Building a Football Prediction model: Week 2

Summary

To the nearest whole number, the model predicted the correct goals scored for 10 out of the 20 teams, giving a strong accuracy of 50%.

Since last week, I’ve made a few changes to my model. I’ve downweighted past results exponentially and added the factor of recent form (last 5 matches) to influence the expected goals (xG).

I added this formula to the model to downweight past results. The decay rate is set to 0.01, so a game 30 days ago is 70% factored in to the teams’ attacking and defensive strength, whereas a game from 180 days ago is 17% factored in. This favours recent form, hopefully increasing the accuracy. I’ve found this formula in a 1997 paper entitled ‘Modelling Association Football Scores and Inefficiencies in the Football Betting Market’.

One team to be studied is Manchester United. Their recent form, due to the new head coach, Michael Carrick, is admirable. A new gaffer often triggers a new manager bounce, an immediate, short-term improvement in results, lasting about 15 games. According to BBCSport, of the last 30 permanent appointments, 18 bettered their team’s points tally in their first five league games in charge compared to the previous five under the old boss.

What went right?

A game that went according to script was Fulham vs Brighton. The model correctly predicted a 2-1 win to Fulham, but this match was decided by a beautiful 90+2 minute free kick by Harry Wilson.

What went wrong?

Watching this on TV, I was truly surprised. After Forest lost 1-0 midweek to Braga without the opposition having a single shot on target (Ryan Yates Own Goal), I had completely written them off for this match. Well done Forest.

Next time

For transparency, I’ll write my next blog alongside xG predictions before next week’s matches. I’ll research the new manager bounce and try add this into my model. Its effect might be minimal, but it’s still worth taking into account.

Instead of calculating the xG of a team overall, I could try to calculate the xG of individual players. This would need me to find a predicted lineup of both teams and factor this into the model.