UPDATE: all tools, charts, tables, and the paper itself
have been updated and now include the 2015 draft
The Automated General Manager
An Unbiased, Backtested Algorithmic System for Drafts, Trades, and Free Agency that Outperforms Human Front Offices
by Philip Z. Maymin, email@example.com, philipmaymin.com
I introduce an automated system and interactive tools for NBA teams to better decide who to draft, who to trade for, and who to sign as free agents. This automated general manager can serve either as an expert-system replacement or a complement to a team's front office, and also as a calibrating benchmark to compare against actual team building performance. Backtested over the past ten years, the automated GM outperforms every single team, and by substantial margins that often represent a major portion of the team's market value. From draft decisions alone, the average team lost about $130,000,000 worth of on-court productivity relative to what they could have had with the automated GM; this shortfall represents about a quarter of the average franchise value. Historically the automated GM's choices would have produced about twice as much as the human choices actually did: approximately one extra win per year per draft pick. The system is calibrated using an innovative extension of traditional machine learning methods, applied to a uniquely broad historical database that incorporates both quantitative and qualitative evaluations, in a way that avoids possible survivorship bias, and for a variety of performance metrics; it is thus robust, comprehensive, realistic, and does not overfit information from the future. I provide virtually all of the interactive tools supporting this paper, including backtesting results, projections, scenario analysis, and more, online, for free, at nbagm.pm.
Available for download in PDF format from SSRN.com.
The remainder of this website is the technical appendix, including virtually all of the interactive tools, reports, and systems.
Reports marked with this icon are in Computable Document Format (CDF). To view and interact with them, download the free CDF player.
- Prospects are evaluated according to a machine learning model trained on historical data.
- Inputs include college stats, combine info, and historical mock drafts, among other things. (See the CDF's for more details.)
- Subsequent NBA performance is measured by WM: Wins Made, Wins Mean, Wins Maymin, etc.
- WM is a combination of three different metrics: Win Shares, Wins Produced, and PER→EWA
- Those three component metrics each estimate how many wins a player contributed to his team that year.
- A past prospect's subsequent NBA success is measured as his average WM over his first three seasons.
- The model backtest results without hindsight or data snooping bias are available below.
- Historical Big Board (31mb)
- Top prospects for past drafts, including both projections and actual production.
- Also includes 2015 draft. See and edit the college stats, combine info, and other data to update projections live.
- Includes both tables of rankings and a graph of projections vs. actual production.
- Historical Draft Pick Performance (35mb)
- How many minutes does the average #31 draft pick play? How does it change as a function of the draft pick?
- Historically, who generates more wins, the #5 overall pick, or the #31, #36, and #48 combined?
Current NBA Players
- Players are evaluated according to a machine learning model trained on historical data.
- Inputs include most recent to date, and best to date, NBA seasons. (See the CDF's for more details.)
- Subsequent NBA performance is measured by the same WM as above.
- Current Player Production and Value (7mb)
- Explore all current NBA players' production, salary, and net value.
- Filter by team, position, career or just 2014 or several years back, average or maximum performance.
- Projections (9mb)
- Project and sort current NBA players out 1-4 years on GP, Min, and WM.
- Historical Projections (5mb)
- Chart of how historical projections actually worked out.
- Stability (5mb)
- Given only a player's stat in one year, how well does that predict that same stat for him the next year?
- Trade Evaluation (9mb)
- Evaluate the basketball impact of trades of players and picks.
- Open multiple copies to evaluate various alternatives side-by-side.
- Note: unfortunately this tool does not work with the free CDF player; the CDF player pro is required.
- Draft Performance Chart (3mb)
- Compare the model choices vs. actual human choices.
- Team-by-Team Table (36mb)
- How did each team do historically compared with the automated GM?
- What portion of the team's franchise value did the losses represent?
- Also optionally show who each team should have taken instead each year.
- Should Have Taken... (36mb)
- What players should each team have taken historically?
- Re-evaluate for different metrics historically.
- Pick prospects based on consensus mock drafts only? Projected total wins? Projected wins per 48 minutes? etc.
- East: ATL BOS BRK CHA CHI CLE DET IND MIA MIL NYK ORL PHL TOR WAS
- West: DAL DEN GOS HOU LAC LAL MEM MIN NOP OKC PHX POR SAC SAN UTH
- Build a Team Through the Draft Only (37mb)
- Suppose you start with zero players and can build only through the draft.
- How would you have done, based on your draft performance metric?
- Or backtest a hypothetical team that every year has picks #15 and #45 (or any other combination of picks).
Last updated: 07/02/15 7:30pm EDT
All content copyright 2014-2015 Philip Z. Maymin