Eigenportfolio Statistical Arbitrage

Pairs trading needs a human to notice that two stocks move together. That doesn't scale past a few hundred names, and the pair you pick might stop cointegrating right after you commit capital. Eigenportfolio stat arb replaces the human's judgment with arithmetic: take the correlation matrix of a few thousand stocks' returns and let it tell you, mechanically, what moves together and by how much. What's left over — the part no common factor explains — is what you trade.

The crowd and the individual

Picture a stadium wave. Most of a spectator's up-and-down motion is the wave itself — everyone rises and falls together, and that shared motion is the dominant, easiest-to-see pattern in the crowd. But subtract the wave and each person still fidgets a little on their own: shifting in their seat, standing to stretch. That private wobble is uncorrelated with the neighbor's.

Stock returns work the same way. Most of a stock's daily wiggle is "the market" moving, then "its sector," then smaller shared currents. An eigenportfolio is the mathematically exact version of "the wave" — a specific weighted basket of stocks that captures one of these shared modes of motion. Once you subtract out the dominant eigenportfolios from a stock's return, what remains is the private wobble: a residual that has historically tended to mean-revert, because nothing systematic is pushing it in one direction.

Building the eigenportfolios

Take daily returns for $n$ stocks over some lookback and compute their correlation matrix, $\Sigma$ — an $n \times n$ table where entry $(i,j)$ is how correlated stock $i$ and stock $j$ have been. Principal component analysis (see PCA (Principal Component Analysis)) decomposes this matrix as

$\Sigma = V \Lambda V^{\top}$

Here $V$ is a matrix whose columns are eigenvectors — each one a set of portfolio weights, one weight per stock — and $\Lambda$ is a diagonal matrix of eigenvalues, each telling you how much of the total variance across all $n$ stocks that eigenvector's basket explains. Order the eigenvectors by eigenvalue, largest first, and the first one is (almost always) a long-only, roughly equal basket across every stock: the market. The next several look like sector or style bets — long energy against short utilities, or high-beta against low-beta — without anyone having labeled them "energy" or "utilities." The matrix found the groupings from co-movement alone.

Worked example 1. Take three correlated energy names — say XOM, CVX, and COP — and assume, for a clean closed form, that each pair has the same correlation $\rho = 0.7$. This "equicorrelated" structure has a known eigen-decomposition: the top eigenvalue is

$\lambda_1 = 1 + (n-1)\rho = 1 + 2(0.7) = 2.4$

with eigenvector $(1,1,1)/\sqrt{3}$ — equal-weighted, exactly the market mode. The other two eigenvalues are both $\lambda_2 = \lambda_3 = 1 - \rho = 0.3$, with eigenvectors that are contrasts (long one name, short another, net zero). The top eigenvalue (2.4) is eight times each of the others (0.3): almost all the shared variance in this trio is one mode, the equal-weighted basket. That's the number PCA is built to produce — a ranked list of how many independent "shocks" are really driving a group of names, and how big each one is.

Trading the residual

Regress each stock's daily return on the returns of the top $k$ eigenportfolios (typically 5–15, chosen where the eigenvalues drop to the noise floor). The regression's residual, $X_t$, is what the stock did that the common factors don't explain. Avellaneda and Lee modeled that residual's cumulative version as mean-reverting and standardized it into an s-score:

$s = \frac{X_t - m}{\sigma}$

where $m$ is the residual's long-run average (usually near zero) and $\sigma$ is its standard deviation. In words: how many standard deviations has this stock's factor-adjusted, cumulative residual drifted from its own normal resting point?

Worked example 2. Suppose a stock's residual has $\sigma = 1.0$ and today $X_t = -1.5$, giving $s = -1.5$. Avellaneda and Lee's rule bought when $s < -1.25$: the stock has fallen further than its factor exposure explains, so buy $1m of the stock and simultaneously sell $1m of the eigenportfolio basket that replicates its factor exposure, keeping the position market- and sector-neutral. If the residual reverts from $-1.5\sigma$ back to $0$ over the next several days while the eigenportfolio hedge earns roughly the market's return on the short leg, the trade nets close to $1.5\sigma$, or 1.5 points of residual reversion, minus financing and the sector hedge's own slippage — on the order of low tens of basis points per name, repeated across hundreds of positions.

What this means in practice, and what killed the easy money

This is a real, still-used technique — Avellaneda-Lee-style stat arb ran (and some variants still run) inside multi-strategy shops and CTAs' equity books. But the edge visible in mid-2000s backtests has decayed hard. Three things did it: speed, because once dozens of funds ran near-identical residual-reversion signals on overlapping universes, the reversion got arbitraged away within hours instead of days; crowding, because in August 2007 many quant books held nearly the same factor-neutral positions, and one large fund's forced deleveraging moved prices against everyone else simultaneously — a stat arb book that was "hedged" against the market lost 20%+ in three days because its hedge assumed normal correlations, not a fire sale; and estimation error, because the eigenvectors themselves are noisy estimates from a finite sample, so the "hedge" drifts as the lookback window rolls (see Random Matrix Theory in Finance for how much of a raw correlation matrix's structure is pure noise).

Drag the correlation on this scatter down toward zero and watch the fitted line flatten — that's the target state of a residual after a good eigenportfolio hedge: whatever correlation existed between two stocks through their shared market and sector exposure should be gone, leaving a cloud with no visible slope.

In interviews

Explain PCA on a correlation matrix in one sentence — it finds baskets of stocks that move together, ranked by how much variance each basket explains — then immediately separate "explains variance" from "predicts direction." Walk through the equicorrelated toy example by hand; interviewers want to see you compute $1+(n-1)\rho$, not just cite it. Close with why the strategy decayed: it's a crowding and estimation-error story, not a story about the maths being wrong.