Dan Falkenheim

One-year WNBA RAPM is noisy. Can encoding prior information, without using box score inputs, help? Yes. Prior-Informed RAPM closes about half of the predictive gap with three-year RAPM, while estimates stay rooted in current season play. Ranks and details below

Taking a step back: RAPM generally improves with at least three seasons in the training window. That makes sense, as the model can work with a larger sample. A drawback of multi-year RAPM, though, is that more stability comes with less sensitivity to current year performance.

I wanted to see if there was a way to both improve one-year RAPM's predictiveness *and* retain its responsiveness To do that, I used each player’s previous three-year RAPM as a weak prior for the one-year model. Doing so provides better predictiveness than ordinary two-year RAPM

A second issue: RAPM wants to lean hard on the prior, which pulls the estimates toward three-year RAPM The prior's penalty share can become a hyperparameter that balances predictiveness with responsiveness Setting it to 1/3 was a sweet spot (Spearman = 0.95 w/ one-year RAPM)

tl;dr: Single season prior-informed RAPM... - Only uses previous three-year RAPM as a prior (no box score metrics) - Caps the prior penalty share at 1/3 (weak prior) - Improves predictive performance and maintains sensitivity to current season

Research here was made possible by the previous work on stabilizing RAPM with priors by Fearnhead and Taylor (2010) and @JerryEngelmann. Engelmann's writeup on incorporating priors back in January formed the foundation for what I did here.

For the full 2026 WNBA regular season prior-informed RAPM leaderboard, check out my writeup for @SInow: si.com/wnba/why-veron…

It's fair to ask why I focused on one-year RAPM when multi-year RAPM is more predictive. I wanted to build a descriptive one-year BPM/SPM model. So, I also wanted a more stable one-year target. For forward-looking uses, time decay or multi-year RAPM is likely the better route.