Adjusted goals projections
I’m using my long-term adjusted goal rating (explained here) to predict final league outcomes.
A note about the model
The projections are generated on a spreadsheet, stochastically simulating 5000 different league outcomes. I’ve modeled the outcome of each match using the rand() random number function to generate goals for and against – these are based on a Poisson distribution using a mean calculated from each team’s attack and defence rating.
Each team’s attack and defence rating is also randomly generated for each of the 5000 seasons – based on a Normal distribution with a mean of long-term adjusted goals and variance of 13% of the mean.
The projections seem to model sensible outcomes – but potential problems are:
- Although the long-term adjusted goal ratings are a good measure of a team’s performance – they don’t quickly adjust to changes (e.g. change of manager or transfer/injury of key player). I could make an adjustment – but this would introduce bias to the model.
- Spreadsheets aren’t a great tool for doing this type of modeling (my PC often crashes, 5000 simulations is about all it can handle). Also the rand() function isn’t truly random – although I don’t think this affects the results much.
- The variability of each teams attack and defense rating could be improved.