Model Journal

How our model gets smarter every week.

Every projection tracked against reality. Every miss investigated. Every improvement documented here — with the data, reasoning, and SHAP analysis to prove it worked.

Looking for the season scorecard? That's on the Transparency page →

80.0%

Current accuracy

v27.1

Model version

Seasons validated

📋Journal vs Transparency page: This journal documents how the model evolves — improvement log, walk-forward validation history, SHAP analysis. The Transparency page is the permanent season record — locked projections vs actuals.

All entries

Walk-forward history

Improvement log

Weekly digestSeason start Oct

Journal entriesClick any entry to read the full analysis

All

Milestones

Walk-forward

Improvements

New Model

The Rookie Identity. Building a player from signals alone.

4 June 2026 · v27.1

v27.1

Every NBA rookie enters the league as an unknown. No career stats. No track record. Just signals. The Rookie Identity model builds a player from scratch — NBA Combine measurements, KenPom-adjusted college stats, and 25 years of historical comps. 550 training examples. FG%/FT% highly predictable. 3PM at 19.0% is the hardest rookie stat in all of sports analytics.

Read full entry →Rookie Identity · /rookie-identity live

Model Update

v27.1: +1pp improvement. Here's what 10 seasons of misses taught us.

4 June 2026 · v27.1

v27.1

Before retraining, we ran a post-mortem across 10 walk-forward holdout seasons (2016–2025). 1,549 player-seasons. Every miss investigated. 67% of all big misses (>30% error) had one thing in common: prior season games played < 55. The injury/availability signal is the single most actionable finding from a decade of honest testing.

Read full entry →10-season avg · 81.6% overall

Model Update

374 features. Real contract data. 80.0% overall. Here's what changed.

4 June 2026 · v27.1

v27.1

v27.1 adds 27 new features across two new passes of Layer 14 — directly addressing the three miss clusters from the multi-season post-mortem. Injury/availability flags, team context, teammate zero-sum signals, and 459 players with real 2026 contract data. 6 of 9 categories improved. Injury-return MAE still 23.8% — honest gap documented here.

Read full entry →80.0% overall · +0.1pp

Walk-forward validation

9 seasons tested. +1.3pp improvement. Here's everything we learned.

1 June 2026 · v26.0

v26.0

Walk-forward validation across 2016–2024. Train on everything before the test year, predict, measure, retrain. The model improved from 80.6% to 81.9% across 9 seasons. Blocks remains the hardest category (68–74%). FG%/FT% consistently the strongest (93–95%). The 2019-20 bubble year shows the expected accuracy dip.

Read full entry →80.6% → 81.9% ↑

Milestone

The model has learned. Here's what it knows.

1 June 2026 · v26.0

v26.0

Nine XGBoost models trained on 3,813 player-season records spanning 2001–2015 with recency weighting. First holdout test on 2016 data. Rankings sanity gate passed — Wembanyama #1, Jokić #3. This is the model's first look at real NBA history.

Read full entry →80.6% overall

Milestone

347 features across 12 layers — the pipeline that feeds the model.

May 2026 · v26.0-pre

v26.0-pre

The complete feature engineering pipeline powering the prediction model. 347 features spanning player talent baseline, rolling momentum windows, opportunity signals, lineup context, injury ripple effects, team and coach context, opponent matchups, schedule effects, age curves, volatility modelling, market signals, and archetype embeddings. Zero DB errors across 150-player integration test.

Read full entry →347 features · 9/10 checks

Walk-forward validation historyThe honest proof — 9 independent test seasons

What walk-forward validation means

Train on data up to 2015. Predict 2016 without ever having seen it. Measure the error. Retrain incorporating 2016. Predict 2017. Repeat through 2024. Nine completely independent test seasons. The numbers below are real — every season we ran, in order, nothing removed.

Independent test seasons

+1.3pp

Improvement across all seasons

2,345

Player-season predictions made

Overall accuracy by season

80.6

15–

81.4

16–

82.2

17–

81.3

18–⚑

82.1

19–

81.4

20–

81.7

21–

82.3

22–★

81.9

23–

⚑ 2018–19 = COVID bubble · ★ 2022–23 = peak season

Season	PTS	REB	AST	STL	BLK	3PM	FG%	FT%	TO	Overall
2015–16	82.3%	80.7%	76.3%	78.3%	68.3%	72.7%	93.8%	94.8%	78.3%	80.6%
2016–17	82.5%	81.1%	77.1%	76.8%	74.2%	73.2%	94%	94.9%	78.7%	81.4%
2017–18	84.7%	82.8%	78%	76.9%	72.4%	75.6%	94.1%	94.7%	80.8%	82.2%
2018–19⚑ bubble	82.4%	82.8%	76.8%	76.9%	71.4%	72.9%	94.1%	94.3%	80.2%	81.3%
2019–20	84.2%	83.5%	78.3%	78.3%	70.6%	76.4%	94%	94.3%	79.1%	82.1%
2020–21	81.7%	83.1%	75.9%	77.5%	71.5%	75%	94%	94.8%	78.7%	81.4%
2021–22	83.4%	81.9%	77.6%	77.8%	69.4%	76%	93.6%	94.8%	81%	81.7%
2022–23★ peak	84%	84.1%	78.6%	79%	70.7%	75.8%	94.2%	95.2%	79.5%	82.3%
2023–24	83.4%	83.4%	77.1%	75.3%	73.6%	76.9%	94.4%	94.3%	78.8%	81.9%
Average	83.2%	82.6%	77.3%	77.4%	71.3%	75%	94%	94.7%	79.5%	81.6%

Season diaryWhat we found and what we changed after each test

2015–16 — First testBaseline · trained on 2001–2015

80.6%

Accuracy by category

PTS

82%

REB

81%

AST

76%

STL

78%

BLK

68%

3PM

73%

FG%

94%

FT%

95%

78%

What we found

Strong baseline. FG%/FT% excellent — percentage categories are inherently stable. Blocks at 68% was our biggest gap — high game-to-game variance makes it genuinely hard. FG% was systematically under-predicting for forwards across all usage tiers. Retrained with 2016 data incorporated before next test.

Biggest miss: Rashad Vaughn FT% — projected 60.6, actual 40.0 (Δ +20.6pp). Erratic young shooter.

2018–19 — Bubble seasonCOVID / Orlando bubble · unusual conditions

81.3%−0.9pp

Accuracy by category

PTS

82%

REB

83%

AST

77%

STL

77%

BLK

71%

3PM

73%

FG%

94%

FT%

94%

80%

What we found

The bubble was genuinely unpredictable. No home court, no crowds, neutral site for all games — conditions our model had never seen in training data. PTS dropped 2.3pp, 3PM dropped 2.7pp. This is an expected and honest result, not a model failure. We flagged 2019-20 as a reduced-weight season in subsequent training. Did not try to “fix” the model on unprecedented data.

Biggest miss: Thabo Sefolosha FT% — projected 71.0, actual 37.5 (Δ +33.5pp). Bubble conditions affected free throw mechanics.

2022–23 — Peak season

82.3%★ best

Accuracy by category

PTS

84%

REB

84%

AST

79%

STL

79%

BLK

71%

3PM

76%

FG%

94%

FT%

95%

80%

What we found

Our best season yet. Broad gains across REB (+2.2pp), STL (+1.2pp), BLK (+1.3pp), AST (+1.0pp). No systematic bias detected in any category — the most balanced result across the full walk-forward run. The archetype clustering (Layer 12) contributed significantly to REB and BLK improvement. No changes required — model incorporated 2023 data and continued.

Biggest miss: Reggie Bullock Jr. FT% — projected 77.4, actual 100.0 (Δ −22.6pp). Perfect FT% shooter we underestimated.

SHAP analysisWhat features drove the model's biggest decisions

SHAP waterfall — FT% projectionHassan Whiteside · 2018-19 season

Our biggest systematic miss category is FT% for erratic shooters. This SHAP chart shows exactly why the model overestimated Whiteside's FT% — it correctly weighted his prior year data, but that data couldn't predict the dramatic drop that followed.

Pushes projection higher

Pushes projection lower

ft_pct_last_season

+9.1pp

ft_pct_career_avg

+6.3pp

ft_pct_3yr_avg

+3.8pp

ft_attempt_rate

+2.2pp

is_known_poor_ft_shooter

−4.8pp

ft_pct_sustainability

−1.9pp

Base value: 68.3% (league avg FT%) → Projected: 66.5% → Actual: 44.9%

Projected FT%

66.5%

Miss: Actual FT% was 44.9% — a 21.6pp gap. The model had no feature to capture the kind of dramatic intra-career FT% collapse that Whiteside experienced. This motivated the ft_pct_sustainability feature investigation.

Current season accuracy

Loading grades…

Model status · v27.1

80.0%

Overall

68.5%

BLK ⚠️

✓

Gate

Injury-return MAE still 23.8% (target 9.5%). Only ~200 injury-return training examples across 24 seasons. Gap narrows each year. v27.2 target: <15%.

Walk-forward validation10 seasons ✓

SHAP explanations374 features ✓

Contract data459 players ✓

Rookie IdentityLive ✓

Injury-return MAE< 15% target

Season projectionsOct 2026

Accuracy by category

v27.1 · 10-season walk-forward

FT%

95%

FG%

93.9%

REB

80.9%

PTS

79.7%

STL

76.4%

AST

75.8%

3PM

73.8%

BLK

68.5%

Version history

v27.1 — 374 features + contract

27 new Layer 14 features. Injury/availability flags. Contract data for 459 players. 80.0% overall.

4 Jun 2026

v26.0 — baseline

9 XGBoost models trained 2001–2015. Walk-forward validated 2016–2024. 81.9% overall.

1 Jun 2026

v26.0-pre — feature pipeline

347 features across 12 layers. 9/10 integration checks. Zero DB errors.

May 2026

Coming next

NextInjury-return MAE to <15% — v27.2 target. More training examples accumulate each season.

ThenHistorical contract data — Spotrac API ($TBD/quarter) closes the 2016–2025 gap.

ThenFantasy value translation — what the raw stats are actually worth in your specific league.

Oct 262026-27 projection lock ceremony. Season projections published and immutable.

What is SHAP?

SHAP (SHapley Additive exPlanations) shows which features drove each prediction and by how much. Blue bars push the projection higher. Pink bars lower. The length shows the magnitude. It makes the model's reasoning visible — not just what we predicted, but exactly why.