Projecting games played and availability

Up: 2026-27 Fantasy Basketball Encyclopedia

Related: 2026-27 fantasy basketball projection methodology · 2026-27 fantasy basketball injury and availability ledger · 2026-27 fantasy basketball rankings · 2026-27 points-league rankings · 2026-27 category rankings · 2026-27 NBA minutes projections · 2026-27 projection uncertainty ledger · 2026-27 fantasy basketball evidence and confidence policy · 2026-27 sleepers, breakouts, and fades · Games played versus per-game value in fantasy basketball · How NBA injury reports should change fantasy decisions

Why GP dominates the error budget

Season total is \(T = G \times v\), where \(G\) is games and \(v\) is per-game fantasy value. In logs the errors separate:

\[\operatorname{Var}(\ln T) \approx \operatorname{CV}_G^2 + \operatorname{CV}_v^2 + 2\rho\,\operatorname{CV}_G \operatorname{CV}_v\]

For a mid-round veteran, a reasonable estimate (not a measured figure) is \(\operatorname{CV}_G \approx 0.20\) — a 65-game mean with a 13-game standard deviation — against \(\operatorname{CV}_v \approx 0.08\) for per-game value, which is a well-behaved quantity built from three or four prior seasons. Ignoring covariance:

\[\frac{0.20^2}{0.20^2 + 0.08^2} = \frac{0.0400}{0.0464} = 86\%\]

Roughly 86% of the uncertainty in a season total is games. Every hour spent refining a usage estimate is an hour spent on the 14%.

How predictive is prior-season GP?

No published NBA study giving a clean year-over-year autocorrelation of games played was located; treat any specific \(r\) you see quoted as unverified. What is published points at a modest but real persistent component:

• A multi-level model of injury events found \(r = 0.672\) between a player's average modeled injury probability and his actual injury count in that season — i.e. observable characteristics do carry substantial signal (Lewis, Sports Health).
• The same work found injury odds rose 2.87% per 96 minutes played and fell 15.96% per day of rest, so GP is partly endogenous: playing more causes injury, which suppresses next season's GP. This mechanically drags the autocorrelation toward zero and is why naive "he's injury-prone" reasoning overfits.
• Minutes per game was the single largest risk factor for season-ending injury, and season-ending injuries clustered in the third and fourth quarters of the schedule (J Oral Maxillofac Surg group, ScienceDirect).

Mechanistic conclusion: prior GP is informative mostly through what caused the absence, not through the raw count. A 55-game season from one 27-game Achilles block and a 55-game season from nine separate two-to-four-game soft-tissue absences are the same number and completely different forecasts. Decompose every prior season into blocks before using it.

Recurrence taxonomy

Blood clots (Wembanyama's 2024-25 DVT) belong in their own bucket: medically managed, low recurrence on anticoagulation, but the anticoagulation window itself can force games off. That is an administrative absence, not a tissue-failure absence.

Achilles and ACL: the actual base rates

Achilles rupture (37 NBA players, 1990–2023, Foot & Ankle Ortho / PMC):

The performance split is the useful part: 100% of reserves regained their level versus 22% of All-Stars and 0% of starters. Elite explosive production is the thing that does not come back. A separate NBA/WNBA comparison found shortened careers post-rupture (3.1 ± 2.3 vs 5.8 ± 3.5 remaining seasons).

ACL reconstruction is materially kinder (Brown University sports injury lab; PubMed 33738307; systematic review, ScienceDirect):

The single most exploitable number here is 48.4% → 62.1%. Year 1 back is a ~40-game season; year 2 is a ~51-game season with restored rates. This is the mechanism behind the classic "fade the return year, buy the second year" heuristic, and it is measured, not folklore.

Critically, the relevant clock is months since surgery, not seasons since last appearance. A player who sat out an entire season is entering his post-op year two even though it is his first year back — a strictly better cohort than someone who returned mid-season on an accelerated ramp.

Load management and the 65-game rule

The Player Participation Policy (2023-24 onward) gates MVP, All-NBA, DPOY and the rest behind 65 games at 20+ minutes. It has not fixed availability. In 2025-26 the ineligible list included Luka Dončić, Anthony Edwards, Cade Cunningham, Giannis Antetokounmpo, LeBron James and Stephen Curry, while Nikola Jokić, Victor Wembanyama and Kawhi Leonard cleared it only in the final week (Philadelphia Inquirer, 2026-04-06; Yahoo Sports).

Three fantasy consequences:

1. The rule creates a floor, not a ceiling. Award-relevant players are pushed toward 65–72 games and away from 55. Players with no award case face no such pressure.
2. It creates a cliff. Once a star is mathematically out of award range — usually late February — his marginal incentive to play collapses. Expect March/April shutdowns for eliminated teams. This is precisely fantasy-playoff season.
3. It creates perverse early returns. Players rush back to bank games, which is exactly the pattern hamstring and calf recurrence data punishes.

Back-to-backs remain the cleanest observable: the 15.96%-per-rest-day figure above means a healthy 30-year-old big on the second night is a genuinely different player. Track team-level B2B rest policy separately from injury.

Age interaction

The PLOS One epidemiology study (625 players, 3,543 injuries, 2017-18 to 2020-21) found the 30–34 bracket carried the highest injury-rate ratio, and that NBA experience above the 4.7-year league average raised risk independent of chronological age. The operational form: age does not raise the probability of an injury event linearly so much as it raises games lost per event and lowers the chance of a full return. Suggested prior, clearly an estimate:

Building an honest GP distribution

Do not project a point estimate. Project a three-state mixture and report the mean and the branch structure:

\[E[G] = \pi_H \mu_H + \pi_N \mu_N + \pi_M \mu_M\]

where \(H\) = clean season (μ ≈ 70–76 — nobody plays 82), \(N\) = nagging season of scattered absences (μ ≈ 55–62), \(M\) = major event (μ ≈ 20–40).

Worked example, Stephen Curry at 38. He played 43 games in 2025-26, including a 27-game runner's-knee block (NBC Sports Bay Area). Age 35+ prior, chronic-knee tag, no award incentive on a 37-win roster:

\[E[G] = 0.20(68) + 0.45(55) + 0.35(28) = 13.6 + 24.8 + 9.8 = 48.2\]

The vault carries 52. The mixture says 48 with an interquartile range of roughly 30–62. That is a two-round difference in a totals league and almost none in a per-game league — which is the whole point of the next section.

The \(\sqrt{GP/82}\) problem

build_encyclopedia.py sets nine_cat_total_z = per_game_z × sqrt(gp/82), and 2026-27 points-league rankings orders by FP/G × sqrt(GP/82). This is a deliberate compromise between per-game (exponent 0) and totals (exponent 1). It matches neither, and the error is large.

The correct season-total formulation for a roster slot accounts for the fact that missed games are backfilled by streamers at replacement level \(v_R\), with coverage fraction \(\rho < 1\) (roster limits, waiver quality, IL slots):

\[V_i = G_i v_i + (82 - G_i)\,\rho\, v_R = 82\rho v_R + G_i\,(v_i - \rho v_R)\]

Worked comparison. Player A: \(v = 1.20\) z-units, \(G = 76\). Player B (Curry-like): \(v = 1.60\), \(G = 52\). Take \(v_R = -0.30\) and \(\rho = 0.8\).

The square root says the fragile star is better; the correct totals math says he is worth 85% as much. That is a 25-point swing and it is systematic — it inflates every low-GP star on the board.

Two further defects. First, the index scales an already-aggregated z-score rather than each primitive category, so it never recomputes \(\sigma\) over the totals distribution of the rosterable population — and \(\sigma^{\text{tot}}_c \neq 82\,\sigma^{\text{pg}}_c\), because totals dispersion absorbs GP dispersion. Second, it treats turnovers symmetrically with counting stats, when in a totals frame extra games strictly add turnovers.

The correct computation: project \(G_i\) and per-game rates separately, form \(T_{ic} = G_i m_{ic}\) for every category, then z-score each category against the totals distribution of the actual rosterable pool, then sum. Ratio categories use summed makes over summed attempts, never a z-score of a percentage.

Which exponent for which format

Two modifiers that matter more than most drafters think: league depth (a 10-team league has a strong waiver wire, so \(v_R\) is high and \(\rho\) near 1, muting the GP penalty; a 12-team league amplifies it) and IL slots (two or more IL slots make a known long absence far cheaper than an unpredictable 55-game grind — predictability is worth more than the raw count).

2026-27: the highest-leverage availability questions

Leverage = draft cost × GP uncertainty. Statuses below are as of the 2026-27 fantasy basketball injury and availability ledger cutoff, 2026-07-18.

Open questions

• No published NBA year-over-year GP autocorrelation was found. Is the raw lag-1 correlation actually above ~0.3 once block-decomposed, or is nearly all the apparent persistence explained by age, minutes load, and a handful of structural injuries?
• What is the true value of \(\rho\) (missed-game coverage) in a 10- vs 12-team league? It drives the entire GP-discount magnitude and could be measured directly from a prior season's transaction logs, but has not been.
• Does the 65-game rule's late-season cliff show up as a measurable March/April GP drop for eliminated teams' award-ineligible stars? If so it should be modelled as a schedule-position effect, not a season-long mean.
• Should the encyclopedia publish two ranking columns — per-game and true totals — rather than one hybrid index? The \(\sqrt{\cdot}\) compromise appears to serve no format exactly.