§ NBA — The record

Every NBA projection, graded in public.

Each projection is saved before the game tips off and never edited. After the game it is graded against the box score, players who did not play included. The good nights and the bad ones stay on the page.

§ 01 — What this page grades

What gets graded

Fantasy points: each projection against the player's DraftKings NBA score from the box score.
Minutes: projected minutes against minutes played.
The 80% range: whether the actual score landed inside it.
The value board: the model's top 10 points-per-dollar plays on each slate, among players projected for 15 or more points, judged on whether they returned 5x salary.

Which games count

Regular-season, play-in and playoff games. Preseason is tested privately.
Each player in each game is graded once, on the last projection saved before his tip-off.

What it leaves out

Players listed Out when the projection is saved. They have no projection, so they are not graded. A player who was projected and then did not play is graded as a miss.
Players with fewer than three games of history. They have no projection.
Ownership and anyone's contest lineups. This page does not track either.

§ 02 — Fantasy points: model vs. the baselines

All three methods are scored on the identical 932 players, so none of them gets an easier test. Lower error is better. The baseline to beat is simply the player's DraftKings season average.

Model error (MAE) 8.43

DK season average error 8.25

80% range held 75%

Players graded 973

Method	MAE (pts)	Bias (pts)	Correlation
This model	8.43	+2.44	0.70
Model without injury logic	8.34	+2.27	0.71
DraftKings season average	8.25	+1.55	0.70

MAE is the average miss in fantasy points. Bias is the average of projection minus actual, so a positive number means the method ran high. The 80% range is the model's floor-to-ceiling band; it should contain the actual score about 80% of the time (973 players here).

§ 03 — Minutes, availability and lineups

Minutes. Projected minutes missed by 6.1 on average, against 5.9 for a plain recent average (973 players).

Availability. Of 695 players projected for 20+ points, 11% did not play: 8% of those with no injury tag, and 82% of the 28 listed as questionable or doubtful. A projection for someone who sits is graded as a miss.

Confirmed lineups. A DraftKings lineup flag existed before tip-off for 56% of projections. Of the 258 players flagged as starters, 100% started, and 100% of players who started had been flagged.

§ 04 — The value board

Each night the model picks its best points-per-dollar plays. A pick counts as a hit when the player returns 5x salary or more; a pick who does not play counts as a miss. The comparison picks use the same rule on DraftKings season averages.

Picks by	Picks	Hit rate	Avg. actual pts
This model	130	22%	13.1
DraftKings season average	130	46%	23.7

Rule: top 10 per slate by projected points per $1k (min 15 projected); a hit is 5x salary or more; a pick who did not play counts as a miss.

§ 05 — Night by night

Date	Players	Model MAE	DK avg. MAE	Value picks	Hit rate
2026-10-24	167	7.85	8.92	30	23%
2026-10-23	279	8.21	8.74	30	13%
2026-10-22	39	8.77	8.97	10	20%
2026-10-21	384	9.10	7.92	40	18%
2026-10-20	104	7.13	7.04	20	45%

§ 06 — The rules this is graded by

Saved before tip-off. Each player is graded on the last projection saved before his own game started. Nothing is edited afterward, and each row carries its save time.
No-shows count. A player who was projected and did not play is a miss, not a quietly removed row.
Same players, same test. The model and each baseline are scored on exactly the same players.
Waits for enough data. No verdict until 30 players have been graded, and the sample size sits next to every number.
Regular season only. Preseason games are tested privately and kept out of the public record.
Fantasy points use DraftKings NBA scoring. The box score is the only source for the actual result.

The full record, one row per graded projection with its save time, is published as scorecard.json so you can recompute every figure above. Last updated October 25, 2026. Model: v1 (2026-10-07): EWMA minutes + absence-band correction x 30-game points/minute.

This is a research record, not a promise of results. DFS involves risk, and a model that is accurate on average can still lose money.