Jason Chao
Work

UK local elections 2026 — a forecast comparator

A static, interactive site comparing eight published forecasts of the 7 May 2026 UK local-election seat outcomes against the actual result. Four ways to read the same data, transparent dated provenance, recomputes three of the eight from current polling against a 2022 baseline. Live.

2026 ·Open-source · live ·Data, Civic tech

Live at uk-elections.jasontc.net; source at jason-chao/uk-elections (open source).

A side-by-side comparator of eight published forecasts of the 7 May 2026 UK local-election seat outcomes, set against the actual results. The same data is shown four ways — each view answers a different question.

At a glance

The At a glance view: two stacked horizontal bar rows labelled Actual and Consensus, each showing seats per party in colour-coded segments. Below, a per-party list shows actual seats, consensus seats, and an off-by delta in green or red.
Actual versus consensus across the chosen methods, shown as stacked bars by party, with per-party off-by deltas underneath. The fastest read of how the methods landed.

Bars per party

The Bars per party view: a vertical bar chart with one cluster per party (Labour, Conservative, Liberal Democrat, Green, Reform UK, Others / Independents). Each cluster shows a hatched bar for the actual result alongside three solid bars for each method's prediction, with 80% confidence intervals as error bars.
Direct comparison: the actual result (hatched) alongside each method's prediction with 80% confidence intervals.

Method spread

The Method spread view: a horizontal range chart with one row per party. Each row shows an IQR grey box, individual method dots, a triangle marking the defending baseline, and a star marking the actual result.
Per-party variance across the chosen methods. IQR box, individual method dots, defending-baseline triangle, actual result as a star. The closer the dots cluster around the star, the better the methods called it.

Numbers

The Numbers view: a detailed table with party columns (Labour, Conservative, Liberal Democrat, Green, Reform UK, Others / Independents) plus a Mean Absolute Error column. Rows show the actual result, the defending-baseline seats, each method's prediction with low-high range and per-party error against actual, and summary rows for minimum, maximum, range, range as percent of consensus, consensus estimate, and 80% interval.
The full table: every method's prediction with low–high range, per-party error against actual, and mean absolute error in the last column. The accountability view.

The eight methods

  • Ward Monte Carlo (PollCheck)
  • YouGov MRP
  • R&T NEV — Rallings & Thrasher National Equivalent Vote
  • BBC PNS — Projected National Share
  • Strong Transition (Electoral Calculus)
  • Fisher Regression (Elections Etc)
  • By-Election Extrapolation (Mark Pack)
  • UNS Baseline — Uniform National Swing

Three of the eight — BBC PNS, R&T NEV, and UNS Baseline — are re-computed from current polling against a 2022 baseline rather than re-published from the original forecaster. The other five reflect each forecaster’s own published numbers at build date. The same comparison runs across National plus five English regions.

Track records

The eight methods aren’t equally reliable. A toggle at the top of every view filters between All, Best 3, and Worst 3 — useful both for forming a credible consensus and for sanity-checking what the weaker methods predicted.

Ranking is by historical track record: each method’s past forecasts are compared with the corresponding actual results, and methods are ordered by how well they called previous elections. The Track record tab on the live site shows the per-method history that underwrites the rank — visible to any reader who wants to audit the ordering.

For the 7 May 2026 election, the Best 3 were Ward Monte Carlo (PollCheck), YouGov MRP, and R&T NEV. The consensus across these three is the more informative default; an average across all eight pulls in weaker methods and dilutes signal.

Keeping it current

The interesting engineering choice is the data refresh. An AI-assisted refresh routine pulls the latest forecasts, re-derives the BBC PNS / R&T NEV / UNS Baseline methods from current polling, regenerates the data files, and reports what moved between runs. The difference between “this gets stale in three weeks” and “this stays current with one prompt”.

What it demonstrates

  • Election-forecasting methodology literacy across eight named methods, including the difference between re-publishing a forecast and re-computing it from inputs.
  • Data visualisation with transparent provenance — every cell traceable to a dated source file in the archive.
  • Method ranking grounded in measured past performance, not editorial opinion.