APEX · Transparency
DATA SOURCES
APEX is built on publicly available data and peer-reviewed research. Every metric source, every academic study that shaped a design decision — documented here with how it was used.
Statistical Providers
Three sources feed the live scoring pipeline. Each is acquired separately and merged by player name before scoring runs.
Basketball-Reference
basketball-reference.com
OBPM
DBPM
AST%
AST/TOV
TOV%
REB%
USG%
FTA/FGA
TS%
WS/48
VORP
GP / MIN
How obtained: Scraped via BeautifulSoup (
Why it matters: The only publicly available source with full historical box-score data back to 1973–74, making it essential for the 40-season backtest. Provides the box-score inputs (AST%, REB%, TOV%) and the OBPM/DBPM pair used as Bayesian stabilizers.
fetch_bref.py). Rate-limited to 4 seconds per request to respect the site's policy. Multi-team players use the TOT (combined) row. Free, no account required.
Why it matters: The only publicly available source with full historical box-score data back to 1973–74, making it essential for the 40-season backtest. Provides the box-score inputs (AST%, REB%, TOV%) and the OBPM/DBPM pair used as Bayesian stabilizers.
BBall-Index
bball-index.com
O-LEBRON
D-LEBRON
Rel TS+
Off. Role
Def. Role
How obtained: Automated via Playwright (
Why it matters: The only public source for LEBRON's offensive and defensive components separately. O-LEBRON and D-LEBRON are the primary win-probability contribution metrics in the Offensive and Defensive Impact pillars respectively. Relative TS+ provides shot-difficulty adjustment that raw TS% cannot. The $5/month Data Tools subscription is the only paid dependency in the pipeline.
fetch_bball_index.py) for the current season — logs in, navigates the Shiny leaderboard, exports all rows. Historical seasons require a manual Excel export saved to data/bball_index_YEAR.xlsx. The pipeline detects both export formats automatically.
Why it matters: The only public source for LEBRON's offensive and defensive components separately. O-LEBRON and D-LEBRON are the primary win-probability contribution metrics in the Offensive and Defensive Impact pillars respectively. Relative TS+ provides shot-difficulty adjustment that raw TS% cannot. The $5/month Data Tools subscription is the only paid dependency in the pipeline.
Dunks & Threes
dunksandthrees.com
O-EPM
D-EPM
EPM (tot)
How obtained: Manual CSV export per season, saved to
Why it matters: EPM is a regularized adjusted plus-minus estimator — it controls for teammate quality in a way box-score metrics cannot. O-EPM anchors the Offensive Impact pillar (40% weight) because it captures offensive win-probability contribution independent of defensive signal. D-EPM anchors the Defensive Impact pillar (43% weight) and was ranked the #1 most trusted public defensive metric by a HoopsHype survey of ~30 NBA executives.
data/dunks_threes_YEAR.csv. The off column is O-EPM; def is D-EPM; tot is the full two-way composite (not scored in v3.0).
Why it matters: EPM is a regularized adjusted plus-minus estimator — it controls for teammate quality in a way box-score metrics cannot. O-EPM anchors the Offensive Impact pillar (40% weight) because it captures offensive win-probability contribution independent of defensive signal. D-EPM anchors the Defensive Impact pillar (43% weight) and was ranked the #1 most trusted public defensive metric by a HoopsHype survey of ~30 NBA executives.
Scored metrics (highlighted above) are directly weighted in APEX pillar scores. Display-only metrics — WS/48, VORP, full two-way EPM/LEBRON/BPM — appear in player cards for reference but carry zero weight. Including VORP in scoring would introduce voter-narrative signal into a model designed to be independent of voters (ACM ICSTPA 2024: VORP is the #1 predictor of MVP voting, not player value).
Academic Research
These studies shaped the model's architecture, metric selection, and known-limitation documentation. Click any entry to read the key finding and how it was applied.
Franks, D'Amour, Cervone & Bornn · JQAS Vol. 12(4) · 2016
Three criteria for useful composite basketball metrics: stability, discrimination, independence
RAPM-family metrics score highest across all three criteria. On/off metrics are discriminative but lower-stability than composite estimators.
How APEX uses it
The primary citation for the five-pillar independence design — each pillar must provide information the others do not. Also the formal basis for DBPM's 10% weight in the Defensive Impact pillar as a Bayesian stabilizer: it prevents D-LEBRON from overreacting to single-season lineup outliers, not because DBPM carries genuine defensive signal on its own.
Terner & Franks · Annual Review of Statistics and Its Application, Vol. 8 · 2021
Counting and ranking: the most comprehensive peer-reviewed basketball analytics survey
DRtg and DWS are "particularly sensitive to teammate performance and thus not reliable measures of individual" defense. RAPM-family metrics are the most valid framework for individual player valuation.
How APEX uses it
The most authoritative single citation in the model. Validates the multi-estimator blending approach, the removal of DWS from the Defense pillar (v1.7), and the known limitation that team context contaminates all on/off-derived metrics including D-LEBRON.
Deshpande & Jensen · JQAS Vol. 12(2) · 2016
Box-score metrics fail to account for game context; RAPM family is the superior individual value framework
RAPM-family estimators are the most predictive of player value among all publicly available basketball metrics.
How APEX uses it
Canonical justification for the Offensive Impact pillar's 32% weight and the decision to anchor it with RAPM-family metrics (O-EPM, O-LEBRON, OBPM) rather than box-score totals. Also the primary basis for excluding PER — a box-score composite — from scoring.
Brill et al. · Quarterly Journal of Economics · 2023
K-means clustering on 48 playing-style variables identifies 8 stable functional archetypes
Traditional 5-position labels are poor proxies for how players actually function. Data-driven archetypes are more stable and predictive.
How APEX uses it
Primary citation for the K-means offensive archetype normalization introduced in v2.3. Shot Quality, Creation & Playmaking, and Physical Contribution metrics are z-scored within each player's archetype peer group (8 clusters) rather than the full pool. This resolved the rim-running big overrating problem: a Rim Runner's TS% is now compared against other Rim Runners, not guards with half the usage.
Jewell, Page & Reese · JQAS Vol. 19 · pp. 293–302
BPM is most robustly significant at high minutes loads; STL%, BLK%, and AST show null significance at the same tier
At high minutes loads — the players APEX evaluates — steals percentage, blocks percentage, and raw assist count add no statistically significant predictive signal over BPM.
How APEX uses it
The peer-reviewed basis for removing STL%, BLK%, DWS, and AST (raw) from scoring in v1.7. These metrics are absent from the model not because they're hard to obtain, but because they do not carry independent information at the minutes thresholds APEX qualifies. The null-significance finding is specific to the ≥1,000-minute player pool.
Byman · MSDS Thesis · 2023
Off-ball defensive movement features dominate OLS coefficients in RAPM prediction (R²=0.848)
Lasso regression using player tracking features predicts RAPM with R²=0.848. Off-ball movement metrics dominate; box-score defensive stats are weak predictors.
How APEX uses it
Grounds the defense known limitation: the metrics APEX can access (D-LEBRON, D-EPM, DBPM) are the best available public proxies, but the dominant predictors of defensive value are off-ball movement features locked behind NBA.com's paid tracking API. This is why the model acknowledges its defensive pillar as its weakest point — not from lack of rigor, but from lack of data access. Also recommends Lasso for future empirical weight optimization.
Kubatko et al. · Journal of Quantitative Analysis in Sports, Vol. 3(1) · 2007
TS% is the most predictive single-metric proxy for offensive scoring efficiency; FTA rate is a persistent offensive differentiator
True Shooting Percentage outperforms eFG%, FG%, and PPG as a proxy for offensive efficiency. Free throw attempt rate is a stable individual differentiator across seasons.
How APEX uses it
Primary citations for TS% and FTA/FGA inclusion. In v3.0, raw TS% was removed from scoring (it's already encoded inside O-EPM, O-LEBRON, and OBPM), but Relative TS+ — a difficulty-adjusted variant — is retained in the Shot Quality pillar because it provides signal not captured in composite inputs. FTA/FGA moved from Shot Quality to Creation & Playmaking in v3.0, where it belongs as a foul-drawing / shot-creation proxy rather than a shooting-efficiency metric.
Dehesa et al. · Kinesiology 51(1):92–101 · 2019
In close games, box-score KPIs show no statistically significant cluster separation; NET/ON/OFF metrics dominate
Cluster analysis of 472 players across 535 close games (margin ≤8 pts). NET/ON/OFF metrics show F > 1,499, p < .001. Box-score KPIs show p > .05, ES ≈ 0 — essentially random in high-leverage situations.
How APEX uses it
Grounds the known limitation that Scoring and Playmaking pillar metrics are less discriminating in high-leverage situations. Also supports the Offensive Impact pillar's 32% weight — RAPM-family metrics (O-EPM, O-LEBRON) remain discriminative in close games where box-score stats fail.
ACM ICSTPA · 2024
SHAP heatmap (CatBoost, 1947–2024 MVP data): VORP is the #1 predictor of MVP voting
Full SHAP feature ranking: VORP #1, PPG #2, PER #3, USG% #4, BPM #5. The model predicted Jokić as 2024 MVP. VORP is more predictive of voter behavior than of player value.
How APEX uses it
The formal basis for keeping VORP display-only. If VORP is the strongest predictor of how voters vote, including it in APEX's scoring engine would make APEX a voter-behavior model rather than a player-value model — exactly what it is designed not to be. VORP appears in player cards as a reference signal, never in the score. GP^0.75 availability modifier is grounded by the v1.4 backtest independently of this paper.
Gong, Whitehead et al. · JQAS · 2024
Hierarchical Bayesian plus-minus across 2015–2022: gap between offensive position groups has narrowed over time
Position-based performance gaps are shrinking as the NBA evolves toward positionless play.
How APEX uses it
Supports position normalization (v1.2) and contemporary z-scoring over fixed historical baselines. As positional gaps narrow, comparing players within archetype peer groups (v2.3) rather than fixed position buckets becomes more accurate. Also supports the known limitation that era-neutral z-scoring has caveats for historical comparisons.
HoopsHype survey of ~30 NBA executives · 2024
D-EPM ranked #1 most trusted public defensive metric among front-office respondents
Endorsed independently by EPM/RPM co-creator Steve Ilardi as "the obvious gold standard" for public defensive measurement.
How APEX uses it
Primary real-world validation for D-EPM at 43% of the Defensive Impact pillar (v2.6). D-LEBRON retains a slight primacy (47%) due to its luck adjustment, which handles single-season shooting variance better than RAPM components. D-EPM's rise from 0% to 43% across versions reflects accumulating evidence — this survey was the final confirmation that the practitioners closest to the game trust it most.
Zimmer, Snyder & Zimmer · Economics Bulletin, Vol. 45, Issue 2 · 2025
Standard plus/minus lacks opponent quality controls
Raw plus-minus is systematically biased by opponent strength, making it an unreliable measure of individual contribution.
How APEX uses it
Peer-reviewed justification for using RAPM-family metrics over raw on/off differential. Raw plus-minus does not appear in any APEX pillar. All impact metrics use regularized or regression-adjusted estimators that control for lineup quality.
Park · NBA Player Salary Determinants · 2025
Salary is most strongly predicted by narrative metrics (PPG, MPG), not impact metrics
Contract value reflects media visibility and scoring volume more than adjusted impact measures.
How APEX uses it
Establishes that salary is an invalid validation target for APEX. A model that correlates well with salary would be measuring narrative, not value. APEX is validated against backtest MVP accuracy (25/40 seasons, 62.5%), not against contract data.
MDPI Big Data & Cognitive Computing · DOI: 10.3390/bdcc9040074 · 2025
Net Rating and Player Impact Estimate consistently correlate with winning; defensive metrics more critical in late-game
AST% peaks in Q2 affecting early momentum. Defensive metrics show greater predictive power in late-game situations.
How APEX uses it
Supports Impact pillar primacy and AST% retention. The finding that AST% is a leading indicator of game momentum (Q2 peak) corroborates keeping it in the Creation & Playmaking pillar. The late-game defensive primacy finding supports the Defensive Impact pillar's weight increase to 28% in v3.0.
Historical Backtest Sources
The 40-season backtest (1985–86 through 2024–25) uses a different source set from the live pipeline. Some metrics are unavailable for older seasons.
| Source | Metrics | Coverage | Notes |
|---|---|---|---|
| Basketball-Reference | BPM, DBPM, OBPM, TS%, AST%, TOV%, REB%, USG%, FTA/FGA, GP, MIN | Full history to 1973–74 | Primary source. OBPM used for backtest Offensive Impact in absence of EPM/LEBRON for pre-2014 seasons. |
| BBall-Index | LEBRON, D-LEBRON, O-LEBRON, Rel TS+ | Limited historical depth | Full LEBRON data available from ~2014 onward. Earlier seasons use BPM family only in Impact pillar. |
| Dunks & Threes | EPM (Season/Actual) | Limited historical depth | EPM data available from ~2014 onward. Pre-EPM era explains the lower backtest accuracy for 1985–2000 seasons. |
The 40-season backtest accuracy of 62.5% (25/40) understates model quality for recent seasons. The 25-season window (2000–01 through 2024–25) shows 76% accuracy (19/25), as RAPM-family metrics become available and the model reaches full specification. The pre-EPM/LEBRON era is a structural data constraint, not a model flaw.