Because I'm sick of these morons misusing SV% and I have a bit of time to kill before my flight...
Over a sufficient data set, with relatively good accuracy, it gives us the individual level of play from a goaltender. Why is a sufficient data set necessary? Because too many things not controlled by the goaltender can create statistical noise within the sample. Examples of this would be PP shots on net (typically of better quality), better 5 on 5 shot quality (which have been shown to smooth out when we start looking at sample sizes in the 1000+ shot range). A goalie may face a higher (or lower) proportion of these higher quality shots in smaller data sets and we simply don't have reliable data on shot quality (yet). Larger sample sizes have been shown to smooth out that statistical noise and give us an accurate depiction of that goalies level of play over the larger time period.
If a goalie stops 25 of 30 shots, and that is the only indicator of his performance that we have, what does this tell us about his performance? Honestly...not a **** of a lot. In smaller samples, shot quality is a big deal. Were those 5 flubbers? Were they 5 breakaways? We don't know...so how can that data be useful to us in determining individual performance? If that data can badly skew a small portion of data selected to stand beside that one game, what value does that data set now have? Not a whole **** of a lot. Utilizing SV% in this manner completely invalidates SV%'s ability to act as a good measure of past performance....and...this is key kiddies....as a predictor of future performance. This is where the 2 and 3 full season splits come in to play. When you have a sufficient amount of data, SV% allows us to predict, again with pretty solid accuracy, the future performance of a goalie over another representational data set (typically 1000 shots or more). Keeping in mind that the difference between a .915 and .920 over 1000 shots is 5 goals allowed, SV% does a tremendous job of predicting future performance as the vast majority of goaltenders will be within 5-10 goals per 1000 shots of their most recent 2-3 yrs worth of performance. When you start breaking SV% down into small, meaningless chunks, it loses all ability to act as a quality predictor of future performance.
