Thursday 15 November 2012

Updates to Model

Anyway who has been following this site from the start a few months ago will hopefully have noticed a few tweaks to the points forecasts along the way.  You should know that my aim has been this season to develop the model along the way and continually improve it.  The changes I've made have not hugely affected the projections if I'm honest, and the top players remain the top players.  However, I will provide a quick update of the most relevant changes here as well as next planned changes:

I will highlight any significant changes to the player rankings in this week's Buying Guide and Select XI.  So far the changes to player projections have been so minor to not notice.

  • A team's average SoT is now entirely determined from this season's data (and has been since 6).  As can be seen from the red line on the first chart below a team's avg SOT converges to the season mean very quickly. 
  • Bonus Points: for a goal increased from 1 to 1.5 for Forwards and Mids and up to 2 from 1 for Defenders.  In effect from about GW7.
  • Goal Conversion rate (Goals:Sot)  adjusted for a player based on his percentage of shots in the box (Sin) compared to total shots (i.e. 80% Sin is better than 40% Sin).  This has been in effect for about 5-6 weeks.
  • Goal Conversion rate adjusted for a player based on his actual conversion rate this season regressed to the mean (currently 31%). By this stage in the season I expect that 50% of team's conversion rates to be similar to that which they'll achieve at the end of the season.  Similarly I expect a player's conversion rate now to be 50% reflective of their end of season average conversion rate.  The model updates each gameweek so a player's conversion rate gets closer to his actual rate as the season progresses.
  • As can be seen from the charts below I believe these are reasonable assumptions for Player and Team Goal Conversion.  The second chart below shows the aggregate conversion rate of 34 top scoring players last season, and this shows it arrives at a stable average of ~35 very quickly.  However, looking at individual players it's clear that taken one by one some players fluctuate quite a lot more early in the season but most have approached 50% variance by GW11.  Take effect from now (GW12)
  •  A player's share of his team's goal threat has been modified to include their share of Shots in the Box on an equal footing to Shots on Target.  This has reduced the point forecast for players shooting outside the box.  This was in effect for GW11.
  • A player's share of his team's Goal Threat now only takes account games in which the player played from 60 minutes or more.  This has reduced the effect of supersubs, most notably Lukaku.   This change has been in effect since GW10.  For further reading on super subs check out the Power of Goal's recent article on te subject.
Rolling Correlation for Team's Avg. SoT and Avg. Goal Conversion Rate. Prem 2011/12
Rolling Average Conversion rate for 34 top scoring players in Prem 2011/12

  • Adjust scoring of model to award Bonus Points relative to probability of winning the game.  Currently BPs are factored into the model based on goals scored so players on the losing team who score get the same BP as players on a winning team.
  • Incorporate "Big Chances" into  the model (thanks to work by SuperGrover).

No comments:

Post a Comment