Originally Posted by SteelSD
Because it's a junk metric that attempts to assign situational importance to an event while disregarding that a baseball game isn't played in a linear fashion.
Basically it just looks at historic play by play data to assign an average win probability to a specific event in a specific context-i.e. a single with bases loaded in the bottom of the eighth while tied increases the team's chances of winning on average over all similar situations in the last gazillion games by X therefore the average value of that single in "wins" is likely "X".
It actually doesn't require that a baseball game be played in a "linear fashion". Since it's largely derived from the application of Markov chains to the play by play data, the sequence of events that lead to a particular state is in fact wholly unimportant.
In other words, the focus is strictly upon the context of the event and the win value associated with the probability of the outcome (i.e. what state you are in, what potential states you can transition into, and the likelihood of each of those possible transitions).
It's really just leveraging the mountains of available play by play data by applying basic probability theory to generate a win expectancy matrix which can be tweaked depending upon the specific players involved (i.e. the probabilities change depending upon whether Coffey is pitching with bases loaded in the 9th or Cordero is on the mound).