The demographic ward model

What this model is

Every other model on this site starts from votes: the last result in a ward, swung by polling. This one deliberately does not. It is shown only the structure of a place, its Census 2021 profile, its occupational mix, its 2024 general election result bridged down to ward level, its neighbourhood, and the national polling level on the day, and it learns how that structure maps to the local Reform UK vote. It models Reform because Reform is where the demographic realignment signal is strongest and where a structure-only model has the most to say; the same machinery can be extended to other parties.

Because the inputs are structural, the model can score wards that have no usable electoral history at all: new boundaries, long-uncontested wards, and the proposed May 2027 shadow unitary wards. That is what it is used for on this site: an independent second opinion alongside the results-based forecasts, labelled separately and never blended into them.

Headline accuracy, tested the honest way

The model was trained only on 2024 and 2025 contests (2,056 ward-elections where Reform stood) and then asked to predict all 2,625 wards Reform fought in May 2026, results it had never seen, in a cooler national environment than the May 2025 peak. On that unseen test:

• 4.4pp average error on Reform vote share (a predict-the-mean baseline manages 11.3pp), R² 0.81.
• 84.7% win/lose accuracy, AUC 0.912, Brier 0.11 (base rate 67.1%).
• Every prediction carries a 90% interval of ±9.6pp; 91.4% of the unseen results landed inside their interval.
• Win probabilities are calibrated: wards scored 50-60% were won 58.8% of the time, wards scored 20-30% were won 24.2% of the time.

A further forward test against 38 English local by-elections from 25 June to 16 July 2026, later than any training or test data, held the ranking power completely (AUC 0.914, accuracy 81.6%). It also measured a real effect: Reform underperforms these scores in low-turnout by-elections by roughly 11 points of win rate. By-election readings of this model should carry that discount.

What the model learned drives the Reform vote

The single strongest predictor, by a factor of roughly two, is graduate share (correlation −0.74 with Reform share). Education, not ethnicity and not deprivation, is the fault line of the realignment at ward level. Behind it, in order of learned importance: no-qualifications share (+0.46), West-European-born share (−0.65), UK-born share (+0.69), the area's UKIP vote from 2013-16 (+0.73, still independently predictive a decade later), occupational class (routine and semi-routine occupations +0.62, managerial and professional −0.44), car dependence (two-plus-car households +0.41), and age (50-and-over wards lean towards, under-35 wards lean against). A ward's neighbours matter too: the surrounding area's profile carries real weight beyond the ward's own.

One negative result worth stating: wards shaped by post-2004 Eastern European migration are not distinctive Reform territory once education and the UKIP legacy are controlled for.

How it works

Data. 19,618 ward-level results 2016-2026 assembled from Democracy Club, official May 2026 declarations and the Local Elections Archive Project, matched to ONS ward codes (95% match rate) including 1,382 county divisions via the ONS ward-to-division lookup. Features are composed population-weighted from LSOA-level Census 2021 data (ethnicity, religion, tenure, education, NS-SEC occupation, age bands, car availability), plus the average profile of the eight nearest non-overlapping wards, the ward-bridged 2024 general election shares, the council area's 2013-16 UKIP average, and the UK Elections 14-day national polling average on polling day.

Model. Predictions average a support vector machine (RBF kernel) and a histogram gradient-boosting model, hyperparameter-tuned with cross-validation grouped by council so no council appears on both sides of a tuning fold. Win probabilities pass through isotonic calibration fitted on out-of-fold training predictions. Intervals are split-conformal, from grouped cross-validation residuals, and checked against the unseen test. The support vector machine also identifies which training wards hold up the win/lose boundary, a direct read-out of where the contest between winning and losing demographics actually sits.

Scope: interpolation, not extrapolation

The national polling feature is what lets one model span different electoral climates, but it earns that honestly only within the range the model has seen: roughly Reform 14% (the 2024 anchor) to 30% (the May 2025 peak). A deliberately broken test proves the point: trained on 2024 alone, the model missed the May 2025 sweep by 18 points, because nothing in a 14%-world tells you about a 29%-world. All published uses of this model sit inside the validated range; if national polling leaves it, these scores should be treated as stale until the model is retrained.

Limitations

• It predicts places, not people. Nothing here supports claims about how any individual or any demographic group votes; that is the ecological fallacy.
• Scores assume Reform stands a candidate and are conditional on the stated national environment.
• Localist parties (residents' associations, town-first parties) are invisible to a census-based model and have taken wards it rated strongly.
• Multi-seat wards use a best-candidate-per-party share approximation.
• England only at present.

Where it is used on this site

The Lancashire unitaries forecast shows this model as an independent demographic second opinion next to the results-based seat forecast, across three named national environments. The two methods are labelled and never blended.