Dynasty Model - Results

NBA dynasty fantasy basketball projection model. Walk-forward cross-validation across 10 seasons (2017-2026).

Architecture

• Stage 0: Feature imputation (tracking stats for pre-2014 seasons)
• Stage 1: LightGBM quantile regressors (14 stats x 5 horizons x 3 quantiles = 210 models) + Transformer trajectory embeddings + learned ensembler
• Stage 2: VORP-based dynasty valuation with positional replacement levels

Conformal calibration via MAPIE on held-out calibration sets.

T+1 Results (Next Season)

Stat	RMSE	MAE	Floor Cal	Ceil Cal
PTS	3.276	2.534	79.2%	69.6%
REB	1.329	0.980	78.5%	77.8%
AST	0.873	0.615	81.7%	79.1%
STL	0.307	0.228	81.2%	78.9%
BLK	0.249	0.169	80.3%	83.7%
TOV	0.451	0.341	80.8%	80.2%
FGM	1.153	0.898	77.5%	74.5%
FGA	2.418	1.867	73.1%	74.5%
FTM	0.718	0.517	83.7%	79.0%
FTA	0.872	0.632	82.3%	79.7%
3PM	0.477	0.348	85.7%	86.2%
3PA	1.173	0.873	79.2%	81.6%
MP	6.058	4.659	87.5%	80.8%
G	23.675	18.612	87.8%	90.8%

T+2 Results (2 Seasons Out)

Stat	RMSE	MAE	Floor Cal	Ceil Cal
PTS	5.426	4.533	89.0%	90.2%
REB	2.149	1.715	88.6%	89.9%
AST	1.456	1.159	89.2%	90.0%
MP	10.321	8.303	89.1%	90.0%

T+3 Results (3 Seasons Out)

Stat	RMSE	MAE	Floor Cal	Ceil Cal
PTS	6.591	5.621	87.0%	91.2%
REB	2.329	1.908	86.9%	91.4%
AST	1.557	1.258	90.7%	88.0%
MP	11.372	9.276	87.3%	91.4%

vs Baseline (T+1 Only, LightGBM-only, Feb 27 2025)

Stat	Baseline MAE	Current MAE	Delta
PTS	2.246	2.534	+0.288
REB	0.894	0.980	+0.086
AST	0.577	0.615	+0.038
MP	4.039	4.659	+0.620

Per-Fold PTS T+1

Year	RMSE	MAE	Floor Cal	Ceil Cal
2017	3.260	2.527	83.4%	65.2%
2018	3.383	2.558	79.4%	74.1%
2019	3.363	2.608	76.6%	67.8%
2020	3.451	2.657	84.7%	71.6%
2021	3.089	2.382	68.9%	75.6%
2022	3.392	2.654	68.1%	75.7%
2023	3.206	2.520	84.2%	69.8%
2024	2.961	2.359	81.5%	68.1%
2025	3.549	2.713	81.2%	67.9%
2026	3.103	2.361	83.7%	59.7%

Notes

• Calibration target: Floor and ceiling should both be ~90% for proper 80% prediction intervals
• T+1 calibration is too tight: Floor avg 81.3%, ceiling avg 79.7% — model is overconfident at 1-year horizon
• T+2/T+3 calibration is good: Both ~89-91%, right on target
• Games (G) is the weakest stat: MAE ~19 games due to inherent injury unpredictability
• Baseline comparison: Current model trades slightly worse point-prediction accuracy for multi-horizon capability (14 stats x 5 horizons vs baseline's 4 stats x 1 horizon)
• Run date: March 1, 2026