204: Players ranked. 5: WAR components. 139: Pending free agents. 6: Front-office verdicts A box score tells you what a player did last season. A roster decision needs something else: what she will do next season, and whether that is worth a contract. The Front Office board answers that for every PWHL player, and the expansion page turns it into protection lists. This post explains how the number is built, because a model that recommends signings should show its work. Last season is the wrong number Take last season's production at face value and two errors creep in. The first is finishing luck. A forward who buries an unusually high share of her chances posts a great scoring line, but shooting percentage is one of the least repeatable things in hockey: next season it drifts back toward normal. Pay her for the hot year and you have overpaid. The second is age. A 23-year-old is probably still improving and a 34-year-old is probably declining, yet both look identical in a raw total. A projection has to correct for both before it is safe to act on. Step one: WAR, and the finishing fix The starting point is WAR, wins above replacement: how many wins a player added compared with a freely available fill-in. Our PWHL WAR has five parts. Even-strength impact comes from RAPM, a regression that isolates a player's effect on shot quality from her linemates and competition. Then penalty discipline, faceoffs, goaltending (goals saved above expected), and finishing. Finishing was a late addition, and it mattered. An audit of the WAR leaderboard found the top ten had no forwards in it at all: it was goalies and defenders. The reason was structural. RAPM measures chances created and suppressed, but expected goals is the same whether a shot is saved or scored, so the model was blind to the one thing elite forwards do best, which is convert. The finishing component fixes that: it is goals scored minus the summed expected goals of a player's own shots, the goals she added by beating the odds. With it, forwards returned to the top of the board where a hockey person expects them. Step two: regression to the mean Each of those five components is then pulled back toward the league average by how much it actually repeats from one season to the next. This is the heart of the projection. Finishing is regressed hard: only about 30 percent of a player's finishing above expected is projected to carry, because most of it is noise. Faceoff skill is the opposite, a genuinely stable trait, so roughly 80 percent carries. Even-strength impact sits in between at about 70 percent. Kelly Pannek is the clearest example on the current board. Her finishing was worth 1.43 WAR last season, the largest finishing edge in the league. The projection keeps only 0.23 of it. That is not a knock on her. It is the model saying her shooting year was partly the bounces, and a team paying for it next season should expect a step back. Step three: the age curve The regressed value is then aged. An empirical curve fit from PWHL season-over-season changes, one per position, says how a player's value tends to move at each age. Below peak the curve rises, above it the curve falls, and it is held monotonic in both directions: a three-season sample throws off enough noise that the raw fit briefly had forwards improving into their mid-thirties, which is survivor bias, not biology. The projection scales each player by how the curve moves from this season to next. A young player gets a small bump; a veteran gets a small discount. Add the regressed, aged components together and that is projected WAR. On the current board the top projection belongs to Haley Winn at 0.89 , down from 1.09 last season once regression and age are applied. Does the projection actually work? 13%: Lower error than naive. 0.38: Projection error (MAE). 0.43: Naive baseline (MAE). 140: Players tested A model that adds steps should beat the simple thing it replaces. The simple thing here is the naive baseline: carry a player's 2024-25 WAR forward unchanged as her 2025-26 projection. To test against it, the projection is rebuilt from the seasons before 2025-26 only, so it is predicting a season it was not allowed to see. Across the 140 players who appear in both seasons, the projection's average error was 0.38 WAR , against 0.43 for the naive baseline: 13 percent lower . Root mean squared error, which punishes big misses harder, moves the same way (0.59 for the projection, 0.64 naive). The projection lands closer to reality because it does not pay full price for a career year: it regresses an unrepeatable spike back down, and players who spiked do come back down. Two honest caveats. The gain is in error, not ranking: the projection and the naive baseline correlate with next-season WAR at about the same level (0.33 and 0.35), and both are low, because next-season WAR is genuinely hard to call on three seasons of league history. And this is one held-out season; a fourth PWHL season will sharpen the test. But on the number that matters for a roster decision, how close the estimate lands to what actually happens, the projection is the better tool. The verdict A ranked number is useful, but a front office wants a recommendation. Every player gets one, from her projected WAR and her 2026-27 contract status. A projection of about 0.70 or more marks a top-tier contributor; about 0.30 or more, a useful regular; below that is depth. Crossed with whether she is signed or a pending free agent, that gives six verdicts: Priority re-sign : a pending free agent who projects as a top-tier player. The re-signing to close first.. Re-sign : a pending free agent who projects as a solid regular. Worth a new deal at a sensible number.. Let walk : a pending free agent projecting near replacement level. The slot and the money go further elsewhere.. Extend : a signed top-tier player. Lock her in longer before she reaches free agency.. Hold : a signed regular at fair value. No action needed.. Monitor : a signed player projecting modestly for her roster spot. Watch whether the contract earns out. On the current board 8 players land in the top tier (Extend or Priority re-sign) and 95 pending free agents come back Let walk. That last number is the useful one: a large free-agent class is mostly depth, and the model is blunt about which names are not worth chasing. The expansion model The expansion page runs the same projection through the 2026 protection rules: three protected players in Phase 1, three more in Phase 3, and the painful cut, the best player left exposed. One wrinkle matters. Phase 1 is ranked by the regressed value before the age discount, so a 35-year-old captain coming off a strong season is judged on what she can still do, not penalized twice for her age. Phase 3 uses the full aged projection and is limited to signed and restricted players, because an unrestricted free agent left unprotected has signed elsewhere by the time Phase 3 opens. What this model isn't It is single-season. The projection starts from one season of WAR. A veteran coming off a quiet year projects below her longer track record. A multi-season blend would soften that, and is the next planned refinement.. It does not price contracts. The verdicts are about value, not dollars. "Re-sign" does not say at what number, because the model cannot see salaries. A GM still has to negotiate.. It rates value, not scarcity. Two players at the same projected WAR are not equally hard to replace. Losing a starting goalie hurts more than losing a depth forward of equal value. The model measures what a player is worth; the front office weighs how replaceable she is.. It is an estimate. A projection built on partial seasons carries real error bars. Treat the top of any tier as a cluster, not a strict order, and the verdict as a starting point for a conversation. Methodology WAR : PWHL-calibrated, five components, all in wins. Even-strength impact (RAPM), finishing (goals above individual expected goals), faceoffs, penalty discipline, and goaltending (goals saved above expected).. Regression : each component is blended toward its league-average value by a fixed weight set from year-over-year repeatability. Finishing is trusted least, faceoff skill most.. Age curve : an empirical per-position curve fit from PWHL season-over-season changes, held monotonic away from the peak, applied as the ratio of next season's multiplier to this season's.. Verdicts : derived from projected WAR and the player's 2026-27 contract status. Thresholds are documented above.. Numbers regenerate on every build. A play-by-play re-sync, a finished game, or a model recalibration reshapes every figure in this post and every row on the board. The live tools are at /front-office (the roster-value board) and /front-office/expansion (the protection lists). Player WAR by season is at /stats .