PCA

What is PCA?

Principal component analysis (PCA) re-expresses a set of correlated variables as a smaller set of uncorrelated ones. Technically it finds the eigenvectors of the data's covariance (or correlation) matrix: each eigenvector is a principal component, a weighted combination of the original variables, and the components are ordered by how much of the total variance they explain. Keeping only the first few compresses the data while retaining as much of the total variance as any linear projection of that dimension can, which is why PCA is the canonical dimensionality-reduction technique.

The method is classical statistics with a long finance career: Karl Pearson introduced it in 1901, Harold Hotelling systematized it in the 1930s, and fixed-income research made it famous by showing that a few components explain most yield-curve movement, conventionally read as level, slope, and curvature. Equity risk models, statistical arbitrage, and factor analysis all run on the same machinery.

Markets give it two natural jobs. Run across a basket of assets, the first component usually captures the shared move (loosely, the market factor), while later components pick up rotation and spread effects; each asset's loading measures how strongly it participates. Run across an indicator stack, PCA collapses dozens of overlapping inputs into a few orthogonal factors, a common feature-engineering step before model fitting, since redundant inputs inflate overfitting risk without adding information.

The caveats are as standard as the method. PCA is linear, so nonlinear structure hides from it; components are ranked by variance, and the most variable directions are not necessarily the most predictive ones for any given label; loadings drift as the estimation window rolls, and each component's sign is arbitrary, so labels like 'risk-on factor' need re-verification rather than faith. When nonlinearity matters, autoencoder-style neural networks generalize the compression at the cost of interpretability.

How to read a PCA tool on a chart

PCA outputs come as explained-variance shares, loadings, and component scores; each answers a different question.

1. 1Check the inputs and standardization first: PCA on raw prices is dominated by scale and trend, so returns (or z-scored features) are the defensible input.
2. 2Read the explained-variance table: how much of total movement the first components capture, and how quickly the shares decay.
3. 3Read the loadings: each asset's or feature's weight in a component says what that component is made of, and whether the 'market factor' story actually holds.
4. 4Track stability: recompute over rolling windows and watch whether loadings and variance shares hold; components that reshuffle every window cannot carry standing interpretations.
5. 5Use the scores: component values per bar serve as compressed features, regime dials, or the basis for spread and rotation reads.

How traders use it

• Extracting a market factor: the first principal component of basket returns typically behaves like an index, and an asset that departs from what its loading implies becomes a relative-strength or pairs-trading candidate.
• Compressing features: replacing many correlated oscillators and averages with two or three components that carry most of the variance, so a downstream model estimates fewer parameters from the same amount of data.
• Reading concentration: when the first component's share of explained variance rises, assets are moving as one block (risk-on/risk-off conditions); when it falls, dispersion is increasing and cross-sectional selection matters more.
• Preparing regime models: component scores make compact inputs for k-means regime labels or Markov-switching models, which behave better on a few orthogonal dimensions than on dozens of correlated raw series.
• Denoising correlation structure: truncating small components filters estimation noise out of correlation matrices, a standard cleanup before portfolio construction and spread selection.

PCA vs neighboring techniques