Asymptotic Principal Components

Ordinary principal component analysis on stock returns runs into trouble when you have thousands of assets but only a few hundred days of history: the covariance matrix between assets is huge and can't be estimated reliably from so few observations. Asymptotic principal components flips the usual approach around — instead of decomposing the (huge) covariance matrix between assets, it decomposes the (much smaller) covariance matrix between time periods, and recovers the same underlying common factors from that.

Concretely, with a returns matrix of $N$ assets over $T$ time periods where $N \gg T$, you compute the $T \times T$ matrix of cross-time-period covariances (comparing how the whole cross-section of returns on one day relates to another day), extract its eigenvectors, and these directly give you the time-series behavior of a small number of common factors. The asset-level factor loadings then fall out from a simple regression of each asset's returns on those recovered factors. The method is called "asymptotic" because its statistical guarantees are proven as $N \to \infty$: the more assets you add, holding $T$ fixed, the more precisely the common factors get pinned down, since idiosyncratic noise across assets averages out.

This is the practical trick behind large-panel factor extraction in equity risk models: rather than needing a huge number of time periods to estimate a huge asset covariance matrix, you lean on having many assets to estimate a handful of common factors from a modest amount of history.