Principal Component Analysis
Principal Component Analysis compresses a return matrix into orthogonal drivers. It is useful for diagnosing concentration, hidden common exposures, and whether a multi-asset set is less diversified than it appears.
On this page
Features
- Covariance vs correlation mode selection
- Explained variance decomposition
- Factor loading heatmaps
- Concentration and diversification metrics to assess portfolio overlap
What it answers
- How concentrated is this basket? A dominant first component means many assets are moving together.
- How many distinct drivers exist? Effective component count and the cumulative variance curve show how quickly the matrix collapses into a few dimensions.
- Which assets load on each theme? The loading table shows which assets define each principal component.
Covariance vs correlation
- Covariance PCA preserves absolute volatility, so high-volatility assets can dominate.
- Correlation PCA standardizes each asset first, making the result about co-movement structure instead of scale.
Use correlation PCA when the question is “how many independent themes are here?” Use covariance PCA when absolute risk concentration matters.
Inputs and workflow
- You can start from explicit tickers or from a Library portfolio. Portfolio mode is useful when you want to diagnose overlap inside a reusable workspace portfolio without retyping its holdings.
- Daily mode is better for short-horizon structure. Monthly mode is better when you want a cleaner read on slower strategic relationships and less day-to-day noise.
- After the run, you can save, share, and reopen the analysis from Workspace.
- Reopening a saved or shared analysis replays the portfolio and the mode it recorded. Weighted-portfolio runs need one fixed set of weights. A portfolio that switches holdings on a signal or changes weights over time reopens with its holdings as tickers, and says why.
- A portfolio whose weights come from a linked saved strategy replays exactly as it ran, but those weights are not editable here. Editing them switches the analysis to its holdings as tickers and says why.
Interpretation
PCA components are statistical constructs, not economic labels. Treat them as clues. If the first component is broad and positive across equities, that suggests a common market driver, not a complete causal story by itself.
How to read the output
- Start with Top component share. If one component explains almost half the variance, the basket is less diversified than the ticker count suggests.
- Then check Effective components and the 80% threshold. These summarize how many dimensions carry the portfolio’s behavior.
- Finally, inspect the loading table. Similar signs and magnitudes indicate common movement. Opposite signs indicate an offset to that theme.
- Weighted-portfolio runs show Portfolio variance by component. It projects your weights onto covariance components and gives each component's share of portfolio variance. It always uses covariance decomposition. Thus, correlation mode can show different components in the loading table.