Current navigation level
Previous navigation level
Next navigation level
Current navigation level

Soccer Analytics - Forecast Quality

On this page we check the quality of previous forecasts. To this end, we compare our forecasts to forecasts from other external sources.

The main objective of the model is to predict the final table. It is probably not possible to fully and most accurately predict it (otherwise football would be a very boring game). Each season has its own surprises and disappointments. Nevertheless we still want to check how precise our forecasts were in the last season. To this end, we compare our forecasts with those of goalimpact.com, F.A.Z., Spiegel Online and fivethirtyeight. Like us, they predicted the ranking of each team in the final table before the start of the season.

We calculate the rank correlation coefficients with the real final table and the respective forecasts. The closer a forecast to the truth, the greater this measure. For season 2016/17, our forecast was only beaten by the forecast of Spiegel Online.

Moreover goalimpact.com and fivethirtyeight predict also the points at the end of the season. The quality of forecasts for expected points is compared using the so-called Root Mean Squared Errors (RMSE). The lower this measure, the closer the prediction is to the true points. In this category our forecast performs better than goalimpact.com and fivethirtyeight.

In season 2015/16 our predictions were at least as good as all considered competitors.

We would be happy to compare our model to other sources. If you are aware of other predictions of the final table, please send an e-mail to soccer.analytics@unisg.ch.

Season 2016/17
Quality measure SEW Soccer Analytics
goalimpact.com F.A.Z. Spiegel Online
538
Rank correlation coefficient
 0.47 0.44 0.29 0.50* 0.42
RMSE 10.14* 10.87 - - 10.83
 * indicates best prediction performance.
 
Season 2015/16
Quality measure SEW Soccer Analytics
goalimpact.com F.A.Z. Spiegel Online
Rank correlation coefficient
 0.64* 0.64* 0.57 0.60
RMSE 8.57* 8.76 - -
 * indicates best prediction performance.