Methodology & Model Performance

Glossary

Key terms used in WIN326 projections and methodology.

UNS

Uniform National Swing — applies the national change in party support to every local ward.

Elasticity

A party-specific multiplier controlling how much national swing is passed to each ward. An elasticity of 1.0 means full swing; 0.75 means 75% of the national swing is applied locally.

Shrinkage

Statistical technique that blends raw projections toward a regional average to reduce noise.

MAE

Mean Absolute Error — the average difference (in percentage points) between projected and actual vote shares.

Winner accuracy

The percentage of wards where the projected winning party actually won.

Calibration

Whether confidence labels (safe, likely, lean, tossup) match observed win rates.

Backtest

Running the model against past elections where the result is known, to measure accuracy.

Baseline

The most recent actual election result in a ward, used as the starting point for projections.

Ward

The smallest electoral area in England and Wales. Each ward elects one or more local councillors.

Constituency

A parliamentary area that elects one MP. Contains multiple wards.

National prior

Blends ward projections toward the parent constituency's general-election-level projection (mu_base weight). Ensures national trends reach wards even without local baseline data.

Nationally significant

Parties polling above 5% nationally are automatically added to every ward's candidate set, even if they did not stand in the most recent local election.

Auto-learning

A daily feedback loop that compares predictions against actual by-election results, logs accuracy, and retrains the model when enough new evidence accumulates.

Candidate standing filter

Removes parties from projections if they are not fielding a candidate in that ward, and redistributes their vote share.

Frequently Asked Questions

How does WIN326 project local election results?

WIN326 applies Uniform National Swing (UNS) from current national polling to the most recent local election baseline in each ward. The swing is modified by party-specific elasticities — how responsive each party is to national trends locally. A national prior blends constituency-level general election projections into ward results, ensuring broader political shifts (such as the rise of Reform UK) are reflected even where parties have no local baseline. Data-driven shrinkage balances national trends against local history.

What data sources does the model use?

The model uses four primary data sources: (1) national opinion polls from all major UK pollsters, aggregated into a recency-weighted daily average; (2) historical local election results from official council declarations and Democracy Club; (3) candidate standing data where available; and (4) by-election results, which are automatically compared against predictions to evaluate and improve the model over time.

How accurate are the projections?

Model accuracy is measured via backtesting against historical elections and through the auto-learning feedback loop, which compares every by-election prediction to the actual result. Key metrics include winner accuracy (how often the projected winner actually wins), mean absolute error (average difference between projected and actual vote shares), and calibration (whether confidence labels match reality). Current metrics are shown in the Model Performance section above.

What is Uniform National Swing (UNS)?

UNS is a method that applies the national change in party support (from polling) uniformly to each local area. For example, if Labour is down 15 points nationally, UNS subtracts 15 points from Labour in every ward. WIN326 modifies this with party-specific elasticities — so a party with elasticity 0.75 only passes 75% of the national swing to each ward, dampening extreme shifts.

What is the national prior?

The national prior blends each ward's UNS projection with the parent constituency's national-poll projection. This ensures that broad national trends — especially the rise of new parties like Reform UK — are reflected at ward level even when the party has no local election baseline. The blending weight (mu) controls how much constituency-level signal flows into ward projections.

How does the model handle new parties like Reform UK?

Parties polling above 5% nationally are automatically added to every ward's standing set, even if they didn't stand candidates in the most recent local elections. This "nationally significant" override ensures parties like Reform UK appear in all ward projections. Their projected vote share comes from the national swing applied through the UNS model and the national prior constituency blend.

What is auto-learning?

The auto-learning feedback loop runs daily. It compares our ward projections against actual by-election results as they come in, logging prediction accuracy for every contest. When enough new by-election evidence accumulates, the model can retrain its elasticities from the fresh data. This means the model continuously improves as more real-world results become available.

What is shrinkage in the context of election forecasting?

Shrinkage blends each ward's raw UNS projection toward a regional or national average. This prevents extreme projections in wards with unusual baseline results. The shrinkage parameter (lambda) is automatically tuned via grid search over historical elections to minimise prediction error.

Why don't the projections cover Scotland?

Scottish local elections use the Single Transferable Vote (STV) system, which is fundamentally different from the first-past-the-post system used in England and Wales. UNS does not apply to STV elections. Scotland is excluded from projections but Scottish national polls are tracked separately.

Are these general election seat projections?

No. WIN326 projections are ward-level local election nowcasts — they project what would happen in local council elections if held today. They are not general election seat predictions. National polling data is displayed separately for reference.

How often are projections updated?

Projections are recalculated daily via an automated pipeline: new election data is imported at 07:00 UTC, the auto-learning feedback loop compares predictions to by-election results at 07:30, national polls are scraped at 08:00, and fresh projections are generated at 09:30. The Data Freshness section above shows exactly when the last projection was run.