The sample is huge, but the sample of polling stations is tiny... what's the point of interviewing 17000 people if you're only going to do it from 130 polling stations?
A polling station could be really good or bad for a particular party even within a relatively balanced constituency, so how do they weight their choice of polling stations accurately? It can't take proper account of marginals at all. I reckon the most accurate poll was whatever that final one of the marginals was...