Which NBA draft-lottery design is worth considering? We tested 48 variants in a basketball simulator. This page shows which ones came out on top.
Every NBA season, teams that fall out of playoff contention have a documented incentive to lose games on purpose. Worse end-of-season records improve a team's odds in the draft lottery. Munro & Banchio (2020) proved that no draft mechanism based only on end-of-season records can both reward poor performance and eliminate this incentive.
Carry-Over Lottery Allocation (COLA), proposed by Highley, Duncan, and Volkov (2026), sidesteps that impossibility by basing draft priority on a team's multi-year playoff history instead of single-season records. COLA is not one design but a family of them: Classic, Simple, Capped, Countdown, plus the NBA's 2026 “3-2-1” reform proposal, which sits nearby in the same design space.
We tested 48 variants from this design space against the ZenGM Basketball-GM simulator (per Prof. Highley's recommendation) and scored each variant on three things a league office cares about:
With three objectives in play, no single variant is “best” on all of them at once. The Pareto frontier identifies the variants that are not dominated by any alternative on all three objectives simultaneously. Variants off the frontier are strictly worse than something else; variants on the frontier are the ones a league office could defensibly choose between.
To make “dominated” concrete: Status quo has 28.5 years between conference finals, 28.6% manipulation gain, and 2.77 rank spread. Capped@150 has 28.0, 1.4%, and 2.87. Capped@150 wins on every dimension, so Status quo is strictly worse and not on the frontier. The same dominance check rules out Classic, Simple, and the NBA's 3-2-1 proposal.
Five of the 48 variants are Pareto-optimal. Only one is a named COLA variant: Capped COLA with cap value 150 (Capped@150).
Each dot below is one variant. The X-axis is manipulation gain (lower is better); the Y-axis is the longest gap between conference-finals appearances (lower is better). Pareto-optimal variants are highlighted in color; named variants (Classic, Simple, Status quo, 3-2-1, Capped@150) are shown as diamonds. Hover over any dot for full details.
Nine variants in detail: the five Pareto-optimal configurations plus the four named variants that are dominated. Capped@150 appears once (it is both Pareto-optimal and named).
| Variant # | Named variant | Eligibility size | Cap | Carry-over scope | Years between conf. finals (median ± std) | Manipulation gain (%) | Rank 1–5 spread (median) |
|---|
The paper proves (Corollary 2.1) that for capped COLA variants, the manipulation-gain bound depends only on the eligibility pool size, not on the cap value. The empirical sweep confirms it: at any given eligibility size, every cap value produces the same manipulation gain.
| Eligibility size | Cap = 100 | Cap = 150 | Cap = 200 | Predicted (0.3 / eligibility size) |
|---|---|---|---|---|
| 14 | 2.1429% | 2.1429% | 2.1429% | 2.143% |
| 22 | 1.3636% | 1.3636% | 1.3636% | 1.364% |
| 16-tiered | 1.8750% | 1.8750% | 1.8750% | 1.875% |
The 48 variants come from sweeping three of COLA's seven design “dials”: eligibility size (14, 22, or a 16-team tiered structure), stockpile cap (uncapped, 100, 150, or 200), and carry-over scope (single-season, unbounded, bounded over 30 years, or reset on championship). The other four dials are held at Classic COLA defaults.
For each variant, we ran 50 simulated 30-season runs (headline run computed 2026-05-26). The lottery itself uses the real ZenGM COLA code unchanged. Regular-season outcomes are synthesized (strength-weighted wins, single-elimination playoff bracket) because the full ZenGM game engine does not run cleanly outside a browser. Team strength persists across seasons via a small Markov model; parameters are calibrated for plausibility, not fit to historical NBA data (see paper §5.1 for the full specification).
Sensitivity passes at 30 and 100 replicates show no greater than 10% mean drift in the primary objective, so 50 replicates are statistically sufficient for the dominance ordering reported here.