Predicting Eurovision with Country-Level Regression

May 16, 2013

There seem to be a few people out there attempting to predict the Eurovision this year using this technique or that, from the Guardian’s tongue-in-cheek averaging out of past results, to Cold Hard Facts using Bayesian inference. I thought I’d throw my hat in the ring, with yet another equally fuzzy model. I don’t expect to do as well as CHF, as I banged mine together in 5 minutes.

Firstly, for those of you desperate to know what I predict, here are my top ten:

  1. Norway: 154
  2. Russia: 147
  3. Estonia: 124
  4. Romania: 114
  5. Georgia: 109
  6. Finland: 108
  7. Azerbaijan 105
  8. Italy: 100
  9. Greece: 97
  10. Latvia: 95

My method was as follows: Using the Guardian’s data set, I performed a linear regression on each country’s previous voting intentions since 1975, changing points back into positions to avoid a skew on the data. This gave me a vague idea of how the relationship changed over time. I then predicted each country’s voting intentions, compiled the voting intentions into points, removed current non-competitors, and ranked the totals. The result is what you see above.

Naturally this does not take into account the quality of the song. Early on, I was a bit worried that the model was predicting an Estonian win, with their dreary power(ish) ballad this year, but once I removed Morocco, Yugoslavia, Turkey et al, the votes changed a bit.

My model is distinct from the others, in that it predicts a relatively poor showing for Azerbaijan. It also has obvious problems, in that it only looks at occasions in which countries placed one another. So really, it’s just for fun, but aren’t they all? I do expect it to correlate with reality come the weekend though, so if you’re looking to place a bet on the UK, I advise against it. But you already knew that.