PASCAL Model
Understanding the model that powers our NBA Draft rankings and projections
By: ReverseEnigma
Purpose
PASCAL estimates each prospect's Estimated Plus‑Minus (EPM) for seasons 1‑5, then ranks the entire class by those projections. Exact EPM values give context, but the ordering, especially near the top, is where the model adds the most value.
Pipeline
1Level the leagues
- • Refresh Layne Vashro's league‑strength scores with 2013‑24 data.
- • Convert every team (NCAA, Liga ACB, G‑League, NBL and others) to a single NCAA‑equivalent strength‑of‑schedule (SOS).
- • Multiply each counting and rate stat by mean_SOS / team_SOS, then z‑standardize so production no longer rises or falls purely because of league quality.
2Add development‑aware features
- • Inputs: pace‑adjusted box stats, Synergy play‑type rates, anthropometrics, consensus board and a survival probability for still being in the NBA by year five.
- • Age handling: teenage output gets a big bonus, the penalty fades after age 21 and is almost gone by 24.
3Predict EPM, then rank
- • Six sub‑models (one for each league–position bucket). XGBoost is used when data are plentiful, Random Forest when they are sparse or noisy.
- • Each sub‑model trains on pairwise rank loss for the five‑year average EPM, learning who should finish above whom.
- • Outputs feed a LightGBM ranker that produces one score for the full class. K‑means on that score creates objective tiers.
4Performance check
- • Validation methodology: 10-fold cross-validation trained on 2020 and prior drafts, with hold-out testing on 2021-24 drafts to ensure temporal validity.
- • Single‑season EPM root‑mean‑square error is 1.47~2.13, solid for pre‑draft data.
- • Ordering is the standout: across drafts 2014‑24, PASCAL's rank has a 0.51 correlation with actual year‑five rank, compared to 0.12 for the real draft order.
- • Normalized Discounted Cumulative Gain (NDCG) echoes the improvement, rewarding PASCAL for identifying the top prospects far more often than the draft itself.
5Outputs
- • Numeric class rank for every prospect plus a tier label that groups similar upside.
- • A projected EPM curve for seasons 1‑5.
Acknowledgments
Model design drew inspiration from Max Savin, Nick Kalinowski and Kevin Pelton.