Factor investing and the modern cross-section

For most of this history the question was about the market as a whole: does it go up, can you time it, what does it pay. This chapter turns the question sideways. Instead of asking how the market moves through time, it asks why, at any given moment, some stocks pay more than others. Line up every stock in the market and sort them by some trait, cheap against expensive, small against large, rising against falling, and check whether the groups earn different returns. Researchers call the lineup the cross-section. The sorting sounds like a parlor game, and for two decades it produced the most useful findings in finance. Then it produced several hundred findings, most of them false, and the field had to learn, in public, how to tell the difference. That second part is the real subject of this chapter, because it is a problem you will meet everywhere data is cheap.

Fama and French, 1993

Eugene Fama and Kenneth French took the two most stubborn patterns in the data, small companies beating large ones and cheap companies beating expensive ones, and turned them into a machine. They built portfolios that captured each pattern in its purest form, long the small, short the big; long the cheap, short the expensive, and showed that these two, together with the market itself, explained most of the differences in return between any diversified portfolios you could construct. Three factors, where before there had been one.

The practical consequence took years to sink in and is now an industry. If most portfolios are just mixtures of a few common ingredients, then a manager’s record can be separated into the ingredients, which are cheap to buy, and whatever is left over, which is rare. Before this paper, a manager who held small cheap stocks in a decade when small cheap stocks did well looked like a genius. After it, he looked like a man who had bought the right ingredients, and the question became what you were paying him for. I ask that question of every fund I am shown, and it is remarkable how often the answer is an ingredient with a markup.

Jegadeesh and Titman, 1993

The same year, Narasimhan Jegadeesh and Sheridan Titman published the pattern nobody wanted. Stocks that had done well over the past six to twelve months kept doing well for a while; stocks that had done badly kept doing badly. Momentum. Value at least had a story the theorists could live with, that cheap stocks carry real risk and get paid for it. Momentum had no such story. If prices reflect what is known, yesterday’s return should say nothing about tomorrow’s, and here was yesterday’s return saying something, decade after decade, in the most heavily studied market on earth.

Thirty years later there is still no explanation everyone accepts, only candidates: people react to news slowly, then chase it late. What I take from momentum is humility about the word impossible. The most awkward fact in finance was sitting in plain sight, in the same data everyone used, until someone bothered to sort it. And a warning that goes with it: momentum earns its return by trading, the portfolio turns over constantly, and costs eat a large share of what the paper measures. A pattern can be real and still be unprofitable for you by the time you have paid to follow it.

Asness, Moskowitz and Pedersen, 2013

Twenty years on, Cliff Asness, Tobias Moskowitz and Lasse Pedersen asked the question that separates a fact from a fluke: does it travel? They took value and momentum out of American stocks and went looking for them everywhere else, stocks in Europe, Japan and the United Kingdom, government bonds, currencies, commodity futures. Both patterns showed up in every market they checked, moving together across markets, which is what you would expect if something real and global were behind them, and which is very hard to explain if the original finding had been luck in one dataset.

This is the test I now apply to any pattern anyone shows me. A discovery made in one market and one period is a coincidence until proven otherwise, because with enough sorting, one dataset will always confess to something. A discovery that repeats across markets that share nothing, Japanese stocks and copper futures, cannot be explained by anything particular to those markets, because they have nothing in common. The only thing they share is the people trading them. So when the same pattern appears in both, it has to come from how people behave: how they react to news, what they chase, what they give up on. That is why the pattern travels. Markets change; the people in them do not.

Frazzini and Pedersen, 2014

Andrea Frazzini and Lasse Pedersen then documented the strangest pattern of the four: the market pays you badly for taking more risk. In theory, return should rise in a straight line with risk. In the data, the line is nearly flat. The safest stocks earn almost as much as the riskiest ones while swinging far less, so per unit of risk, the calm end of the market is the good end.

Their explanation is about constraints. Most of the money in markets is barred from borrowing: pension funds, mutual funds, most individuals. When a constrained investor needs a high return, he cannot lever a safe portfolio, so he does the only thing available and buys the riskiest assets he is allowed to hold. The risky end gets crowded and overpriced; the safe end is left cheap for whoever can borrow prudently against it. It is a premium paid by the impatient and the constrained to the patient and the solvent, and it has been visible in every asset class they measured. When someone tells you that earning more always requires risking more, this paper is the polite way to say it is more complicated than that.

Fama and French, 2015

Twenty-two years after their three factors, Fama and French returned and revised themselves. Two more patterns had proven too strong to ignore: profitable companies beat unprofitable ones, and companies that invest conservatively beat companies that spend aggressively. So the three-factor model became five. And in the new model something awkward happened: their own value factor, the one that had launched the industry, became largely redundant once profitability and investment were in the room.

I find the episode more instructive than the model. The two most cited empiricists in finance looked at twenty more years of data and rewrote their signature work, in public, at the cost of their earlier conclusions. That is what the discipline looks like when it works. Hold every model the way they held theirs: firmly enough to use, loosely enough to revise. Twenty more years of data will change this one too.

Harvey, Liu and Zhu, 2016

Then came the reckoning. Campbell Harvey, Yan Liu and Heqing Zhu counted every factor the academic literature had discovered and published: more than three hundred. Three hundred separate traits, each claiming to predict returns, each with a backtest behind it. They then made a point that every investor should have tattooed somewhere. If thousands of researchers run thousands of tests on the same data, ordinary standards of proof collapse, because run enough tests and chance alone will hand you dozens of patterns that look highly significant and mean nothing. By their arithmetic, a large fraction of everything published, perhaps half, was likely false. The profession had built a zoo and stocked it, in large part, with animals that do not exist.

The lesson reaches far past finance. Any time someone mines a large pile of data for patterns and shows you the winners, you are seeing the survivors of a tournament of noise. Before you admire the winner, ask how many patterns were tried to find it.

McLean and Pontiff, 2016

David McLean and Jeffrey Pontiff finished the job by asking what happens to a pattern after its discovery. They took nearly a hundred published anomalies and measured their returns in three periods: inside the original study’s sample, after the sample ended but before publication, and after publication. Returns fell at each step, modestly out of sample, which measures how much of the original finding was luck, and sharply after publication, which measures something more interesting: investors read the paper, trade on it, and their trading shrinks the very pattern they read about. Publication is an act of arbitrage. A significant share of the average anomaly survived publication, which suggests many were partly real; much of each one did not.

So the zoo sorts itself, slowly, out of sample and in public. What remains after that sorting is a short list: the handful of premia that repeat across markets, survive their own fame at reduced strength, and come with a reason, either a risk someone must be paid to carry or a mistake people keep making because they are people. Everything else was weather in the data.

Three habits

This chapter compresses into three habits I use constantly.

First, decompose before you admire. When a fund, a strategy or a person shows you a track record, ask which of the known ingredients explains it, and what those ingredients cost separately. Pay active fees only for what is left after the ingredients are accounted for, which is usually little.

Second, discount every backtest, and discount it twice. Once for luck, because you are seeing the winner of a tournament you were not shown, and once for fame, because whatever the pattern earned in the past, it will earn less now that it is known. If a strategy only makes sense at the full advertised return, it does not make sense.

Third, demand a reason. Before trusting any premium, name who is on the other side and why they stay there: a risk they refuse to carry, a constraint they cannot escape, a mistake wired into being human. If the only argument is the data, you have met an animal from the zoo.

Gala: when a pattern makes money on paper, ask who has to keep paying for it, and why they would keep doing that after the pattern is famous. If nobody can answer that, do not put your money in it.