Factor analysis

Why does factor analysis matter for enterprise AI?

Factor analysis lets teams work with the small number of variables actually driving a set of measurements instead of treating every observed variable as independent. In market research, psychometrics, and risk modeling, that underlying structure is usually what the business needs to act on.

In financial services, factor analysis underlies risk factor models used in portfolio and credit risk management, where a handful of latent factors, like market risk, credit risk, and liquidity risk, explain correlated movement across many individual positions. Documenting which factors a model relies on, and how they were derived, is part of the model documentation examiners expect under guidance like SR 26-2, the Federal Reserve's current model risk management framework.

In life sciences, factor analysis validates patient-reported outcome (PRO) instruments and clinical rating scales, confirming that a set of survey items actually measures the underlying construct, like pain or quality of life, that a trial claims to measure. A regulatory submission that relies on a PRO instrument typically needs to show this validation work, not just the raw survey results.

How is factor analysis different from PCA?

Principal component analysis (PCA) finds the smallest set of dimensions that capture the most variance in a dataset, with no assumption of measurement error. Factor analysis assumes some of that variance is measurement error or noise, and models a smaller number of latent factors that explain the correlations among observed variables, a form of dimensionality reduction. Where PCA compresses data, factor analysis explains it.

Exploratory vs. confirmatory factor analysis

How do you implement factor analysis?

1. R: the psych package includes a dedicated function for exploratory factor analysis.
2. Python: scikit-learn's sklearn.decomposition.FactorAnalysis class covers standard exploratory factor analysis.
3. Python: the factor_analyzer package adds confirmatory factor analysis options that scikit-learn doesn't cover.

After extracting factors, most EFA workflows apply a rotation, like varimax, to make the factor loadings easier to interpret without changing how much variance the factors explain.

A minimal example using scikit-learn:

from sklearn.decomposition import FactorAnalysis

fa = FactorAnalysis(n_components=3, random_state=0)
factors = fa.fit_transform(X)

Industry use cases

Financial services

A single risk model might track hundreds of positions; factor analysis is what lets a risk team explain portfolio-level exposure through a handful of interpretable factors instead of tracking every position as its own independent risk. That factor structure is what turns a documentation requirement into something examiners can actually follow.

Life sciences

A single PRO instrument can include 20 or more individual questions. Factor analysis is what confirms those questions collapse into the two or three constructs, like physical function and emotional well-being, the instrument claims to measure, rather than measuring something else entirely.

Public sector

Public agencies use factor analysis on citizen or employee survey data to identify the underlying attitudes, like trust in a service or perceived responsiveness, driving satisfaction scores across dozens of individual survey questions. That underlying structure is what informs which service or policy changes are likely to move the numbers that matter.

Basic factor analysis illustration

Source: DataCamp

FAQ

What is factor analysis used for?

Factor analysis is most common in survey design, psychometrics, and market research, where a business wants to know the small number of real drivers behind dozens of individual questions. It shows up anywhere a set of correlated measurements is standing in for a handful of variables that can't be measured directly, like customer satisfaction or brand perception.

What's the difference between factor analysis and cluster analysis?

Factor analysis groups variables that move together to find the latent factors behind them. Clustering groups observations, not variables, based on their similarity to each other. Factor analysis explains why variables correlate; clustering finds which rows of data resemble each other.

How many factors should you extract?

Common approaches include the Kaiser criterion (keep factors with an eigenvalue greater than 1), a scree plot of eigenvalues looking for an elbow, and parallel analysis, which compares real eigenvalues against those from random data. All three are heuristics, and results are usually checked against domain knowledge of what the factors should represent.

Is factor analysis machine learning or statistics?

Factor analysis originated in statistics and psychometrics, decades before "machine learning" was common terminology. It's used today inside machine learning workflows, most often for dimensionality reduction and feature engineering ahead of a downstream model, but the technique itself is a statistical model with well-defined assumptions.

What are factor loadings?

Factor loadings are the coefficients that describe how strongly each observed variable relates to a latent factor, similar to how a regression coefficient describes a predictor's relationship to an outcome. A loading close to 1 or -1 means the variable is a strong indicator of that factor; a loading near 0 means it isn't related to that factor at all.