Two weeks ago, I made some predictions about what would happen in the Italian elections. It’s only right that I should revisit them now, and evaluate their accuracy — and engage in a bit of self-criticism.

First — and opting to emphasise the positive — I predicted that the centre-left would `win’ the Camera, and would become the largest group in the Senate. So far, so good. However, the centre-left’s margin of victory in the Camera was extremely small, at around 0.4%. I had previously speculated that only a “melt-down” comparable to 2006 could prevent the left from winning a majority in the Camera, and it seems that it was in fact comparable.

Second, the seat predictions I made in the Senate were badly off. Here’s a table of my predicted seat shares against actual seat shares.

The mean absolute error, taking all parties listed, is 11.5; the mean absolute error, for only those parties which actually entered the Senate, is 15. Given that the range of outcomes I canvassed involved the centre-left being anywhere from ten seats short to ten seats over (depending on Lombardy), an error of that magnitude means that what I did wasn’t that useful for forecasting.

Third, the predictions I made were off because the polls were off. If you look at the regression tree I made, then my description for node 7 is roughly accurate:

“If the centre right gets more than 28%, and the centre-left gets less than 32.6%, then the centre-left has done as badly as it can and still win the national majority in the Camera. There’s no chance of a Senate majority, and they’ll be more than ten seats short”

So, had the exit polls given us reliable information, I could have made an instant prediction that would have been proved right. As it was, the exit polls were wrong, and badly so. This, to me, suggests that the polling industry has made a collective mistake. Gianfranco Pasquino was on Rai News 24 gamely arguing that the polls hadn’t got it wrong, that late (blackout) polls were showing the M5S increasing its vote share. Maybe he was talking about polls like this Ispo poll. If that’s the case, then Pasquino’s idea of getting it right involves an error of 4.8%. He must share his ideas of success with Mario Monti.

What’s the lesson from this? Add a large random-walk for the blackout period? Introduce a term capturing the probability of groups of parties being systematically underestimated? My guess is that, if a fix were easy, the polling companies would have done it already. For the moment, I’m going off to read this book, and see what I can learn from 2006 and 2013.