Optimizing NBA Roster Moves for Postseason Value

The question

Which realistic, same-position roster swaps create the greatest expected postseason value after accounting for probability gain and salary cost?

The money makes it worth answering: postseason home games can generate roughly $1–2.5M each, playoff relevance lifts merchandise demand, and the 2025 postseason averaged 6.13M viewers per game (75M for the U.S. Finals). The project used a base-case postseason value of $10M per team-season.

Core assumptions

• Swaps happen without contract, cap, or trade-rule constraints, and are one-for-one, same-position
• The replacement player takes over the incumbent's minute role
• Team postseason probability can be estimated from historical team-season features
• Postseason probability converts to commercial value via a fixed value estimate
• Observed player-season performance is a reasonable proxy for swap impact, without explicitly modeling chemistry or coaching

Approach

Modeling. A logistic regression estimates team postseason probability with postseason success as a binary target, a natural fit that keeps coefficients interpretable. The final model draws on payroll, roster structure, efficiency, defense, and positional features; minutes-weighted age, assist/turnover ratio, and total salary were the strongest predictors. Generalization was tested with leave-one-season-out cross-validation, and accuracy stayed stable across seasons.

Prescriptive swap engine. On top of the model sits a five-step tool that turns a prediction into a decision:

• Select team context (team, season, position)
• Generate one-for-one, same-position swap candidates
• Simulate the role-based replacement (insert the candidate into the incumbent's minutes)
• Recompute the team outlook and new postseason probability
• Convert the probability change to expected value, compare against salary, and rank the swaps

Results

The model generalized well enough to drive the prescriptive tool: about 76.0% leave-one-season-out accuracy and an F1 of roughly 0.772, with raw probabilities outperforming post-hoc calibration methods on Brier score and log loss. The end result identifies realistic same-position upgrades, quantifies the competitive-versus-financial tradeoff, and flags vulnerable roster spots, and the project placed 3rd in a project showcase.