puntgm

The measurement

Every claim below is a replay: draft a team with a strategy, play the season that actually happened, count who won more categories. Three completed seasons, twelve draft seats, four seeds — about 29,700 category decisions per arm per seed.

Correcting for availability is worth +11 to +20pp

Category win-rate difference, in percentage points, against the per-game z-score that free ranking lists publish. Positive means this board won more categories.

Pair2025-262024-252023-24
null — the same board in both seats +2.6 −0.7 +2.1 +1.0 −0.2 −0.6 −0.2 +0.6 +0.1 −0.2 +0.3 −0.5
this board vs per-game z-score +17.9 +19.6 +17.5 +18.6 +15.9 +16.1 +15.9 +17.1 +10.9 +13.1 +12.0 +11.9
this board vs total-value z-score −1.2 −0.9 −1.3 −1.6 −0.7 −0.7 −1.0 −1.0 −3.8 −3.8 −3.4 −3.8

The first row is a control: the identical board drafted against itself. It is there so you can see the noise floor rather than take our word for it. The largest number in it is 2.6pp; the smallest in row two is 10.9pp.

And the part most people would leave out

Read the third row. Against a total-value z-score — z computed on season totals, which Basketball Monster and Hashtag Basketball both expose behind a toggle — this board loses, by 0.7 to 3.8pp, in all twelve runs.

That is not a caveat we were forced into. It is the finding. The edge over free rankings is availability, not some superior metric, and if you already pay for a tool with a total-value mode you already have most of it. What we will not do is quote you the big number and hope you never check the small one.

Where the number comes from

Each run drafts two strategies head to head in a twelve-team snake, nine ADP-driven bots filling the rest, then grades every team against every other team every week of the real season. Seats are mirrored: each arm drafts from all twelve positions, and each rotation is run again with the draft order reversed.

That mirroring is not decoration. Before we added it, a snake over an odd number of rounds was handing the arm listed first up to +9.5pp for nothing — an effect the size of the ones we were trying to measure. The full account is here, including the other error it hid.