Reform UK returned 1,448 councillors across 14 English councils on Thursday 7 May 2026, finishing 894 seats below the PollCheck and YouGov final MRP projection of 2,342 published 24 hours before polls opened . The shortfall is a 38 percent undershoot, the largest projection error in modern UK polling.
MRP, the polling technique that combines a national vote-share estimate with constituency-level demographic data, came of age in 2017, when Chris Hanretty's model called a hung Parliament against the consensus. It has since become the trusted projection method for Westminster elections. The 7 May ballot was its first major five-party fragmented-FPTP test. First past the post (FPTP), which awards each council ward to the candidate with the most votes regardless of national share, behaves predictably under two-party competition. Under five parties polling at 26, 18, 17, 17 and 16 percent, ward-level outcomes turn on a few hundred votes either way, and uniform-swing assumptions, which redistribute votes proportionally across every ward, stop describing reality.
The operational cost landed within 48 hours. Reform had stood up transition teams for 22 councils and inherited 14. Civil-service induction packs sized for 22 went unused; coalition arithmetic prepared for 22 collapsed at 14. The party's projected 26 percent national vote share was confirmed at the count; what the model could not project was how that share converted to ward wins under fragmentation. The Greens, working from the same family of FPTP models, hit the inverse problem: 18 percent vote share projected to 696 council seats, against an actual 543 (event-06).
The consequence runs further than this council cycle. Every commercial and academic Westminster MRP currently in the field for the next general election uses the same uniform-swing under-the-hood assumption that produced the 894-seat miss. The 2026 polling industry now owes its readers a structural review of MRP performance under five-party fragmentation before the autumn party conference cycle, or its projections will be priced at a 38 percent error bar by every campaign that has read this result.
