For those of you who don't know how MRP modelling works, it's based on real polling, but the aim is to project the likely winner in each individual constituency, and also to project the most likely overall number of seats for each party. (For statistical reasons, the overall number of projected seats for each party may confusingly not always tally up exactly with the number of individual seats that party is projected to win.) Because there aren't enough YouGov panellists in each constituency to directly measure support for parties at that local level with confidence, the process partly depends on assuming that similar types of voters in similar types of constituencies will behave in similar ways. In general those assumptions have been relatively well-founded, and in 2017 YouGov's MRP modelling was famously much more accurate than their conventional polling, even though they stated explicitly that they expected the reverse to be true.
There will be exceptions, though. You might remember that the BBC/ITV/Sky exit poll in 2017 produced some wildly inaccurate predictions at Scottish constituency level because it was assuming that voters in former Lib Dem constituencies would all behave in much the same way, whereas in fact the trends in (for example) Gordon and Edinburgh West were radically different from each other. The projection in the latest YouGov modelling that I'm most sceptical about is the Labour gain in Na h-Eileanan an Iar. It's not that the SNP aren't under threat in that constituency - they probably are to some extent. But the seat has changed hands between Labour and the SNP three times in recent decades (in 1970, 1987 and 2005) and on all three occasions the result would not have been predicted by the national trend. So if YouGov are projecting an SNP loss based partly on what is happening "in other similar SNP-Labour battleground seats", they're likely to have gone awry, because there aren't any seats that are similar to Na h-Eileanan an Iar.
YouGov projection of Scottish seats in the next UK general election based on voting intentions of 3586 respondents, collected 10th April-21st May 2023:
The BBC have been fools trying to set themselves up to a God-like status of infallibility and absolute truth. People might forgive "BBC News" misleading them or even lying to them occasionally, but the same from "BBC Verify" will provoke anger and mockery.https://t.co/7cUn4EydH2— James Kelly (@JamesKelly) May 24, 2023
* * *