The headline outcomes in all three contests. Our central scenario gave Reform UK a 93% probability of winning the most seats in the English locals, Plaid Cymru a 69% probability of being the largest party in the Senedd, and the SNP a 99.9% probability of being the largest party at Holyrood. All three were correct.

The Labour-second-place call in England. The model assigned a 93% probability that Labour, despite their losses, would finish second in seats won on the night ahead of the Liberal Democrats. Most of the public commentary in the run-up to polling day framed this as a tight race between Labour and the Lib Dems for second place. Labour finished comfortably second with 1,068 seats, 224 clear of the Lib Dems on 844. This was the model's most distinctive call against consensus, and it landed.

Calibration of the confidence intervals. All twelve of our party-by-party seat predictions for the Senedd and the Scottish Parliament fell inside the 90% confidence interval we published. That is genuinely good calibration on what those intervals are meant to represent, though it does suggest the intervals may be slightly wider than they need to be. We would rather err on the side of honest uncertainty than false precision.

The sensitivity framing. We published three scenarios for the English locals — conservative, central, and aggressive — and the actual result landed between them and closest to central. That is exactly what scenario-based forecasting is supposed to do when the underlying uncertainty is high. It does mean readers had to engage with three numbers rather than one, which is a credibility trade-off we are still thinking about.

Comparison with other public projections, Reform seats won in England (actual 1,453):

Liberal Democrats were the biggest single miss. Our central scenario projected the Lib Dems would lose 240 seats net. They gained 155. In absolute terms our seats-won projection was 463; the actual was 844, off by 45% and in the wrong direction on net change.

This is bigger than any other error in the model and the one we are taking most seriously. Lib Dem performance has a known pattern of incumbency stickiness and ground-game concentration in target wards which the model did not capture well. There is also a Lab-to-Lib-Dem switching channel that we under-weighted: when our central scenario redistributed collapsing Labour vote, it sent most of it to Reform and the Greens. In reality a meaningful chunk went to the Lib Dems in shire districts and county councils where they were already well-placed to receive it.

Reform UK won lots of seats but flipped fewer councils than the model expected. Our central scenario projected Reform would take control of around 30 councils. They took control of 14. Reform won seats almost everywhere but did not concentrate them densely enough to clear majorities outside their strongest areas. Many councils tipped into no overall control rather than into Reform control. This is a distinction the GE model needs to be sharper on — “Reform wins most seats” and “Reform wins a majority” are very different propositions in a five-party FPTP system.

Welsh Greens. The English Green prediction was almost exact (581 forecast, 587 actual). The Welsh Green prediction was significantly off (7 seats forecast, 2 actual). This looks like a late-campaign Welsh-specific story rather than a structural model error — the English Green vote behaved as we expected, but the Senedd Green vote softened in a way we did not anticipate. We will check this against the Senedd polling timeline before drawing any model conclusions.

Labour total seats over-projected. Our central scenario had Labour on 1,349 seats; the actual was 1,068, an over-projection of 26%. This is the flip side of the Lib Dem under-projection and the same underlying error — when Labour lost vote share, more of it went to the Lib Dems than our model expected.

Conservatives in Scotland. We projected a Conservative mean of 16.7 Scottish Parliament seats; the actual was 12. Still inside the 90% confidence interval but at the lower end, suggesting our Conservative floor is set slightly too generously. Worth re-examining in the GE model.

A few things the May results do support:

Wide, honest confidence intervals matter. Twelve of twelve party seat totals inside the 90% CI is a vote of confidence in our uncertainty quantification. Models that report narrow ranges tend to look smarter when things go their way and badly wrong when they don't. We would rather show the genuine range of plausible outcomes.

Vote distribution matters more than vote share. Reform converted around 28% of the projected national share into roughly 41% of contested seats. That kind of FPTP geographic conversion is exactly what our seat model — which uses empirically calibrated within-council variance rather than uniform swing — is built to capture, and it largely worked.

The sensitivity-scenario approach held up. The actual result landed between our conservative and aggressive bookends and closest to central. That is what bookending is supposed to do when you know the underlying parameters could go several ways.

A few things the results challenge:

The Lib Dem channel is a real model gap. The Reform-most-seats-versus-Reform-majority distinction needs to be sharper. The Conservative floor may be set too high.

We are deliberately not over-reacting to a single election with unusual features — a mid-term low for the governing party and a Reform baseline near zero (because they were a new and small party in 2022, when most of these seats were last contested) that exaggerated headline gains numbers. Most of those won't generalise. But the Lib Dem error pattern is structural enough that it warrants a model rework rather than a wait-and-see.

Priority — Lib Dem model rework. Re-examine how the model handles Lib Dem incumbency, target-ward concentration, and Lab-to-Lib-Dem vote switching. Cross-check our seat projections in current Lib Dem targets against YouGov MRP and More in Common segmentation. This is the most important diagnostic before anything else changes.

Decouple “wins most seats” from “wins a majority” for Reform. The GE forecast will explicitly separate the probability of Reform winning most seats from the probability of Reform forming a majority government. In a fragmented FPTP system these are very different outcomes, and the May results confirm Reform's vote is efficient enough to top tables without being concentrated enough to clear majorities in most places.

Pressure-test the Conservative floor. Check whether the GE model carries the same slightly-too-generous prior on Conservative resilience that we now think showed up in the Scottish locals.

Stress-test tactical voting both ways. Locals are normally a low-tactical-voting environment, so May 7 tells us less about General Election tactical behaviour than the headlines suggested. The more interesting question for the GE is whether anti-Reform consolidation in marginal seats could produce a record-high level of tactical voting. We will run sensitivity scenarios in both directions.

What we are explicitly not changing on the basis of one election: protest vote discount, Reform-local discount, recency weighting on polls, or the locals-to-General-Election translator. These need more calibration points than a single cycle.

ReadingSignal exists because we think the polling community deserves a forecast that is honest about what it doesn't know and explicit about what it's doing under the bonnet. The model waterfall, the sensitivity scenarios, the published confidence intervals, and now this track record audit are all part of the same commitment. When we are wrong we will say so, name the error, and either fix it or explain why we are choosing not to.

The General Election forecast will publish in the same form: central projection, three sensitivity scenarios, party-by-party seat ranges with explicit confidence intervals, and a methodology page that documents the assumptions. The model is iterating; the discipline of being held to specific numbers is not.