Surface PCA & Eigenmodes

An implied-volatility surface looks like it has a thousand moving parts. Every strike, every expiry, its own number. When the surface shifts overnight you are staring at hundreds of cells that all changed at once, and the question that matters for trading is brutally simple: did the surface actually do something, or did a few quotes wiggle and drag the rest along for the ride? You cannot answer that by watching a thousand numbers. You answer it by discovering that the thousand numbers are, almost always, secretly only three or four numbers wearing a costume.

That discovery has a name. It is principal component analysis, and it is the single most important dimensionality tool you will use on any surface, curve, or panel in this institute.

Start with the intuition before any algebra. Picture every daily snapshot of the surface as one point in a very high-dimensional space — one axis per strike-expiry cell. Day after day, you drop points into that space. What you find, empirically, is that the cloud of points is not a fat blob filling all the dimensions. It is a thin, stretched, cigar-shaped cloud. Most of the variation lives along one long direction; a bit more lives along a second direction perpendicular to it; a little along a third; and after that, almost nothing. Principal component analysis is nothing more than finding the directions in which that cloud is stretched, in order, from most stretched to least. Each direction is an eigenmode — a single coordinated pattern of movement across the whole surface — and each one comes with a number telling you how much of the total variation it carries.

Now the math, and it is cleaner than you expect. Take your data matrix, one row per day, one column per surface cell. First center it: subtract the mean of each column so you are working with deviations from the average surface, not the surface itself. Centering is not optional — principal components describe how things vary, and variance is defined around the mean. Call the centered matrix X. Form the covariance matrix, which is X-transpose times X divided by the number of observations minus one. That matrix is symmetric and positive semi-definite, which guarantees it has real, non-negative eigenvalues and orthogonal eigenvectors. The eigenvectors are your principal components — the directions of stretch — and the eigenvalues are the variances along those directions. Sort them from largest eigenvalue to smallest, and you have ranked your eigenmodes by exactly how much of the surface's daily motion each one explains.

There is a second road to the same destination, and in practice it is the one you should drive. Instead of building the covariance matrix, apply the singular value decomposition directly to the centered data X. The SVD factors X into U times Sigma times V-transpose. The columns of V are the right-singular vectors, and they are identical to the eigenvectors of the covariance matrix — the principal components. The singular values in Sigma are the square roots of the eigenvalues, scaled, so each singular value squared, divided by the number of observations minus one, gives you back the variance of that mode. Why prefer SVD? Because forming X-transpose-X squares your condition number and can throw away precision on a near-degenerate surface, whereas SVD works on X directly and stays numerically honest. Same answer, steadier hands.

The proportion of variance explained by the k-th mode is simply its eigenvalue divided by the sum of all the eigenvalues. Add the first three and, for a liquid equity-index volatility surface, you will typically find they account for the overwhelming majority of all daily variation. That empirical fact — that three modes nearly suffice — is the celebrated result of Cont and da Fonseca, who studied implied-volatility surfaces and found a strikingly stable, low-dimensional structure underneath them. And the three dominant modes are not abstract. They have shapes you can name.

The first eigenmode is level: a vector that is positive across essentially every cell, so loading on it pushes the entire surface up or down in parallel. The second is skew: positive on one side of the moneyness axis and negative on the other, so loading on it tilts the surface, steepening or flattening the smile's asymmetry. The third is curvature: positive in the wings and negative at the center, or the reverse, so loading on it pinches or puffs the smile — changing how convex it is. Level, skew, curvature. Three coordinated motions that, between them, reconstruct most of what the surface does on any given day.

This is where it stops being a math exercise and becomes evidence. When tomorrow's surface arrives, you project its deviation onto these three modes and read off three scores. If a move loads cleanly and largely on level, skew, and curvature — if the truncated three-mode reconstruction reproduces the observed surface with small residual — then the move is real, structural, the kind of coordinated shift that says the market repriced something. If instead the move lives mostly in the residual, in the fourth mode and beyond, you are looking at noise: a few stale quotes, a thin strike, microstructure litter. The eigenmodes are your filter for telling a genuine signal that the asset is about to move from a cloud of meaningless wiggle.