Not as well as 308.com's, where Eric Grenier pretty much nailed it. Of course, people will still call him a left-wing hack.
However, this left-wing hack didn't do too bad either.
Of all 57 ridings, I made total calls in 48, and labelled it too close to call in 9. Of those calls, 7 of the 9 TCTC races were, in fact, pretty well called; and of the 48 called races, 3 were TCTC.
Of the races that I did call, I called 27 for the NDP, 20 for the PCs, and 1 for the Liberals. Of those calls, the model predicted 47 with the correct party winning, and one was inaccurate.
Of the 9 races I called as too close, I called 5 as leaning for the NDP, 3 for the PCs, and 1 that was absolutely tied. The model guessed correctly the NDP would win 5 of those ridings, and the PCs 3 - though not exactly the same ones as predicted. The one "tied" result, Kirkfield Park, which the model pegged at 44-44, is tied at 46% for the PCs and 46% for the NDP.
Overall, I predicted 33 seats for the NDP, 23 for the PCs, and 1 for the Liberals. Last night's results have been 37 NDP, 19 PCs, and 1 Liberal. The model guessed incorrectly in only three ridings, all of which predicted a slim PC win, but ended up being NDP holds.
In terms of the popular vote, I predicted the NDP at 42.9%, the PCs at 45.9%, the Liberals at 8.9%, and the Greens with 2.3%. The results were 46% NDP, 43.9% PC, 7.5% Lib, and 2.5% Green.
Had these results popped up in the model, the model would have gotten only two ridings incorrect, one being Southdale, which was too close to call anyways, going PC instead of NDP; the other being Swan River, also too close to call.
Overall, I'd say the projection model did very well. Here's just the list below of the "too close to call" ridings, plus the one called riding I got wrong: