Question. Per team-season 2021-22..2025-26: does the coach's leverage tilt (leverage-weighted TOI share minus raw TOI share) line up with out-of-sample player quality?
Method
- Skaters only, regular-season rated games; team spine
roster_toi.csv. - Split-half identification: each team-season's games are ordered chronologically; quality is measured on odd games (shrunk relWPA v3 per game, 35 pseudo-games of league average), deployment tilt on even games. Both metrics derive from on-ice events and quality earns leverage, so scoring deployment against quality measured in disjoint games evaluates the decision against information the coach could have had, not against the same minutes.
- Skaters need >= 5 GP in each half to enter the team correlation; teams need >= 8 such skaters.
- DeployEff = Spearman rank correlation across the team's skaters between shrunk quality and leverage tilt.
- Wins on table = conservative counterfactual: base TOI shares and the observed tilt distribution are held fixed; only the matching is changed so the largest tilt goes to the best player. A tilt unit for player i is valued at his quality per unit of leverage share (calibrated in the quality half, share floored at 0.5/roster). Assumption: per-game relWPA scales linearly with leverage share. relWPA pp / 100 = wins. Because quality is heavily shrunk and only the tilt is permuted, this is a deliberate lower-bound flavor of the true cost; it is also an upper bound on what re-matching alone can buy (some tilt is structural: PP units, matchup roles).
League picture
- 160 team-seasons. DeployEff mean +0.236, range -0.657 .. +0.797.
- Wins on table: mean 0.49 wins/season, range -0.11 .. 1.29.
5-season team ranking (mean DeployEff)
| Rank | Team | DeployEff (avg) | Wins on table (avg/season) | Seasons |
|---|
| 1 | Florida Panthers | +0.628 | 0.42 | 5 |
| 2 | Colorado Avalanche | +0.560 | 0.44 | 5 |
| 3 | New Jersey Devils | +0.535 | 0.27 | 5 |
| 4 | Carolina Hurricanes | +0.479 | 0.32 | 5 |
| 5 | Ottawa Senators | +0.475 | 0.41 | 5 |
| 6 | New York Rangers | +0.437 | 0.31 | 5 |
| 7 | Edmonton Oilers | +0.415 | 0.56 | 5 |
| 8 | Tampa Bay Lightning | +0.393 | 0.43 | 5 |
| 9 | Vancouver Canucks | +0.364 | 0.61 | 5 |
| 10 | Toronto Maple Leafs | +0.362 | 0.63 | 5 |
| 11 | Winnipeg Jets | +0.350 | 0.30 | 5 |
| 12 | Vegas Golden Knights | +0.349 | 0.58 | 5 |
| 13 | Dallas Stars | +0.339 | 0.48 | 5 |
| 14 | Minnesota Wild | +0.311 | 0.67 | 5 |
| 15 | Washington Capitals | +0.225 | 0.53 | 5 |
| 16 | Los Angeles Kings | +0.217 | 0.46 | 5 |
| 17 | Pittsburgh Penguins | +0.194 | 0.52 | 5 |
| 18 | Buffalo Sabres | +0.192 | 0.36 | 5 |
| 19 | Utah Mammoth | +0.178 | 0.40 | 5 |
| 20 | Boston Bruins | +0.160 | 0.46 | 5 |
| 21 | New York Islanders | +0.116 | 0.49 | 5 |
| 22 | Nashville Predators | +0.088 | 0.53 | 5 |
| 23 | Anaheim Ducks | +0.088 | 0.36 | 5 |
| 24 | St. Louis Blues | +0.064 | 0.40 | 5 |
| 25 | Chicago Blackhawks | +0.059 | 0.45 | 5 |
| 26 | Columbus Blue Jackets | +0.051 | 0.72 | 5 |
| 27 | Detroit Red Wings | +0.049 | 0.62 | 5 |
| 28 | Montreal Canadiens | +0.037 | 0.57 | 5 |
| 29 | Calgary Flames | +0.028 | 0.74 | 5 |
| 30 | San Jose Sharks | -0.040 | 0.60 | 5 |
| 31 | Philadelphia Flyers | -0.068 | 0.46 | 5 |
| 32 | Seattle Kraken | -0.090 | 0.51 | 5 |
Best and worst single team-seasons
| Team | Season | DeployEff | Wins on table |
|---|
| best | Florida Panthers | 2023-24 | +0.797 | 0.16 |
| best | Florida Panthers | 2025-26 | +0.786 | 0.66 |
| best | Dallas Stars | 2022-23 | +0.731 | 0.20 |
| best | Colorado Avalanche | 2021-22 | +0.728 | 0.28 |
| best | Carolina Hurricanes | 2025-26 | +0.727 | 0.22 |
| worst | Anaheim Ducks | 2022-23 | -0.369 | 0.53 |
| worst | Montreal Canadiens | 2021-22 | -0.370 | 1.09 |
| worst | Seattle Kraken | 2024-25 | -0.430 | 0.40 |
| worst | St. Louis Blues | 2022-23 | -0.435 | 0.53 |
| worst | Philadelphia Flyers | 2022-23 | -0.657 | 0.56 |
Most under-deployed player-seasons (good + negative tilt, >=40 GP)
| Player | Team | Season | GP | Quality (relWPA pp/gm, shrunk) | Tilt (lev share − TOI share) |
|---|
| T. Liljegren | Toronto Maple Leafs | 2021-22 | 61 | +2.941 | -0.27 pp |
| S. Noesen | Carolina Hurricanes | 2022-23 | 78 | +2.779 | -0.26 pp |
| J. Mahura | Florida Panthers | 2022-23 | 82 | +1.746 | -0.28 pp |
| C. Ruhwedel | Pittsburgh Penguins | 2021-22 | 78 | +1.590 | -0.29 pp |
| P. Hornqvist | Florida Panthers | 2021-22 | 65 | +2.499 | -0.18 pp |
| J. van Riemsdyk | Boston Bruins | 2023-24 | 71 | +2.147 | -0.21 pp |
| J. Manson | Colorado Avalanche | 2025-26 | 79 | +1.717 | -0.26 pp |
| D. Stepan | Carolina Hurricanes | 2022-23 | 73 | +1.552 | -0.28 pp |
| S. Noesen | Carolina Hurricanes | 2023-24 | 81 | +1.942 | -0.22 pp |
| B. Kulak | Edmonton Oilers | 2023-24 | 82 | +1.228 | -0.34 pp |
Most over-deployed player-seasons (bad + positive tilt, >=40 GP)
| Player | Team | Season | GP | Quality (relWPA pp/gm, shrunk) | Tilt (lev share − TOI share) |
|---|
| D. Savard | Montreal Canadiens | 2022-23 | 62 | -4.538 | +0.38 pp |
| M. Matheson | Montreal Canadiens | 2023-24 | 82 | -4.006 | +0.34 pp |
| C. Fowler | Anaheim Ducks | 2022-23 | 82 | -4.007 | +0.32 pp |
| C. Parayko | St. Louis Blues | 2023-24 | 82 | -3.486 | +0.29 pp |
| L. Crouse | Utah Mammoth | 2022-23 | 77 | -3.791 | +0.26 pp |
| N. Leddy | St. Louis Blues | 2023-24 | 82 | -4.172 | +0.22 pp |
| M. Seider | Detroit Red Wings | 2023-24 | 82 | -3.695 | +0.25 pp |
| C. Fowler | Anaheim Ducks | 2021-22 | 76 | -2.872 | +0.31 pp |
| E. Gudbranson | Columbus Blue Jackets | 2022-23 | 70 | -4.212 | +0.21 pp |
| K. Guhle | Montreal Canadiens | 2022-23 | 44 | -2.858 | +0.31 pp |
Validity check
- corr(DeployEff, standings points) across 160 team-seasons: r = +0.589.
- corr(wins on table, points): r = -0.192.
- We expected a weak positive; the observed r is stronger than pure deployment skill should produce. Two inflating channels: (1) good teams have clearer stars, so the quality-vs-tilt ranking is higher signal-to-noise (noise attenuates DeployEff toward 0 on flat rosters); (2) good teams spend more time protecting leads, and the lead-protecting minutes both carry leverage and go to players who rate well defensively in relWPA. Read DeployEff as coach aligns leverage with measurable quality, not as a pure coaching-IQ score.
Caveats
- Leverage tilt is partly structural (PP specialists, defensive-zone matchup roles, injuries mid-season), so a negative tilt on a good player is not always a coaching error.
- lwtoi weights high-leverage games as well as high-leverage minutes; the tilt therefore also reflects lineup choices in big games.
- Split-half kills same-minute endogeneity but not slow feedback (a player who was good in odd games was probably also good in even games the coach watched). That is fine: we want the coach to use that information -- the metric asks whether he did.