Not very

I was thinking "about as accurate as tossing a coin," but that seems unfair to the coin:

According to National Weather Service statistics for southwest Pennsylvania over the past 20 years, Punxsutawney Phil's record is 4 wins, 5 losses, and 11 ties.

Does that mean he's right 20 percent of the time? Let's plunge ahead:

A "tie" occurred when February's and March's temperatures differed greatly. For example, in 2011, temperatures in southwest Pennsylvania were below-average in February but above-average in March. So I counted that as a tie.

What does "greatly" have to do with being above or below average? Would two SDs below the mean in February but only -1 SD in March count as differing greatly? How about the other way around? What happens if Phil bets the mean and hits it?

Actually, Phil does better at predicting March temperatures than February.  If you compare his prediction with one month's temperatures, he had a 7-13 record in February but 11-9 in March.

If we had a consistent idea of what "predicting temperatures" meant, we could conclude that Phil's not very good in either one.

The point of the story seems roughly the same as it is with self-selecting polls and other stuff that finds its way to the homepage: It's a lot of fun, and if the groundhog's forecast suits your tastes, it also brings a bit of hope. That's as may be. But once you've told us you believe in magic, it's hard to know when we're supposed to take things seriously again.