A simulation study of tanking in the NBA · Algorithmic Game Theory

Why is losing
the right move?

Every year, NBA teams deliberately lose games. The draft rewards losing. We ran the math and simulated an NBA designed to stop it.

30 Teams simulated

50 Seasons per run

5 Mechanisms tested

3 Agent types

Scroll to learn how we stop tanking

About this study

Testing the debate, not just joining it

The NBA is adopting a new draft lottery mechanism this season, the 3-2-1 proposal, meant to reduce tanking incentives. Academics have proposed their own alternatives for decades. The league has tried reforms before. Everyone has a theory about what works. Almost no one has tested them against each other.

We ran 50 consecutive simulated NBA seasons under each of five mechanisms simultaneously, measuring how often strategic teams chose to lose deliberately and whether the draft was actually delivering picks to the teams that needed them most. Then we replaced every front office with an AI agent to see if the same incentive structures held under qualitative reasoning.

For the first time, the leading academic proposals and the NBA's own reforms are tested head to head in the same simulation framework.

What follows

1 How the draft works The lottery turns losing games into a strategic choice 2 The value of a great pick NBA 2K data puts a number on why pick 1 changes a franchise 3 The crossover point The exact record at which losing becomes more valuable than winning 4 A season, decision by decision One simulated team. 82 decisions. The exact moment the math flips. 5 Five proposed fixes From COLA to 3-2-1, each mechanism targets the same problem differently 6 What we found Tanking rates, draft fairness, and AI front offices, all five compared

Step 1

First, how does the NBA draft work?

The system

It's all about the… ping pong balls?

14 ping pong balls are numbered from 1–14. Before the draft, the league pulls 4 of these balls at random and the team with the corresponding ticket gets the pick.

One combination

This is one of 1,001 possible combinations

Draw 4 balls from 14 and you get a unique ordered set. There are exactly 1,001 ways to do it. The league provides all 16 non-playoff teams with a set of these combinations. The worst teams hold hundreds but better teams hold just a handful.

The full picture

More combinations, better odds. Simple enough.

That's the whole system. The worse your record, the more lottery combinations you hold. Each square here represents 5 combinations. The worst three teams, shown in red, hold more than 40% of all combinations combined. Once teams figured that out, some started losing on purpose.

The breakdown

The worst team sits at the top. The 17th sits at the bottom.

Combinations rearranged by NBA rank. The three worst teams hold the most, so their stacks are widest. From there, each slot gets fewer, until the 17th-ranked team barely registers.

Step 2

Why would a team want to lose?

One thing drives tanking: the value of a great pick. NBA 2K translates real player performance into numerical ratings. We use those ratings as a proxy for skill. The highest-rated players are the kind of franchise-altering talent that almost always goes first overall.

#1 PICK

Click to flip ↻

peak NBA 2K rating

The data

LeBron James. Pick #1, 2003. Peak NBA 2K rating: 99.

That bar is his peak NBA 2K overall rating. Use the buttons above to toggle between peak rating, first-season rating, and career average. He wasn't the best player in the league from day one, but he had the highest potential of his draft class. And he delivered.

The full picture

Now here's every lottery pick, averaged over 25 draft classes.

The bars show average NBA 2K ratings by pick position, from 2000 to 2024. Pick #1 averages 88 at peak: reliably elite, not always LeBron. The trend isn't perfectly smooth. Pick #3 historically outperforms pick #2. But the direction is clear: the top few picks offer a far higher chance of landing a franchise star.

The introduction of the lottery

Hakeem Olajuwon. Two championships.

Hakeem was the first pick in 1984. Houston won back-to-back titles with him. He predates NBA 2K but he's the reason the lottery exists. In 1984, the worst team got the best pick automatically. Houston, after losing to the Lakers, saw the incentive and tanked the rest of the season, knowing Hakeem would be available. The league instituted a lottery so no team could guarantee the top pick.

The lottery at work

Victor Wembanyama. Another generational talent.

The entire league knew Wembanyama would transform a franchise. Several teams positioned themselves accordingly. San Antonio finished with the second-worst record and won the top pick at 14% odds. Many others made questionable roster decisions in an obvious tank effort. They ended up with nothing to show for it.

Outliers exist

Steph Curry far exceeded expectations at the 7th pick.

The green bar is pick #7. That's where Curry landed in 2009. His first-season rating was 77, mid-lottery standard. His peak hit 98, one of the highest ever. Every team that passed on him is still thinking about it. The chart shows averages. Averages can't see Curry coming.

Mid-lottery

By pick 10, you're mostly getting role players.

Andrew Bynum. Richard Jefferson. Solid careers. Pick #10 and #13 average well below the top of the lottery. These picks matter, but our simulation models the expected value, the average outcome over many draft classes, not the lucky exception.

Our model

This is what our simulation uses.

The orange line represents expected pick value in our simulation. It measures average player contribution per draft position in a typical year. Not NBA 2K ratings. Actual team impact. Pick #1 is assigned a baseline value of 100. A Curry or LeBron would warrant far more. Every simulated team weighs this when deciding whether to compete or tank.

Step 3

The crossover point

Making the playoffs pays. Two to four extra home games with sold-out arenas. National TV appearances that drive sponsorship deals. A merchandise surge that lasts for years. Economists estimate the total franchise value boost at around $200 million per appearance. In our model, that's worth 200 points.

Now compare that to the lottery. The average lottery pick is worth about 24 points in player value. Even the number one pick, at 100 points, is only worth that if you win the lottery outright. So why would anyone tank? Because at some point, your playoff odds get low enough that competing isn't worth it anymore. Drag the slider to find that crossover.

Your playoff probability 22%

No chance (0%) 50/50

What's the top pick worth? Normal draft year (pick #1 worth 100 pts)

Normal draft year (100 pts) Generational talent (200 pts)

Value of Competing

points

Value of Tanking

points

How to read this: competing has value because you might make the playoffs (200 pts). But you still get a lottery pick even if you miss. Tanking accepts a lower playoff chance in exchange for better lottery odds. The crossover happens when those two paths are worth the same.

Try sliding "top pick worth" to 200 (generational talent year). Watch how the math changes.

Compete. Playoff value beats the lottery at these odds.

Step 4

A season, decision by decision

Here's a real simulated team navigating one full 82-game season. Each decision point is where the team ran the math and chose whether to compete.

The full season

Rank over time, game by game.

The line traces this team's position in the league standings as the season unfolds. The dashed line at rank 16 is the playoff cutoff. Above it, you're in. Below it, you're in the lottery.

Quarter of the season

On the bubble. What's the call?

About 20 games in. The team is still within reach of the playoffs. Here's what the simulation decided:

Halfway through

The gap is growing.

At the midpoint of the season, the playoff picture is coming into focus. The team recalculates. Here's the decision:

Three-quarters in

The math has spoken.

With about 20 games left, there's little ambiguity. The calculation is straightforward. Here's what happens:

Every decision

Green means compete. Red means tank.

Here are all the decisions across the full season, plotted on the rank line. Green dots late in the season appear when the team recalculates and finds its lottery position is effectively set. No more ground to gain or lose. A red dot that follows means the next calculation found a small window to slip one more rank. They didn't see a way to improve their lottery position at that moment. Then the math changed again. Hover any dot for the full reasoning.

Step 5

Five attempts to fix tanking

The NBA and researchers have proposed several mechanisms to reduce tanking. Each takes a different approach. Click to explore how each one works.

NBA Lottery (Current Rules, 2019)

The three teams with the worst records each get a 14% chance at the number one pick, a flat zone designed to stop one team from hoarding the best odds. Teams ranked 4th through 14th worst get declining odds, from 12.5% down to 0.5%. Only the top four picks are lottery-determined. Picks 5 through 14 go by record.

This still incentivizes tanking. A worse record always means more lottery balls.

The baseline. Better than before 2019, but tanking still pays off.

Bilevel Ranking

Draft positions are locked in at game 70 of the 82-game season. Whatever your record is when standings freeze determines your lottery slot. The final 12 games can only affect playoff seeding, not draft position. After the lock, losing has exactly zero lottery benefit.

Why it cuts tanking: once standings lock, there's nothing to gain from losing. The last third of the season becomes honest competition.

Eliminates late-season tanking. Some early-season incentive still remains.

COLA: Carry-Over Lottery Allocation

Every non-playoff team earns 3 tickets per missed season. Record does not matter. Follow one team, two universes, three seasons.

Tanks

Plays hard

1 of 4

For reference: cost of winning a top pick

Landing a top pick resets your ticket count. The better the pick, the steeper the cost.

#1Full resetback to 0

#2−75%keep a quarter

#3−50%keep half

#4−25%keep three-quarters

#5–14No costkeep everything

Fall to pick five and tanking bought you nothing anyway. Land the best player and you start over. You earned that pick the same way whether you won 20 games or 40.

Weighted Loss Mechanism

Lottery position is based on a cumulative loss score where early-season losses count far more than late ones. The weight decays exponentially. A loss in game one is worth twice as much as a loss in game 20, and four times as much as a loss in game 40. By midseason, tanking buys almost no lottery benefit.

Why it reduces tanking: the payoff from losing shrinks continuously. There's no single cutoff to game the system around.

Continuous decay. No hard cutoff to time tanking around.

NBA 3-2-1 Proposal

The NBA is introducing this as a three-year pilot. The incentive is deliberately inverted. Teams ranked 4th through 10th worst get 8.1% odds at the first pick, more than anyone. The three worst teams get only 5.4%. Finishing last is now actively bad for lottery odds. The proposal expands the lottery to 16 teams, adding Play-In losers as eligible teams at 5.4% and 2.7%. Our simulation includes all 16 teams.

Finishing last is bad for your odds. The worst three teams hold about 5.4% each. The middle group holds 8.1%. Teams compete to stay out of the bottom, not reach it.

The risk stays at the boundary. A bubble team in a strong draft year might still decide a missed playoff is worth more than a first-round exit.

Lottery odds at number one pick, by draft rank

Season timeline: standings lock at game 70

Ticket accumulation over 5 missed seasons

Loss value by game number (half-life = 20 games)

Odds at pick 1, by team rank

current rules, 2026

the worst teams just lost their advantage

The Simulation

How much does each reform actually cut tanking?

We ran 50 simulated NBA seasons under each mechanism with strategic teams that calculate their odds every game. Scroll to reveal each mechanism.

Mechanism 1 of 5

NBA Lottery (Current)

6.9%

The current system is the baseline. Strategic teams tank 6.9% of game-day decisions. Every time they can win, they calculate whether it's worth it. That's the incentive the lottery creates.

Mechanism 2 of 5

Bilevel: Lock at Game 70

3.0%

Cutting the season in half for lottery purposes nearly halves tanking. The remaining 3.0% happens before the lock point. Teams still tank early to secure a better lottery slot before standings freeze.

Mechanism 3 of 5

Weighted Loss

2.3%

The exponential decay squeezes tanking windows throughout the year. Strategic teams still find pockets to optimize, but the payoff gets smaller the longer they wait.

Mechanism 4 of 5

NBA 3-2-1

0.9%

When being in the bottom three gives you fewer lottery balls, teams near the absolute bottom compete harder to avoid that zone. Only 0.9% of decisions are tanking. Near the noise floor.

Mechanism 5 of 5

COLA

0.5%

When this season's record has zero effect on lottery odds, rational teams essentially stop tanking. The 0.5% comes from early-season decisions before teams have enough data to rule out a playoff run.

When During the Season Do Teams Tank?

Each mechanism creates a different incentive window. Watch how tanking pressure builds, peaks, and cuts off under each set of rules.

Beyond Tanking

Does less tanking mean fairer drafts?

Suppressing tanking only matters if the mechanisms also help the teams that genuinely need good picks. We track two things for every mechanism.

How well do standings reflect true team quality?

We compare each season's final standings against each team's true underlying skill. Zero means the best teams won the most. A higher score means more disorder between talent and results. All five mechanisms land in a tight band. The reform design barely affects competitive balance at these tanking rates.

Average draft pick for the 8 weakest teams

Lower is better: weak teams get earlier picks.

The weakest quarter

3-2-1 gives the weakest teams the best picks.

Looking at the 8 teams with the lowest true skill each season, 3-2-1 delivers the lowest average pick number. Under 3-2-1, all 14 picks go through the lottery. Even middling weak teams get a real shot at any slot, rather than being locked into picks 5 through 8 by record.

Now zoom in on the bottom 3

COLA is hardest on newly terrible teams.

COLA rewards history, not this year's record. A team that was great for years and just fell apart this season hasn't accumulated many tickets yet. It could be the worst team in the league and still sit at the back of the lottery queue, behind franchises that have been bad longer. COLA is slow to recognize newly terrible teams.

Which picks do the weakest teams actually receive?

Since our simulation assigns each team a true skill level, we know which three teams were genuinely the weakest each season, not just which three lost the most. Among those teams, what pick slots did they end up with? Darker cells mean that range is more common. The row labels tell you whether dark is good or bad. A well-designed mechanism has its darkest cells at the top.

The AI Experiment

What if AI ran every front office?

We replaced every team's front office with a Claude Haiku language model, the same AI that powers many commercial chatbots. Each AI received real standings, game counts, and the rules of the mechanism, then had to decide: compete or tank? No hints. No training on the right answer.

NBA Lottery

8.0% vs 6.9% for strategic teams

A real front office doesn't run expected-value calculations on every game. A GM reads the standings, weighs fan expectations, thinks about whether a season still feels salvageable. The AI does the same. It applies a qualitative "we're out of it" threshold and drops effort sharply. That's closer to how front offices behave than a strict probability calculation. It's also why the AI tanks slightly more than the math demands. Loss aversion and narrative framing push GMs to give up earlier than optimal.

Bilevel

6.2% vs 3.0% strategic

The AI grasps the deadline mechanic: tank before Game 70, or lose the window entirely. It responds to bounded deadlines more aggressively than a pure optimizer would. Real GMs show the same behavior. Thaler calls it present bias. The value of an action feels larger when time is running out, even when the math doesn't change.

Weighted Loss

3.4% vs 2.3% strategic

The AI picks up the decay mechanic and pulls back on late-season tanking. But it still overshoots. Qualitative factors push it further than the formula says: the appeal of attaching your name to a high pick, the fear of disappointing fans who expected a tank-for-the-top narrative, the sheer momentum of a committed strategy.

NBA 3-2-1

~0% matches strategic

The AI correctly identifies that landing in the bottom three is actively bad for lottery odds and competes accordingly. The inversion is intuitive enough that qualitative reasoning handles it.

COLA

0.0% matches strategic

When your record this year has zero effect on lottery position, there is nothing to rationalize. The AI reads the rules and doesn't tank. Not because it calculated expected values. Because it understands the situation doesn't reward it. This matters: even behaviorally messy decision-makers stop a behavior when the incentive is fully removed.

Policy Recommendation

What should the NBA do?

No mechanism eliminates tanking entirely. Any lottery that links record to picks will always have some team somewhere with a reason to lose. The goal is to make tanking costly enough that it only makes sense in extreme situations.

What we support

3-2-1 cuts tanking from 6.9% to 0.9% in our simulations. All 14 picks go through the lottery, so even mediocre non-playoff teams get a real shot at any slot. Teams near the bottom three threshold have genuine reason to compete instead of lose. And it requires no new infrastructure. Just a rebalancing of balls.

0.9% tanking rate

What we worry about

3-2-1 closes the door on tanking for teams already out of playoff contention. But it leaves another door open. In a year with a generational prospect, a bubble team might decide the pick is worth more than a first-round exit. Miss the playoffs on purpose, collect your three lottery balls, take your shot. The system cannot stop that calculation. And if it happens, the cruel irony is that the worst teams in the league get fewer balls than the teams that chose to lose. The prospect most needed at the bottom goes somewhere else.

One more thing: the NBA is treating 3-2-1 as a 3-year experiment. If it doesn't reduce tanking in practice, they'll revisit.

Our results assume a playoff berth is worth 200 simulation points. Raise that threshold, or introduce a generational-class prospect, and the calculus shifts. The mechanism's robustness to extreme draft years remains an open question.

Tier A 2 balls vs Tier B 3 balls: the inversion

If 3-2-1 proves effective, two options are worth building on. A Bilevel standings lock at game 70 would cut tanking further with no new infrastructure. Or the league could move toward COLA, where lottery position is based on years out of the playoffs rather than current record. That approach carries the strongest theoretical guarantee: tanking is never the rational move.

Teams are smarter, they are creative, and they respond. We move, they move. So we're always looking to see whether there's yet a better system.

Adam Silver, NBA Commissioner

Grant Valentine · Topics in Algorithmic Game Theory · May 2026
Simulation: 30 teams · 50 seasons · 3 agent types (Strategic, Honest, Claude Haiku) · 5 mechanisms
Built with D3.js and Scrollama

Go Deeper

Want to read more?

This essay is the readable version of our formal research paper. If you want the proofs, the model details, or just to check our work, here's everything we read to build it.

Paper

Valentine & Dhinakaran (2026)

Our full paper. The math behind all of this, with proofs and complete simulation results.

Full paper ↗

Assumption used throughout

A note on simulation assumptions

All tanking rates in this essay assume a playoff berth is worth 200 simulation points, calibrated to reflect typical franchise value and competitive incentives. This is consistent with Kazachkov 2020. In a year with a generational prospect, the value of the top pick can exceed this threshold and the tanking incentive returns even under 3-2-1. The mechanism's robustness to extreme draft years remains an open question.

Code

AGT Tanking Simulations (GitHub)

All simulation code, data, and this visual essay. Open source.

GitHub repository ↗ Proposal

Banchio & Munro (2020). "A No-Tanking Draft Allocation Policy"

Proposes the COLA mechanism. Proves that carry-over tickets effectively remove the tanking incentive when record has no effect on lottery position.

MIT Sloan paper ↗ Proposal

Kazachkov & Vardi (2020). "On Tanking and Competitive Balance"

The original proposal for locking draft position mid-season. The theoretical foundation for the Bilevel mechanism we test here.

PDF ↗ Proposal

Highley et al. (2026). "Carry-Over Lottery Allocation"

Full theoretical treatment of the COLA mechanism, with proofs of incentive compatibility and simulation results across league configurations.