What Works on Wall Street — O'Shaughnessy: COMPLETE

The setup is pure mechanics. Every year on 31 December the entire universe is divided into ten equally large groups — deciles — by a factor, for example P/S. Decile 1 contains the ten percent of stocks with the lowest value on the factor, decile 10 the ten percent with the highest. You buy the decile portfolio, hold for exactly one year and repeat the procedure.

No opinions, no adjustments along the way, no knowledge of the future. Then you compare the deciles' aggregate results with the universe as a whole over decades. This is how you test a factor — and it is exactly the way AK1A works when a variable is to earn its place in the system.

Two universes underlie almost all tests. All Stocks is all companies of sufficient size — thousands of stocks in modern editions. Large Stocks is the larger and more traded companies. During the main period 1951–1996, All Stocks returned about 13,6 percent per year while Large Stocks returned about 11,9 percent — the small-cap premium is thus already in the base data, something chapter 10 returns to.

The fourth edition extends the tests to 2009 and in several chapters back to 1927 — and the main conclusions hold essentially unchanged through the IT bubble and the financial crisis.

The book's most important methodological choice concerns survivorship bias. An ordinary stock database contains only companies that still exist — those that went into bankruptcy or were delisted have disappeared. If you test factors on such a sample, the result is systematically overstated, because all the dead companies are missing from the statistics. O'Shaughnessy uses the database's research file, which includes historical, now deceased companies, precisely so that the result will mirror reality: whoever bought decile 1 in 1955 could not know which companies would survive to 1996.

In addition, he looks for MONOTONIC relationships — that the result improves or worsens evenly across the deciles — because a gradual pattern over ten steps is hard to explain with luck, while a single top reading is easily data mining.

The result in plain text. During the main period 1951–1996, decile 1 — the ten percent of stocks with the lowest P/S — returned about 16 percent per year. Ten thousand dollars invested in that decile in 1951 became about 8,2 million dollars at the end date, against about 3,1 million for the same stake in All Stocks as a whole (≈13,6 percent per year). And at the other end of the scale? Decile 10 — the most expensive stocks measured by P/S, that is the glamour stocks — returned barely five percent per year.

That is worse than government bonds over the same period. O'Shaughnessy's conclusion is undramatic but total: the relationship between share price and sales is the strongest single factor in the entire study.

Why precisely P/S? Because almost everything else can be tinkered with. Net income — the basis of P/E — is affected by depreciation rules, impairments, tax items and one-off items; managements have great freedom to massage the result. Sales are harder to fake over time: either the revenue is in the invoices or it is not. P/S also measures the size of the whole business — what you pay per unit of currency the till passes through — not only the sliver left after all costs.

Connect to Graham's legacy: margin of safety without numbers is precisely this — paying little for much fundamental per unit of currency. Low P/S is the most reliable price form of that margin of safety.

The behavioral economics explanation is overreaction. The market loves stories and extrapolates them: hyped companies can pay any valuation multiple (decile 10), while boring, out-of-fashion or temporarily disfavored companies are thrown in the wastepaper basket (decile 1). The research that losers beat winners over time — DeBondt and Thaler's overreaction hypothesis — is the theory behind the data: because people systematically exaggerate glamour and misery, low P/S buys mispriced pessimism and does not sell it until it has been corrected.

Boredom pays. In AK1A this means that V04 should carry the most weight in the entire valuation dimension — but the book's own warning applies: alone it is strong but not unbreakable. In 1996–1999 low P/S underperformed three years in a row while IT glamour rose and rose. The solution comes in chapter 7: combine.