The tool was made using adjusted fenwick save percentage data from Emmanuel Perry's amazing Corsica.hockey. Adjusted fenwick save percentage shows if a goalie is saving more or less shots than an average NHL goalie would be expected to save if they faced the same quality of unblocked shots.
The distribution of possible talent level of the goalie was created using Bayesian statistics. The problem we are trying to solve is "What is the probability that a goalie is actually a certain talent level given their performance in the NHL so far?". As there is a lot of luck involved in goaltending, a goalie may save more or less shots over the course of a season than their true talent would reflect. For example, a goalie whose talent level is the exact NHL average has about a 12% chance of actually saving 5 or more goals over the course of 500 shots than average goalie typically would. By combing this with knowledge of the distribution of talent of NHL goalies in general, we can generate an estimate of a goalie's true talent level.
For a mathematical example: the probability that a goalie is actually a 930 save percentage talent goalie given they have a 920 SV% over 500 shots = % of goalies with 500 shots against who historically turn out to be 930 SV% goalies multiplied by the chance a truly 930 SV% goalie would have a 920 SV% over the course of 500 random shots divided by the % of goalies who have faced 500 shots that have a 920 SV%.
As you can see from the graph above, there is significant overlap across almost all goalies in talent level, which acknowledges how similar most NHL goalies are and that saving one more goal out of a 100 is the difference between a good and bad goalie.
Goalies with less than 10 life time games played were excluded from the creation of Bayesian priors. Three different groups of priors were created based on the amount of NHL experience each goalie has. This helps to account for the fact that high quality goalies are more likely to have survived to face more shots.
A potential future improvement is to use beta-binomial regression to give each goalie their own unique prior instead of just having three for all NHL goalies.
The previous four seasons of data was used for each goalie, with older years being weighted less based on analysis by Eric Tulsky