What we tested, and what actually worked

Every valuation measure we could test, inside every kind of business, over 213282 PIT company-quarters 2006-2026, delisted included. A measure only counts when the cheapest fifth of companies beat the dearest by a margin that held in every period we tested and that survives resampling. Most do not. The failures are listed below the successes, deliberately: a list of only the winners would tell you nothing about how hard they were to find.

Study AV-01 · generated 2026-09-20 · benchmark: S&P 500 total return over the same window · 206 combinations tested, 2 beat the index, 48 were real but still lost to it.

What actually beat the market

The hardest bar, and the only one that would justify buying something: the cheap group had to beat the S&P 500, not merely fall less than the expensive group.

Real, but not a reason to buy: "lost less"

These 48 findings passed every statistical test — the gap between cheap and expensive is real and held across periods. But both groups lost to the index; the cheap ones simply lost less. Among small and unprofitable companies that is what "value" mostly means, and it is not a case for owning them. We list them because the gap is genuine and because leaving them out would make the section above look easier to find than it was.

What did not work at all

These were tested on exactly the same data and are shown so you can see the denominator. A measure fails either because it worked in one period and reversed in another, or because the result cannot be separated from chance.

How this was measured

• Every figure is taken as it stood on the day the filing reached the SEC — never a restated number, and never one published after the date it is used on.
• The price history includes companies that were later delisted, carrying whatever actually happened to them. A study run only on today's survivors would flatter every result on this page.
• Companies are ranked against others of the same business type in the same quarter, so a cheap market cannot be mistaken for a cheap company.
• The periods are tested separately. A measure that works in one and reverses in another is marked as such, not averaged into looking useful.
• Confidence comes from resampling whole quarters, which keeps months that moved together from counting as independent evidence.

Testing many combinations at once means a few will look good by luck. That is why agreement across periods is required as well as statistical significance, and why the count of everything tested is printed above rather than only the winners.