Category correlations and roster construction

Up: 2026-27 Fantasy Basketball Encyclopedia

Related: 2026-27 volume-corrected nine-category board · Projection engine defects and corrections · 10-team versus 12-team fantasy basketball strategy · 2026-27 category rankings · 2026-27 fantasy basketball rankings · 2026-27 points-league rankings · 2026-27 fantasy basketball tiers · 2026-27 fantasy basketball player projections · 2026-27 fantasy basketball projection methodology · 2026-27 NBA minutes projections · 2026-27 sleepers, breakouts, and fades · 2026-27 fantasy basketball injury and availability ledger · 2026-27 fantasy basketball evidence and confidence policy · Percentage categories and volume · Robust punt builds in fantasy basketball · Games played versus per-game value in fantasy basketball

Data basis — rebuilt on corrected data 2026-07-18 The correlation matrix (§1) and the concentration table (§2) are computed from the counting-stat columns of player-projections-2026-07-18.csv across the 168 players with at least 40 projected GP, 24 MPG, and complete counting lines. Those two sections were re-run against the corrected export and reproduce to the second decimal, unchanged. That is not luck: the July 18 defect fabricated shot attempts, and 3PM/PTS/REB/AST/STL/BLK/TO are none of them. The correlation matrix never touched the broken columns.

Sections 3, 4 and 5 did, and have been rebuilt on Data/player-projections-corrected.csv — see 2026-27 volume-corrected nine-category board and Projection engine defects and corrections. §3 in particular no longer proposes a repair; the repair has been carried out.

Correction 2026-07-19: the whole note was recomputed once more, end to end, against Data/player-projections-corrected.csv. §1, §2, §4 and §5 reproduce to the digit shown. §3 did not survive intact: four of the twelve "underpriced" rows were players whose attempts were stated in the July 18 export — the proxy never fired for them — and have been replaced. See the dated correction in §3.

These remain correlations among this desk's projections, not among realized NBA outcomes — projections are smoother than reality, so true in-season correlations will be somewhat weaker. Treat the signs and the ordering as robust; treat the second decimal as decoration.

1. The correlation matrix

Pearson \(r\) across the 168-player pool:

Four structures fall out.

The perimeter cluster (3PM–PTS–AST). Threes correlate \(+0.48\) with points and \(+0.33\) with assists. Buying one buys some of the others. The cost is the mirror image: 3PM correlates \(-0.48\) with rebounds and \(-0.39\) with blocks. A roster that wins threes by construction is losing the frontcourt categories by construction.

The interior cluster (REB–BLK–FG%). \(+0.66\) between rebounds and blocks is the tightest positive pair outside the turnover column. It is also the cluster that carries FG% along with it — the FG%-impact leaders on the 2026-27 category rankings board are almost entirely the same names as the block leaders.

AST and TO are effectively one category. \(r=+0.79\). There is no meaningful population of high-assist, low-turnover players; Mikal Bridges and Davion Mitchell are notable precisely because they are rare. PTS–TO at \(+0.74\) says the same thing about scoring. In practical terms, 9-cat is closer to 8-cat than the category count suggests, because the ninth category is a near-deterministic tax on two others. This is the structural case behind the 8-cat versus 9-cat debate: if you draft for counting stats, turnovers largely draft themselves.

STL is the free variable. Mean absolute correlation with the other six counting categories, lowest to highest:

| Category | Mean |r| | Strongest entanglement | | --- | ---: | --- | --- | --- | | STL | 0.184 | AST (+0.39) | | BLK | 0.253 | REB (+0.66) | | REB | 0.279 | BLK (+0.66) | | 3PM | 0.342 | PTS (+0.48), REB (−0.48) | | PTS | 0.352 | TO (+0.74) | | TO | 0.379 | AST (+0.79) | | AST | 0.412 | TO (+0.79) |

Steals are the only category you can add to a roster without materially moving anything else. That is what "efficiently acquired" means — not that steals are cheap in ADP, but that a steals purchase has almost no side effects. Dyson Daniels (2.0), Ausar Thompson (2.0), and Cason Wallace (1.9) sit at the top of an otherwise flat distribution, and none of them forces a positional or category commitment the way an elite blocker does.

2. Scarcity: concentration, not level

Concentration across the same 168-player pool. Gini here measures how unequally a category's total production is distributed — higher means more top-heavy.

The operational reading:

• Blocks are a different kind of category. Twelve players hold 22% of the pool's blocks, and the leader produces 6.4× the median. Victor Wembanyama at 3.2 is worth roughly six median players in that column. This is why one specialist wins blocks: the category is so concentrated that a single acquisition moves your weekly total by more than a full standard deviation. It is also why blocks cannot be repaired late — see §5.
• Points and steals are the flattest distributions. A Gini of 0.166 for points means the difference between an elite scorer and a median rostered scorer is 1.9×, not 6×. You cannot win points with one player; you win it with nine or ten players each a little above median. Same shape, different scale, for steals.
• Steals are flat and independent. That combination is the reason the category is winnable cheaply: the marginal cost of upgrading a 1.0-steal player to a 1.4-steal player is small, and it does not drag anything else with it.
• Assists have the sharpest early cliff despite a middling Gini, because the top of the distribution is role-locked. Only five players project at 9.0+ and twelve at 7.0+. Losing a lead-guard role collapses the line in a way that losing rebounding minutes does not.

3. Ratio categories: attempts are the whole story

FG% and FT% are not player attributes that average. They are team-level ratios, computed from summed makes over summed attempts. A player's contribution is therefore his surplus makes relative to a reference rate:

\[ I_{FG} = FGA \times (FG\% - .472), \qquad I_{FT} = FTA \times (FT\% - .780) \]

The references are the neutral yardsticks the projection engine itself uses; the roster-relevant benchmark shifts as the roster does.

Worked example — the 90% shooter. Two players both shoot .900 from the line:

• 2.0 FTA: \(2.0 \times (.900 - .780) = \mathbf{+0.24}\) surplus makes per game
• 9.0 FTA: \(9.0 \times (.900 - .780) = \mathbf{+1.08}\) surplus makes per game

Identical rate, 4.5× the leverage. On corrected attempt projections this is why Cam Spencer's .920 on 1.4 attempts (+0.196) is a rounding error while Shai Gilgeous-Alexander's .880 on 8.9 attempts (+0.890) is a roster-defining asset — a lower rate producing 4.5× the impact. The sharpest illustration on the corrected board is Payton Pritchard: .889 from the line, which sounds like an elite ratio asset, on 1.6 projected attempts, which makes it worth +0.17 — less than a fifth of SGA's contribution at a higher percentage. Pritchard fell 36 → 57 on the corrected board almost entirely for this reason.

And the damage runs the same way. Devin Booker contributes \(7.5 \times (.875-.780) = +0.712\). One high-volume FT drag can erase him outright — Giannis Antetokounmpo at .665 on 11.6 projected attempts is \(\mathbf{-1.33}\), which cancels Booker and half of SGA together — and the proxy had charged him barely half that volume (≈5.9 FTA), so restoring it is the single reason he fell 21 → 46 on the corrected board despite his FG% term gaining from the repair. This is the arithmetic that makes "punt FT%" a coherent build rather than a resignation: once the drag is large enough, the surplus shooters you would need to offset it cost more than the category is worth.

The vault's attempt gap — closed 2026-07-18

An earlier version of this section argued that the encyclopedia's ratio categories were computed on a fabricated attempts proxy, and proposed a reconstruction. The gap has since been closed. This section now records what the repair found rather than proposing one.

The defect: only 120 of 549 rows in the July 18 export carried real FGA and FTA. For the other 429 the build script silently substituted

\[ FGA \approx \frac{PTS}{2.15}, \qquad FTA \approx \frac{PTS}{4.8} \]

and then scored two of the nine categories against those invented numbers. Full diagnosis in Projection engine defects and corrections.

The repair, in Data/recompute_corrected.py: real 2025-26 per-game attempts were recovered from ESPN and StatMuse for 262 pool players, free-throw volume was projected forward by scaling with minutes, and field-goal attempts were then solved from the scoring identity rather than guessed:

\[ PTS = 2 \times FGA \times FG\% + 3PM + FTA \times FT\% \;\Longrightarrow\; FGA = \frac{PTS - 3PM - FTA \times FT\%}{2 \times FG\%} \]

Of the 262 players, 54 already carried stated attempts, 194 were repaired with recovered data, and 14 rookies without an NBA season fall back to the pool-median free-throw rate and are flagged estimated in the export. Nothing was invented.

Validating the old proxy against the 248 players who now have non-estimated attempts — a five-fold larger test than the 45-player check the earlier version of this section could run:

The earlier estimate held up: the FTA proxy is roughly unbiased on average while being badly wrong for individuals, and the FGA proxy understates attempts by about four per game for nearly everyone. One earlier claim did not survive. The proxy's FGA error was reported as correlating \(+0.458\) with FG%; measured against real recovered attempts it correlates \(\mathbf{-0.361}\). The sign was backwards, so the mechanism described — "the proxy shrinks every FG% impact toward zero" — was right about the shrinkage but wrong about who it favoured.

Delta below is corrected combined ratio impact minus the impact the proxy assigned. Ranks are corrected → the old defective rank in brackets.

Underpriced by the proxy:

Correction 2026-07-19: an earlier version of this table also credited Nikola Jokić, Mark Williams, Zach Edey and Walker Kessler as proxy-underpriced. All four carried stated attempts in the July 18 export — the fabrication clause never fired for them, so their old and corrected ratio impacts are identical and their true Δ is zero; the "proxy impact" previously shown for them was a counterfactual the defective board never used. Their rank moves (Edey 53 → 72, Mark Williams 63 → 79) are displacement — the players around them were repriced, they were not. The four slots now hold the largest genuine movers: Wembanyama, Robert Williams III, Anthony Davis and John Collins.

Overpriced by the proxy:

The systematic direction the earlier version identified was correct and is now measured rather than estimated: efficient interior volume was underpriced by the mechanical rank, inefficient perimeter volume was overpriced. Zubac (93 → 73) and Poeltl (153 → 135) are the cleanest gainers; Lillard (81 → 101) and Pritchard (36 → 57) the cleanest losers.

One earlier example inverted. Rudy Gobert was listed above as underpriced by +0.50, on the reasoning that his FG% impact was being computed on too few shots. That was true of his field-goal term in isolation, but it ignored the free-throw side: the proxy assigned him roughly 2.1 FTA when his real projected volume is 3.8, and at .560 that extra volume is a large negative. Netted correctly, Gobert is one of the board's biggest losers from the repair — 99 → 138, the single largest fall in the pool. The lesson is that the two ratio terms have to be corrected together; fixing FGA alone would have moved him the wrong way.

4. Roster construction

Both Yahoo and ESPN default to 13 players — 10 starters, 3 bench. Yahoo's shape is PG, SG, G, SF, PF, F, 2×C, 2×Util; ESPN's is PG, SG, SF, PF, C, G, F, 3×UTIL. Two mandatory centre slots on Yahoo is a real constraint: it forces exposure to the REB–BLK–FG% cluster whether you wanted it or not.

Given the correlation structure, a balanced 13-man build:

One placement here is new with the repair. Trae Young rose 55 → 33 on the corrected board — the pool's top assist projection (10.1) had been paid for roughly 4.6 phantom free-throw attempts when his real volume is 8.1, and at .875 that restored volume nearly doubles his FT% surplus (+0.44 → +0.77). He is now the cleanest one-stop assist purchase available, rather than a mid-round luxury.

10 versus 12 teams. Rostered depth is \(10 \times 13 = 130\) against \(12 \times 13 = 156\). The marginal drafted player in the vault's rank order:

Twenty-six extra rostered players is a modest depth change but a large strategic one. In 10-team, replacement level is high enough that a hard punt is usually unnecessary — the waiver wire still holds 18-point scorers, and balanced construction is the default recommendation for shallow formats (Yahoo's 9-cat primer). In 12-team, the wire thins fast enough that committing to the correlation structure — deliberately buying one cluster and conceding the other — becomes the higher-EV play. See Robust punt builds in fantasy basketball before hard-punting anything, and 10-team versus 12-team fantasy basketball strategy for the full replacement-level computation on corrected data.

5. What can and cannot be streamed

Treating the 12-team waiver wire as corrected ranks 157+ with 50+ GP and 18+ MPG (68 players), counting how many clear a useful threshold:

• Threes are the most repairable category in the game. Klay Thompson (2.7), Max Strus (2.5), Tim Hardaway Jr. (2.5) and Grayson Allen (2.3) all sit outside a 12-team roster. The correction deepened this wire from 10 names to 12, because volume three-point shooters were precisely the group the proxy had overpaid on phantom free-throw volume. Never pay an early premium for ordinary threes.
• Points got harder to stream, not easier. The wire's 15-point sources fell from 5 to 3 (RJ Barrett 18.5, Darius Acuff Jr. 17.8, Shaedon Sharpe 16.8), because Jalen Green and Bennedict Mathurin both rose onto 12-team rosters. This is the one streaming conclusion the correction moved materially.
• Elite blocks cannot be streamed. The wire tops out at Brook Lopez and Isaiah Stewart at 1.2. Against a Wembanyama roster that is not a repair, it is a rounding adjustment. Combined with the 0.382 Gini, this is the strongest single argument for drafting blocks early or conceding them deliberately.
• Steals are the exception that proves the distribution point. Only Kris Dunn (1.4) and Jeremiah Fears (1.2) clear 1.2 on the wire — yet steals remain the easiest category to fix, because the distribution is so flat (Gini 0.151, median 1.0) that the wire's 1.0–1.1 options are already near league-average. You are not streaming steals to gain an edge; you are streaming them to avoid a deficit, and that is cheap.
• FT% impact cannot be streamed at all, and the corrected attempt data makes this quantitative rather than assertive. Impact scales with attempts, and the entire 68-player wire projects low attempt volume. A .900 shooter on 1.5 FTA moves your team ratio by +0.18 makes a game; Giannis alone is −1.33. It would take seven such shooters to offset him, and the wire does not hold seven. Ratio categories must be solved in the draft.

Open questions

• These are correlations among projections, which are smoother than realized seasons. What do the same seven pairs look like on actual 2025-26 per-game data, and how much does projection smoothing inflate the AST–TO and REB–BLK figures specifically?
• Does the identity-solve hold as tightly for low-usage bigs? Partly answered 2026-07-18. The repair no longer relies on an FTA regression at all — free-throw volume is scaled from each player's own recovered 2025-26 rate. The residual question is narrower and now concerns the 14 rookies with no NBA season, who take the pool-median free-throw rate and are flagged estimated. For those 14 the ratio categories are still, honestly, a guess.
• The repair scales attempts linearly with projected minutes. That assumes free-throw rate per minute is stable across a minutes change, which is plausible for role continuity but doubtful for a player moving from bench to starter, where usage typically rises faster than minutes. Nobody has measured the size of that bias on this pool.
• The FGA proxy error was reported here as correlating \(+0.458\) with FG% and measures \(-0.361\) against real attempts. The earlier figure came from a 45-player subset that skewed toward high-usage perimeter players. How much of the sign flip is sample composition rather than the earlier method being wrong?
• The streaming table assumes the wire equals ranks 157+, which ignores in-season injury and role churn — historically the largest source of streamable value. How much does a realistic mid-season wire differ from a static July rank cutoff?
• Does the G-score framework, which adjusts z-scores for week-to-week variance, change the specialist-versus-balance conclusion for the two highest-variance categories, STL and BLK?