Insights Blog

On target? Our goal predictions for 1. FC Nuremberg

Exasol stolzer Technologiepartner des 1. FC Nürnberg

Exasol has been a proud technology partner of 1. FC Nuremberg since 2020. But for the 2021/2022 season, we decided to take our fandom a step further – predicting 1. FC Nuremberg’s expected goals (or ‘xG’) for the club’s matchday previews. This prediction is based on a clever neural network model that can predict the xG for various European leagues. In this article, we’ll take a look at the performance of the neural network and see how we can improve it in future.

This comparison is based on all of 1. FCN’s games so far this season – nine home matches and nine away. It should be mentioned at this point that the model makes a strict distinction between home and away matches. Home obviously bringing an advantage. In addition to the complete first half of the season, we’ve included the first match of the season’s second half, between Erzgebirge Aue and FCN, which took place at the end of 2021. So, how did it do?

1. FCN – total predicted goals

In total, the model predicted 49.7 goals for the 1. FCN matches, which corresponds to an average of 2.76 goals per match. We reckon we can round that up to 50. In reality, 48 goals have been scored in those matches – 2.67 per match. Not bad! With a deviation of just 3.5% from the actual number of goals, this prediction has proved very accurate. Take a closer look at the details, it’s easy to see which areas the model made better or worse predictions in. Let’s dive in:

1. FCN vs. all opponents

Let’s look at 1. FCN’s predicted goals and the predicted goals for 1. FCN’s respective opponents:

Exasol 1. FCN predictions

It quickly becomes obvious that our model slightly underestimated 1. FCN and overestimated their opponents. Overall, however, this balances out.

The reason for this deviation? Our model reacts very late to changes in form and 1. FCN are in a better position this season than in the last (27 points after the first half of the 21/22 season compared to 20 points in the first half of the 20/21 season). We deliberately modified our model in this way, as it had been shown that rapid responses to changes in form tend to lead to poorer predictions. 

Home vs. away matches

When comparing 1. FCN’s home and away matches, a clear trend in the performance of the neural network can be seen:

Home matches

Exasol 1. FCN predictions

Away matches

Exasol 1. FCN predictions

While 1. FCN’s away matches were predicted very accurately by the model, there are larger deviations in the home games. Here, 1. FCN were strongly underestimated, while their opponents were heavily overestimated.

One explanation for this is 1. FCN’s much better home record in the 21/22 season compared to the previous season – which could be explained by the partial return of fans to stadia. In the 20/21 season, 1. FCN had a fairly even record with five wins, seven draws and five defeats and finished 14th in the home table. With five wins, two draws and two defeats already in the 21/22 season (6th in the home table), 1. FCN have been decisively better at home.

Away from home, 1. FCN’s performances have been very similar between last season and this season:

  • six wins, four draws, seven defeats (8th in the away table) in 20/21
  • three wins, four draws, two defeats so far (7th in the away table) in 21/22.

Again, this shows that form changes are incorporated too late, but with consistent performances the model delivers very good predictions.

The verdict

The neural network, implemented and regularly retrained by Exasol, has delivered very good predictions for the first half of the season. Although in individual matches the predictions have deviated from the result, over the entire period it’s proved accurate.

The model does take time to integrate changes in form, like the improvement in 1. FCN’s home record. In this area, several optimizations of the neural network are possible. We’ll therefore fine-tune our algorithm in order to be able to make even better predictions. It’s time for a team talk.