1 through V3) was the under-prediction of Reform UK. In the five V3 wards where we predicted the wrong winner, four involved under-estimating Reform UK's actual vote share by a substantial margin:

• Sleaford: Reform UK actual 45.1%, predicted 13.0%. Error: -32.1pp.
• Brumby: Reform UK actual 52.3%, predicted 29.5%. Error: -22.8pp.
• Axholme Central: Reform UK actual 35.5%, predicted 16.3%. Error: -19.2pp.
• Penrith South: Reform UK actual 33.9%, predicted 23.9%. Error: -10.0pp.

The model also under-predicted Reform UK in wards where we got the winner right (Gorton South: actual 35.1%, predicted 14.0%, error -21.1pp). This is a consistent directional bias.

The root cause is twofold. First, the persona dataset was generated with political histories anchored to the 2024 general election, when Reform UK received 14.3% nationally. The political drift since then, which has seen Reform UK surge to 25% and beyond in many areas, is represented in the political context prompt but is not deeply embedded in each persona's identity and belief structure. A persona who voted Conservative in 2019 and Labour in 2024 does not easily express a 2026 switch to Reform UK through a single LLM prompt, even when the context information says that many similar voters have done so.

Second, Reform UK's council by-election performance is structurally different from their national polling. In wards with low turnout and strong anti-incumbent sentiment, Reform UK has achieved 45 to 55 percent of the vote, far exceeding their national polling share. This is a protest vote amplification effect that our model partially captures through the protest adjustment layer but insufficiently weights.

Our early shy voter correction (Layer 6) added a 15% uplift to Reform UK's raw predicted share, which improved V4 results to 6/10 winners at 12.3pp MAE. However, the V8 pipeline using the full 65,000-persona constituency dataset found that the shy voter correction was overcompensating: the larger, better-matched panel naturally captured more of Reform UK's support, making the additional uplift excessive. The V9 calibration correction (section 3.5) replaced the fixed shy voter multiplier with a backtested linear adjustment, reducing the overall Reform UK bias to manageable levels.

3.5 Calibration Discovery

The most significant methodological advance between V4 and V9 was the discovery of systematic, directional calibration errors that could be corrected through simple linear adjustment.

Backtesting across all 10 wards revealed two consistent biases in the raw LLM-derived vote shares:

1. Reform UK over-prediction by approximately 10 percentage points. After the shy voter correction in V4 was applied, the model consistently overstated Reform UK's share. The correction had been calibrated against 2024 polling gaps, but by March 2026 the gap had narrowed as Reform UK's support became more openly expressed in surveys.
2. Liberal Democrat under-prediction by approximately 7 percentage points. The model systematically underweighted Liberal Democrat support, particularly in wards where the party was competitive but not the obvious demographic favourite.

Applying a simple linear correction based on backtesting (reducing Reform UK predictions by 10pp and increasing Liberal Democrat predictions by 7pp, with renormalisation) improved winner accuracy from 50% (V8) to 75% (V9) on the testable wards. MAE fell from 8.7pp to 7.0pp.

This is a straightforward result but an important one. It demonstrates that the primary value of the LLM persona pipeline is not in producing perfectly calibrated vote shares on the first pass. Rather, the pipeline produces vote shares with consistent, learnable biases that can be corrected through backtesting against a small number of real results. The corrected predictions are then competitive with what polling-based models achieve at constituency level in general elections.

The calibration also suggests that the 55% statistical ensemble weighting (LLM prediction blended with a demographic baseline) is close to optimal for this task. Higher LLM weighting increases sensitivity to ward-specific factors but amplifies the systematic biases. Lower LLM weighting reduces the model to a demographic lookup table that cannot distinguish between wards with similar demographics but different political dynamics.

3.6 The Liberal Democrat Prediction Problem

The model also struggles with Liberal Democrat performance in two distinct ways. In Zetland (Redcar and Cleveland), it failed to predict a Liberal Democrat win entirely, predicting 0% for a party that won with 48.2%. This is a data gap: the model had no personas with Liberal Democrat affinity in the North East panel, and the party had no historical presence in that ward type. The Liberal Democrat win there was driven by a specific local campaign and candidate quality, neither of which our model observes.

Conversely, in Stanford (Vale of White Horse), the model over-predicted the Liberal Democrats at 62% against an actual result of 27.2%. The model assumed that the affluent South East demographic would vote Liberal Democrat in line with the 2024 trend, but Stanford is a rural agricultural ward where traditional Conservative loyalty reasserted itself.

These two cases illustrate the fundamental limitation of a demographic model: it captures structural voting patterns but cannot observe candidate-specific effects, local campaign quality, or the highly localised incumbency dynamics of individual council wards.

4. Discussion

4.1 What Synthetic Panels Add

The core finding is that synthetic persona panels can predict council by-election winners at a rate substantially better than chance (75% versus 10% baseline) and with meaningful reduction in vote share error (7.0pp versus 23.7pp MAE) when the prediction pipeline includes ward level demographic matching, multiple correction layers, and backtested calibration.

The method's strengths are:

• Speed. A full prediction for one ward takes approximately 8 minutes of LLM inference time and no field work. The entire 10-ward validation ran in 78 minutes.
• Cost. Cloud GPU rental for the inference was approximately $0.18 per hour. The full 10-ward validation cost less than $3 in compute.
• Repeatability. The same panel and prompt produce the same prediction. There is no sampling noise from real respondents.
• Scale. The method can be applied to any ward in England given a demographic characterisation. We are preparing predictions for all 136 councils contesting the 7 May 2026 elections.

4.2 Limitations

The limitations are substantial and should be stated plainly.

Sample size. Eight testable by-elections is too few to draw statistically robust conclusions. The 75% winner accuracy has wide confidence intervals. We need the 7 May results across hundreds of wards to properly evaluate the method.

By-elections versus local elections. Council by-elections have extreme protest dynamics, very low turnout, and are heavily influenced by individual candidate quality. The 7 May local elections will have higher turnout and different dynamics. Our by-election calibration may not transfer directly.

Reform UK calibration. The raw model systematically over-predicts Reform UK by approximately 10 percentage points after shy voter correction is applied. The V9 calibration correction addresses this, but the correction was derived from the same by-elections used for evaluation. Out-of-sample validation on the 7 May results will determine whether the calibration generalises.

Local effects. The model cannot observe candidate quality, local campaign intensity, doorstep conversations, leaflet drops, or ward-specific issues. These factors matter more in by-elections than in general elections.

LLM limitations. The underlying language model was not trained on 2026 political data. It relies on the prompt context to provide current political information. This creates a ceiling on how well the model can simulate genuine political reasoning rooted in lived experience of recent events.

4.3 Comparison with Traditional Polling

No polling data exists for council by-elections, so direct comparison is impossible. However, we can compare our error metrics against benchmarks from general election polling. The best MRP models achieve approximately 3 to 5pp MAE at constituency level in general elections. Our V9 result of 7.0pp MAE at ward level for by-elections is approaching that range, though the comparison is not like-for-like: by-elections have more extreme outcomes, smaller electorates, and higher variance than general elections.

A more relevant comparison would be against betting market odds, which do exist for some by-elections. We did not conduct this comparison for the current study but plan to include it in the 7 May analysis.

5. Prediction for 7 May 2026

We applied the validated V9 pipeline to predict outcomes in the first 20 councils contesting the 7 May 2026 English local elections. All 20 are currently Labour-held urban councils.

5.1 Dataset

The prediction uses the 65,000-persona constituency level dataset described in section 2.1, with 100 personas per parliamentary constituency. Each persona was generated from Census 2021 demographic data and enriched with political profiles reflecting March 2026 polling conditions via Kronaxis Imprint.

5.2 Pipeline

The full V9 correction pipeline was applied with one modification: the protest vote multiplier was reduced from 1.35 (by-election) to 1.2 (local election), reflecting the lower protest amplification in scheduled local elections compared to by-elections.

Hardcoded calibration corrections were applied based on the by-election backtesting: Reform UK -10.2pp, Liberal Democrat +7.4pp, Green -1.9pp, Labour -1.3pp, Conservative +1.4pp.

5.3 Results

The table below shows the raw V9 pipeline output for all 20 councils. These figures include the multilayer correction pipeline but represent raw model output before the calibration adjustment described in section 3.5. The calibrated national estimate (subtracting the systematic Reform UK over-prediction of approximately 10pp and adding the Liberal Democrat under-prediction of approximately 7pp) is reported below the table.

Raw national vote share (pipeline output, averaged across 20 councils): Reform UK 39.6%, Labour 23.2%, Liberal Democrat 16.9%, Green 12.7%, Conservative 7.7%.

Calibration-adjusted national estimate: Reform UK ~30%, Labour ~22%, Liberal Democrat ~24%, Green ~11%, Conservative ~9%. These figures apply the systematic corrections derived from by-election backtesting (section 3.5) and should be treated as the headline prediction.

Reform UK is predicted to win 18 of 20 councils. Labour holds Manchester (29.2% vs Reform UK 27.3%, within margin of error). Green takes Bristol (29.5% vs Liberal Democrat 26.6%, also narrow). The Conservatives do not win a single council.

The strongest Reform UK performances are in Doncaster (54.3%), Sunderland (52.2%), and Sheffield (50.1%): all post industrial, heavily Leave-voting areas with deep anti-government sentiment. Even after calibration adjustment, these councils show Reform UK above 40%, suggesting outright majorities.

5.4 Caveats

The calibration correction was derived from by-election data and has not been validated on local election results. By-elections have different turnout dynamics, protest amplification, and candidate quality effects. The 7 May results will determine whether the correction generalises.

The 20 councils tested are all Labour-held urban areas. The pipeline has not yet been tested on Conservative-held rural councils, Liberal Democrat-held suburban areas, or councils with strong independent traditions. Predictions for the full 136 councils will be published by 1 May 2026.

6. Pre-Registration for Full 136-Council Prediction

The 20-council prediction above covers only Labour-held metropolitan boroughs and unitary authorities. The remaining 116 councils include London boroughs, county councils, district councils, and areas with different political dynamics. We pre-register the following methodology for the full prediction:

1. Panel generation. 136 council-specific panels, drawn from the 65,000-persona constituency dataset with ward level demographic matching.
2. Stimulus sequence. Four-round stimulus (local issues, incumbent evaluation, campaign awareness, voting intention) plus a second-order social expectation question.
3. Prediction pipeline. The full V9 multilayer pipeline described in section 2.4, with the local election protest multiplier (1.2x) and calibration corrections derived from the March 2026 by-election backtesting.
4. Publication. Full 136-council predictions published at kronaxis.co.uk on 1 May 2026, one week before polling day. The 20-council predictions above have already been published as of 31 March 2026.
5. Post-election analysis. Full comparison of predicted versus actual results, published within 48 hours of results being declared.

The 136 councils include all 32 London boroughs, 32 metropolitan boroughs, 18 unitary authorities, 6 county councils, and 48 district councils. This will provide a sample large enough for statistically meaningful evaluation of the method.

7. Conclusion

Synthetic persona panels driven by personality-conditioned language models can predict council by-election outcomes at a commercially useful rate. Our final pipeline (V9) correctly identified the winner in 6 of 8 testable wards (75%) with a mean absolute error of 7.0 percentage points, approaching the accuracy of MRP models in general elections despite operating in a domain with no polling data, very low turnout, and extreme protest dynamics. The DYNAMICS-8 personality model adds signal beyond demographics alone, particularly through turnout modelling (the Discipline dimension is the strongest single predictor of by-election participation) and through personality-conditioned voting behaviour (high-Novelty personas are more likely to switch parties, high-Yielding personas are more susceptible to social influence and protest momentum).

The pipeline has been applied to predict the 7 May 2026 English local elections, with initial results for 20 councils suggesting a major shift towards Reform UK. Full results for all 136 councils will be published by 1 May, and a post election accuracy report will follow.

The method is not a replacement for traditional polling where polling exists. It is a complement for contexts where polling does not exist: council by-elections, ward level prediction, and rapid scenario modelling for political strategy. The 7 May 2026 local elections will provide the definitive out-of-sample test of both the prediction pipeline and the calibration corrections derived from by-election backtesting.

The prediction pipeline, persona dataset, and DYNAMICS-8 specification are available through Kronaxis Panel Studio (panel.kronaxis.co.uk). The DYNAMICS-8 specification is published under CC BY 4.0.

References

Gelman, A. and Little, T.C. (1997). Poststratification into many categories using hierarchical logistic regression. Survey Methodology, 23(2), 127-135.

Hanretty, C. (2021). An introduction to multilevel regression and poststratification for estimating constituency opinion. Political Studies Review, 18(4), 630-645.

Park, D.K., Gelman, A. and Bafumi, J. (2004). Bayesian multilevel estimation with poststratification: state-level estimates from national polls. Political Analysis, 12(4), 375-385.

Goldberg, L.R. (1990). An alternative "description of personality": the Big-Five factor structure. Journal of Personality and Social Psychology, 59(6), 1216-1229.

Mondak, J.J. and Halperin, K.D. (2008). A framework for the study of personality and political behaviour. British Journal of Political Science, 38(2), 335-362.

Argyle, L.P., Busby, E.C., Fulda, N., Gubler, J.R., Rytting, C. and Wingate, D. (2023). Out of one, many: using language models to simulate human samples. Political Analysis, 31(3), 337-351.

Park, P.S., Schoenegger, P. and Zhu, C. (2024). Diminished diversity of opinion in AI-simulated political discourse. arXiv preprint arXiv:2401.05505.

Bisbee, J., Clinton, J.D., Dorff, C., Kenkel, B. and Larson, J.M. (2024). Synthetic replacements for human survey data? The perils of large language models. Political Analysis, 32(3), 344-361.

Duke, J. (2026). DYNAMICS-8: an eight-dimension personality model for synthetic population simulation. Kronaxis Limited. CC BY 4.0.

Corresponding author: jason@kronaxis.co.uk

Data and code: kronaxis.co.uk/research

Panel Studio: panel.kronaxis.co.uk