At NYA we spend most of our time evaluating players and professional clubs rather than trying to predict the outcomes of international matches. Prediction can be especially messy in knockout tournaments, where a single game isn't always enough for teams to display all of their qualities. But because it's fun, we still give it a shot.
Our World Cup model is very, very simple. It uses only five explanatory variables, with none based on detailed data about players or teams. It doesn't know even know who's going to Brazil and who's out injured. We developed it in 2010 for a laugh, and it happened to do pretty well. In fact, it beat (or tied, if we disregard statistically negligible differences) many of the other models out there. For example, Nate Silver's Soccer Power Index had a correlation of 0.58 between its pre-tournament ranking of teams and FIFA's post-tournament ranking; ours had a correlation of 0.59.
We calibrated our 2010 model with data from the 2002 and 2006 World Cups. This year's version uses the same five variables from all three of those previous tournaments. We didn't change anything else.
So, were we lucky last time, or are we really onto something? We revived the model this year to find out. Here are our rankings going into Brazil 2014:
| 1. Brazil | 9. Chile | 17. Belgium | 25. Ecuador |
| 2. France | 10. Switzerland | 18. Croatia | 26. Iran |
| 3. Portugal | 11. England | 19. United States | 27. Nigeria |
| 4. Germany | 12. Russia | 20. Mexico | 28. Honduras |
| 5. Uruguay | 13. Japan | 21. Netherlands | 29. Cameroon |
| 6. Argentina | 14. Algeria | 22. Bosnia and Herzegovina | 30. Australia |
| 7. Spain | 15. Ivory Coast | 23. Greece | 31. Costa Rica |
| 8. Colombia | 16. Italy | 24. Ghana | 32. Korea |
Do these rankings look a little odd, given the publicity surrounding the teams? Yes, they do. But if models can't tell us anything that goes against our prior beliefs, then there's no point in using them at all.
That said, we don't expect any model's predictions to be 100% accurate in any tournament. Models are designed to have a high success rate in prediction over a large number of tournaments. In other words, a good model might not win a single World Cup betting pool, but it would have more correct predictions than most people over a long series of pools.
Here are the model's predictions for the games of the group stage:
We'll have more predictions when the knockout rounds begin. Enjoy the tournament!