Methods and limitations

The metric

The core metric is expected versus actual time on ice, evaluated at contract level.

Expected usage is modeled from AAV, normalized as cap share, with position controls.

Actual usage is post-signing average time on ice per game for observed seasons.

Time on ice is used because it is available and reflects coaching preference at scale.

The expected-versus-actual framing also handles zero-TOI outcomes as a floor, not an error.

The data

• Contract data: a GitHub-hosted community dataset snapshot from Chief-Zach Sports-Data, accessed through the contract source seam adapter in extraction. Current signing-year window is 2012 to 2025.
• nhlscraper for skater performance and player bios.
• Wikipedia captaincy lists, used under Creative Commons attribution (CC BY-SA).
• Current skater season coverage: 2009-2010 through 2024-2025.
• Current contract signing coverage: 2012 through 2025.

Scope and exclusions

• Eligible walk-year sample size is 1,624 contracts.
• Eligible overpay sample size is 1,760 contracts.
• Coverage is uneven by signing year. 2012 to 2014 are sparse at 72, 89, and 106 signings, 2025 is partial at 40, and middle years run about 188 to 443.
• Scope is skaters only. Goalies are excluded.
• Entry contracts are flagged and excluded from walk-year effect analysis.
• Extension deals are flagged and excluded from walk-year effect analysis.
• Players without NHL performance rows are out of scope because no usage can be measured.

What the model cannot see

• Captaincy captures captains only. Alternate captains are not covered in the source.
• Intent is unmeasurable. The metric captures revealed outcomes, not motivation.
• Captaincy is an association variable. It is not a causal claim.
• The captain-as-discount archetype was tested and did not generalize in this sample. See Findings for the captain lens results.
• Some iconic older contracts predate the current contract source window. They are narrative anchors, not model rows.
• Time on ice captures coaching usage and trust. It does not capture every form of contribution.

Same-team and new-team contracts were not randomly assigned. Teams choose which of their own players to re-sign, and they tend to keep the players they expect to deliver. So part of the same-team advantage may reflect selection rather than a discount teams choose to give. The model controls for tier and trajectory, and the effect holds within those groups, which makes pure selection an incomplete explanation. But selection cannot be fully ruled out with observational data. The finding is a robust empirical regularity, not proof of a behavioral mechanism.