Skill vs. luck decomposition

Four components

At 5v5 we break a player's on-ice production into four parts: chance creation (their own expected goals per 60), own finishing (their goals versus their own expected goals), teammate finishing (how well linemates converted while they were on the ice), and goalie luck (how their goaltenders did versus expectation behind them). The first two are things the player largely controls; the last two are things that happen around them.

How skill and luck are combined

The luck score is teammate finishing plus goalie luck - effects that reliably wash out over time. The skill score adds chance creation and own finishing, but as within-season standardized scores rather than raw numbers, because a per-60 rate and a goals count live on different scales and summing them raw lets one drown the other out. Each season's expected goals are also re-calibrated to that season's actual scoring, so luck stays an approximately zero-sum quantity across the league rather than absorbing an era-specific bias.

Does it hold up?

Tested year over year across the shift-data era, the pieces behave exactly as the split predicts. The things a player controls carry forward; the things riding on others don't. Chance creation is the anchor - a player who generated dangerous looks last season overwhelmingly does so again this season. Own finishing is far streakier, which is why it earns less weight. And the two luck pieces barely repeat at all, which is the whole point of calling them luck.

Standardizing the two skill pieces before summing them is not cosmetic: it lifts the combined skill score's year-to-year stability from 0.196 (raw sum) to 0.575 (standardized), while the luck score sits at 0.081 - close to noise, exactly where it should be. The stability numbers above are measured on 6,457 pairs of consecutive qualifying player-seasons (at least 500 5v5 minutes in both years).

The split also earns its keep looking forward. The skill score correlates +0.565 with a player's next-season 5v5 goals and +0.468 with next-season points; the luck score correlates +0.028 with next-season goals - essentially zero. When both are put in the same forecast, skill does nearly all the work and luck contributes almost nothing. That is the honest version of the claim: skill predicts, luck doesn't.