I am developing a machine learning model aimed at predicting football match outcomes based on past team performances. The model incorporates data from the last 10 home games for the home team and the last 10 away games for the visiting team, in addition to all past confrontations between the two teams, regardless of the number. However, I am facing a few challenges:

Missing Data: Sometimes, a team hasn’t completed 10 previous matches before the given match.

Variable Number of Past Direct Confrontations: The number of times the two teams have previously faced each other varies, with some pairs never having met before.

How can I structure my neural network to accommodate these inputs with variable sizes? Are there specific techniques or methods for integrating this kind of uneven or missing data effectively without impairing the model's performance?

Any advice or resources you could recommend would be greatly appreciated.

  • $\begingroup$ You can simplify things by introducing a home advantage as the square root of the average number of home goals divided by the average number of away goals in one season. You can then normalize the outcome to one on neutral ground. You should, however, take into account that the home advantage is constantly decreasing for all professional soccer leagues (and probably others). So if you train your model with long-term data, you should estimate the home advantage with a linear regression first. $\endgroup$ Commented May 28 at 16:20
  • $\begingroup$ Have you considered RNN's? Would that work? $\endgroup$
    – Miss Girl
    Commented Jun 9 at 19:08


You must log in to answer this question.