Why Linear PCA May Miss Stock-Pricing Factors

Summary

The document considers whether principal component analysis across stocks can uncover pricing factors for a long-short strategy that buys shares deemed underpriced and shorts those deemed overpriced. Its answer highlights a key limitation: standard PCA maps high-dimensional data into a lower-dimensional representation through a linear transformation. Relevant structure that depends on nonlinear relationships may therefore remain undiscovered.

This is a caution about treating PCA components as a complete set of economically meaningful pricing factors. The response points toward comparing linear and nonlinear dimension-reduction methods, but it does not specify a particular alternative, provide empirical tests, or show that PCA-derived components support profitable trades. In practice, factor interpretation and valuation would require additional modeling and validation beyond extracting components from stock data.

Key ideas

• Standard PCA finds linear combinations of the input variables.
• Nonlinear relationships may not be captured by the resulting components.
• PCA components alone do not establish that a stock is underpriced or overpriced.
• The response recommends considering nonlinear dimension reduction but gives no empirical comparison.

Tags

# Why cant I use PCA to find all stock factors?


# Why cant I use PCA to find all stock factors?












Why cant I run all the stocks in the stock market thru a PCA model, and use the resulting principal components to create a factor model to price stocks and then buy stocks under priced and short stocks overpriced according to said model? Does it have to with the fact that PCA is unsupervised?

## Answer by Myoujin (score 1, accepted)

https://quant.stackexchange.com/a/72035

The PCA is applying the mapping from the high dimensional space to the lower dimensional space linearly. Thus, any relevant factors which are not linear will not be discovered. Please consider reading the following: Linear vs Non-linear dimension reduction.

Shown in full with attribution under the source's licence. Licence: CC BY-SA 4.0 (Stack Exchange)

This summary was written by Stratmill's research agent from the original; it is not a copy of the source.