Skip to main content

Orthogonalizing Correlated Factors

Updated Jun 24, 2026 at 2:35pm

Research Draft High 1,222 words

When you estimate a security's sensitivity to several macro drivers at once — say, growth, inflation, real rates, the dollar, and oil — the drivers themselves overlap heavily. A weaker dollar tends to coincide with rising oil and rising breakeven inflation; growth surprises move rates. Regressing returns on all of them simultaneously produces multicollinearity: the betas become unstable, their standard errors balloon, and the model cannot cleanly attribute how much of the return came from which driver. Orthogonalizing correlated factors is the family of techniques that transforms the overlapping factor set into an uncorrelated (orthogonal) one, so each transformed factor carries only its own non-redundant information. The core tension is that you buy statistical stability and clean attribution at the cost of economic interpretability — the orthogonalized betas no longer answer the original question "how does the asset respond to oil?" in a way you can act on.

How it's calculated / formed

Two main families exist, distinguished by whether they treat factors asymmetrically (one privileged) or symmetrically (all equal).

Residualization / Gram–Schmidt (sequential, asymmetric). You pick an ordering. The first factor is left untouched. The second factor is regressed on the first, and its residual replaces it. The third is regressed on the first two, its residual kept, and so on. Each step strips out the part of a factor already explained by the earlier ones (StockCharts/financial-education and CRAN matlib both describe this as "replacing each variable by its residuals from a least-squares regression on the previous variables"). The resulting design matrix has a diagonal covariance matrix — correlations fall to numerical zero.

A crucial fact underpins the technique: the Frisch–Waugh–Lovell (FWL) theorem. The coefficient on any target regressor in a full multivariate regression is identical to the coefficient you get by regressing the residualized outcome on the residualized target. So residualizing the control factors against your factor of interest, while leaving the target factor whole, preserves the target's beta exactly and only redistributes the rest. Total R² is unchanged.

Symmetric orthogonalization (PCA / Löwdin). Instead of privileging an order, you transform all factors at once. Principal Component Analysis rotates the factor set onto the orthogonal eigenvectors of its covariance matrix; the components are uncorrelated by construction but are linear blends of the originals. Löwdin (symmetric) orthogonalization produces orthogonal factors that are, in a least-squares sense, as close as possible to the originals — useful when you want decorrelation without committing to an ordering or losing factor identity.

How it's used in practice

In macro factor-sensitivity work the most common pattern is asymmetric residualization with a chosen primary factor. If your thesis is about oil sensitivity, you orthogonalize the other factors against oil (not oil against them). FWL then guarantees the oil beta is the true partial sensitivity, and the other betas describe the dollar/inflation/rates after their oil-overlap has been removed. This keeps the one number you care about interpretable.

A second pattern decomposes total explained variance: orthogonal factors let you attribute R² cleanly to each driver without the cross-terms that make correlated-factor decompositions ambiguous (Klein & Chow, Quarterly Review of Economics and Finance, 2013, build orthogonalized factors specifically to make this systematic-risk decomposition well-defined).

A third use is building uncorrelated macro indices as inputs to a portfolio model — e.g. an "augmented" market factor that is the excess market return orthogonalized against observable macro fundamentals, isolating the market-specific variation. PCA-based decorrelation is favoured when the goal is dimensionality reduction (a handful of statistical macro factors) rather than preserving named economic variables.

Adoption, debate & evidence

Orthogonalization is standard, well-understood machinery in econometrics and quant finance — it appears in CFA Level II curriculum material on multicollinearity, in the MOSEK portfolio cookbook, and across academic factor-model papers. What is contested is whether it should be used at all for interpretation.

The honest landscape: residualization changes nothing about model fit (R², fitted values, and the target coefficient under FWL are invariant), so it is not a way to "improve" a regression — a frequent misconception. Critics (e.g. the Portfolio Optimizer note, and a broad statistics literature on "spurious" interpretability) stress that the technique is order-dependent and can manufacture nonsensical loadings. Their worked MSCI example residualized both World and Emerging-Markets factors simultaneously and got betas of ~308% and ~217% with essentially unchanged R² (~99.99%) — statistically valid, economically meaningless. The measured lesson: orthogonalization solves a variance/attribution problem, not an estimation-bias problem, and only the privileged (un-residualized) coefficients retain a clean reading.

Strengths & limitations

Works when: you have a single clear factor of interest among correlated controls (asymmetric residualization with FWL gives you a stable, interpretable target beta); or you genuinely want uncorrelated inputs for variance decomposition or as model features (PCA/symmetric methods).

Fails / misleads when: you orthogonalize all factors and then try to read every beta economically — the loadings become artefacts of the chosen order. The #1 misuse is treating residualization as a fix for multicollinearity that "lets you keep interpreting all coefficients normally." It does not: a residualized factor's beta now answers "response to the part of X uncorrelated with the other factors," which is rarely the question a trader or allocator actually asks. PCA components suffer a related problem — they may not map to any nameable economic driver, defeating the purpose of a macro sensitivity study. Finally, none of this rescues a spurious time-series regression: orthogonalizing non-stationary, trending macro series does nothing about the underlying unit-root problem (use returns/differences or test for cointegration first).

Sources

Dispute flagged: practitioners disagree on whether orthogonalized macro betas should ever be interpreted economically beyond the single privileged factor; this doc takes the cautious (FWL-grounded) view that they generally should not.