Before the World Cup I started to take a look back at the league projections that Aaron Nielsen and I made at the start of the 2013-14 European season. I took a break during the World Cup to concentrate on the Soccermetrics Connect API, and then took some time away. With the 2014-15 European season about to start, it’s time to finish taking a look back before we look further ahead.
So here is a summary of how we did for each league that we projected. I’m going to split this post into parts.
Overall the RMSE of the expected model was about eight points, but I don’t think that’s good enough for a ten-team competition. We were quite good at predicting the finalists as we got five of the top six. There were some mitigating circumstances with Melbourne Heart — Orlando Engelaar and Harry Kewell were out of the squad for extended periods due to injuries — but perhaps we should not have projected such a high number of goals scored for a squad with two aging players. The difference in point totals has much to do with the difference in projected goals allowed. If it is very large, as it was for Brisbane and Newcastle, there will be a large difference in points as well as final position.
This projection was for Torneo Inicial only. The RMSE of the expected model was low (6.58 points), but it’s a deceptive figure in a short tournament. Argentina’s short tournaments are perhaps the most chaotic that they have ever been in terms of the performance of the major clubs, and that volatility was reflected in a poor expected league table. On average, we were off by seven places in the final league table. It was fitting that the league title was won by San Lorenzo de Almagro, who we projected to finish in mid-table. (San Lorenzo just won the Copa Libertadores two days ago, in a final against the only team seeded below them at the conclusion of the group stage. So South American club football was chaotic this year.) Better local knowledge of the squads would have helped here.
This was about as close as you could get to a perfect league projection. Our points RMSE was 6.3 points, which is as good as it gets in a 34-match season. (By comparison, our English Premier League projection had a RMSE of 10.9 points.) Our final position RMSE was 2.3 places, which might have been smaller if Utrecht’s defense hadn’t cratered. We nailed the top two teams in the league, Twente surprised us again (but they seem to surprise everyone once every few years), and we predicted two of the three relegated teams. And to top if off, we won Simon Gleave’s Eredivisie prediction challenge. That and a couple of bucks will buy coffee at Dunkin or Tim Hortons, but being first is nice. What was amusing about the search was that we were all over the place in terms of predicted goals scored and allowed. Speaking of which, OH MAN! Do any Eredivisie teams play defense in that league? It’s pretty clear that if you manage an Eredivisie side and keep your defense just a little snug, you have a very good chance of either winning the league (Ajax), entering European competition (Twente), or avoiding relegation (Breda).
Of course the downside of doing so well is that it will be very difficult to repeat in future seasons. But who knows, maybe we can catch lightning in a bottle again.
Jeez, we were way off in terms of expected points and places (RMSE of 10.5 and 4.4, respectively). We got Györ, Videoton, and Ferencváros among the top four, but we missed badly on Debrecen. Diósgyör was another bad miss, and the two teams we projected for relegation reached safety (albeit in the lower half of the table). The two teams that were relegated were projected to finish comfortably in mid-table. So, not a good projection.
Celtic were always going to win the league, and Hearts were always going to finish bottom, so there’s no skill in predicting either end of the table. We did project three from the top four (Celtic, Aberdeen, Dundee United), but Hibs’ relegation was a bad miss. There were more goals than expected, not as many as Holland but in contrast to the Eredivisie, the variance in goals scored and allowed reflected itself in wide differences between expected and actual points.
We thought the title would come down to Red Star and Partizan, but there’s no real skill in that. Our predictive model actually did very well with Partizan’s goals scored and allowed, which was reflected in the very small difference between expected and actual points. But it failed to pick up the 20 extra goals scored by Red Star, which converted what we thought would be a comfortable Partizan title into a narrow Red Star win. The model performance was worse everywhere else and failed to predict goals scored or allowed with any accuracy. Again, local knowledge would have helped.
Another one of those non-European leagues that follows a European calendar. Our projection captured five out of the top six teams with the exception of Platinum Stars who finished eighth (we projected them to finish fourth). Mamelodi Sundowns were the surprise team of the season relative to our expectations. Below sixth place, finishing position was very sensitive to goals scored and allowed, so an estimate far away from actual totals resulted in large discrepancies in point totals and final positions. There’s also a ‘luck’ factor that needs to be considered when comparing the expected table. For example, AmaZulu had exactly the same number of projected goals scored and allowed, yet they scored nine more points than expected.
This league projection was almost as perfect as the Eredivisie’s, and even surpassed it to some degree. Our projected champion was correct and we captured the teams in the top and bottom three, but we failed to pick up Luzern. Nevertheless, the points and final position RMSEs were excellent for a competition of this size.
We captured the top three teams, but the implosion of the Kayseri teams hurt our overall RMSE (Kayseri Erciyesspor finished 19 points and 10 places below expectations, and Kayserispor finished 27 points and 13 positions below expectations). We missed the relegated sides, Elazığspor was kind of close but not really. Karabükspor was another big miss, we projected them to finish bottom yet they finished seventh! Turns out that their defense ran well ahead of expectations.
Given the administrative rulings in the first half of the season, and the turmoil in eastern Ukraine in the second, I’m amazed that the projection for Ukraine’s Premier League turned out so well. But perhaps that’s because the top five or six teams are pretty much predictable. Our picks for relegation probably would have gone down were it not for Tavriya Simferopol and Arsenal Kyiv being dissolved or ejected from the league. When Arsenal were ejected, their results and remaining schedule were expunged, so all teams played 28 matches instead of 30. This is reflected in our RMSE calculations. A points RMSE of 8.7 and a positions RMSE of 2.6 is pretty good, but I haven’t decided if that’s the result of a good expectation or a predictable league competition.
A statistical summary:
|Total Goals||Pred Goals||Pct Change||RMSE Pts||RMSE GF||RMSE GA||RMSE Pos|