Ah, so they apply the general variance to the previous election across individual constituency results. That makes some sense, though it seems fairly lazy. A lot can change locally in five years, but at least the UK isn't redefining its constituencies for every election.
I mean, if we had better data we'd do more with it. You need a sample size of around 1,000 people to have a poll accurate to within 3 points. There are about 50 important marginals in the country. That means you need to sample about 50,000 people to get a good idea of the marginal situation. That is absolutely ludicrously expensive. Conducting an accurate national poll needs just a 1,000 people; and your predictions don't end up being significantly different. It's not laziness so much as making the best out of what you have available.
We also do redefine our constituencies fairly often for changing populations; every other election normally.
I would disagree with that qualification and call it highly inaccurate. It may somewhat confidently predict who will be the largest, but it seems particularly unfair to smaller parties. These polls must be frustrating for them.
If you use UNS from the last elections for England and Wales and Scotland based on the current polling average, you get this as a result:
269 CON
285 LAB
18 LD
54 SNP
1 GRN
2 UKIP
I'd be surprised if that wasn't fairly accurate. It's probably going to be a little bit inaccurate because the Liberal Democrat vote has collapsed differently in some areas to others, but if any of the major parties were more than ~20 seats away from this, and any of the minor parties more than ~5 away, I would be shocked.
