The key to FPL Reviews predictive algorithm is the use of bookmaker odds to forecast player and team performance. The value of the data on the site of course depends on how well it is performing- perhaps it is all just a waste of time. This is what this post aims to understand by use of mounds of data collected.
There are two layers to be investigated before a conclusion can be reached:
- How accurately does the FPL Review Team Planner tool forecast future odds? This will be investigated in this post.
- How well do bookmaker predictions correlate to FPL Performance? This analysis will be released shortly in a separate post.
Both of these layers must show strong predictive qualities to make the tools on this site worthwhile.
Using & Forecasting Bookmaker Odds
The first step is to understand how I performed this analysis and also how the algorithm works.
Bookmaker odds can be used to forecast player goals, clean sheets and team goals- though margins need to be carefully removed. FPL Review uses this data for 1 GW ahead predictions and historic xG and match data to predicts assists, cards and saves. A bespoke algorithm is used to determine expected bonus points.
For GWs for which the bookmakers odds have not been released the FPL Review algorithm is used. Simply put this algorithm learns from past bookmaker data how to predict future odds.
An initial 5 GW forecast using the Team Planner prior to GW18 is the data which is being investigated. This means GWs 18,19,20,21 & 22 have projections for every FPL player.
GW18 projections are from bookmaker odds, however GWs 19,20,21 & 22 are all deduced from the FPL Review algorithm. To check the accuracy of this data I had to wait until the odds for GW22 had been released. This meant I could compare the data available prior to GW18 to all the actual bookmaker odds.
The data collected as described above is plotted below. Each data-point represents the 5 GW forecast of an individual player.
There is an incredibly strong correlation (a factor of 0.976 for FPL points, 0.986 for goals and 0.960 for clean sheets) between the FPL Reviews deduced forecast and the actual odds that the bookmakers used. The mean accuracy of the forecast points per player for the full 5 GWs was 95.8%, the median was 96.9%.
This indicates that a 5 GW forecast using the Team Planner tool would typically be expected to achieve ~96% accuracy per player.
For a 5GW forecast the Team Planner tool achieved an average accuracy of 95.8%. With this kind of accuracy it is fair to say the FPL Review bookmaker learning algorithm is doing what it has set out to do and the Team Planner tool has an exceptional value.
This data should give regulars users of the FPL Review tool-set an extra sense of confidence when using these tools to aid decisions.
My next post will look into just how effective using bookmaker odds is in FPL.