SALO has reached the state (version 1.0) that it will remain in for the 2017 offseason, but it will be developed over time.

In addition to a major release each summer, SALO will see minor updates once or twice each year, around opening day and the trade deadline if possible.

Only with each year’s major release will the likelihood change (if at all). Minor releases will add new features without changing the interpretation of the (underlying) model coefficients themselves.

Features targeted for introduction in the next minor release (version 1.1) include:

Features targeted for introduction in future minor releases, but not necessarily version 1.1, include:

Each major release will mean an expansion of the set of events whose probability is modeled in terms of skater abilities. Features targeted for introduction in a major release (version 2.0 or later) include:

What format results will be reported in upon future releases remains to be seen. In particular, converting model coefficients to wins above replacement without ignoring effects already modeled in SALO will rapidly become a more challenging task as the model gets more complicated.

Lastly, in parallel with the effort to add new features, a second aim is to compute model estimates faster and for more years by identifying and using alternatives to Stan. It may prove feasible to overcome the extreme memory usage and resistance to parallelization that characterize Stan by instead hand-coding the likelihood and prior and drawing from them with a stand-alone implementation of the No-U-Turn sampler. The target date for a recoded SALO (or for giving up on one if it won’t help) is the release of version 2.0 in summer 2018, with a stretch goal of getting it done by the minor release at the 2018 trade deadline.