Put to the Test

In order to get the best possible predictions for fantasy football points we need to test what we are doing, and learn from the results.  Myself and Chris Glover at Premier League Fantasy Blog have set up a very simple test to begin with.

We are each going to predict performance stats and fantasy points for the following 5 players:

• GK John Ruddy (Norwich)
• DEF Jonas Olsson (West Brom)
• MID Leon Osman (Everton)
• MID Antonia Valencia (Mancheter United)
• FWD Luis Suarez (Liverpool).

These 5 were deliberately chosen to represent a balance of teams in the division and not too focused on the big-hitters.

The following will be predicted for each player over a 5 week game span, GW11 to GW15.

• Total Shots
• Total Shots in the Box
• Total Shots on Target
• Goals Scored
• Key Passes
• Assists
• Clean Sheets
• Save Points (GK)
• Fantasy Points (inc & exc. Bonus Points)

Things like yellow and red cards, injuries, non-starts, will be excluded from the points if possible or the prediction for the player abandoned if he will miss a few games.  We're not trying to be too clever here, we just want to complete a simple test.

Note that Chris and I have developed different models with focus in different areas.   Each of us will probably have a better angle on differing aspects.  Be sure to check out Chris' forecast for this case study too.

What I hope to achieve from this:

1. Predictions of player points over 1 game are always going to be inaccurate.  Will 5 games be enough?  I tend to think so but let's see.
2. Shot data is much more predictive than goals and assists.  I hope to the most see accurate predictions in the Total Shots data, then less accuracy as we move down through the categories towards goals scored as the sample size reduces.  The amount the accuracy of each data type varies should give us more insight into how the number of games affects.
3. Identify important differences between the results of the two models and with the actual results... and then improve.