The Superinvestors of Graham-and-Doddsville

Definition

Warren Buffett's appendix essay answers the efficient-market claim that market-beating investors are simply lucky. His counter is a test of origin, not of any single track record: if a coin-flipping contest among 225 million people leaves 215 winners after twenty straight correct calls, that is exactly what chance predicts. But if 40 of those 215 turn out to have come from the same small zoo, you no longer need to invoke luck — you go ask the zookeeper what he's feeding them. A disproportionate concentration of long-run winners tracing to one shared "intellectual origin" is the signature of a real, transmissible method.

In the Book

Buffett names a specific "intellectual village" — investors trained directly under Graham at Graham-Newman or through his Columbia classes — and profiles several by name and audited record: Walter Schloss, running a fund with no formal education beyond Graham's course, compounding well above the S&P 500 for decades by buying statistically cheap stocks with no inside information or industry specialization; Tom Knapp and Ed Anderson at Tweedy, Browne; Bill Ruane at the Sequoia Fund; Charlie Munger and Rick Guerin, each independently applying Graham's margin-of-safety framework to different portfolios of different companies, often with low overlap in actual holdings between them. Buffett stresses the guardrail that makes the test valid: these investors were identified as Graham disciples years before their results were tallied for this talk, ruling out the after-the-fact cherry-picking that would let anyone construct an impressive-looking cluster from a large enough population.

Why It Matters

This is a general method for telling skill from luck in any domain with a large population of independent actors: don't ask whether one result beat the odds, ask whether the winners cluster around a shared, pre-identified method or origin more than chance would predict. The pre-identification requirement is the load-bearing part — without it, the same clustering argument can be manufactured retroactively around almost any pattern, so the test only works if you can show the "village" was named before you knew who would win.