Every portfolio runs four bets. Beta only shows one.

The one number — and what it hides

Beta is a covariance slope: it measures how a portfolio’s returns co-move with the market. Estimate it by regressing portfolio excess return on market excess return, and read off the coefficient — β = Cov(r_p, r_m) / Var(r_m). The regression also produces a residual: the part of the return stream the market did not explain.

A low beta is not a low-risk certificate. The same low beta can describe a genuinely diversified book or a concentrated bet wearing a hedge — and beta cannot tell them apart, because everything the market does not explain is swept into one residual term it never reports.

The whole picture is a single identity

One consequence is worth stating plainly: you do not have to trust a model to see past beta. Once the regression is run, the residual is orthogonal to the market by construction — that is what least squares does. Orthogonal terms have variances that add. So the fitted-component variance and the residual variance sum exactly to the total: the identity at the top of this page.

That identity is not a forecast and not a proprietary score. There is nothing to calibrate and no black box. The systematic share and the residual share are two slices that must sum to one. A variance decomposition is read like any set of shares — a few pieces that account for the whole — which is the first reason the method is easy to work with: the hard object, total risk, reduces to components that are required to add up.

Four layers, in the right order

The two-way split is correct but coarse. The residual term σ_ε² is not noise — it has structure. Most of it is sector and subsector exposure that a market factor was never going to explain. The instinct is to add factors. That is where most tools go wrong: throw raw sector and subsector returns into one regression and they overlap — with the market and with each other — so the coefficients double-count and the shares stop summing to anything meaningful.

The fix is sequence, not quantity. Estimate the market first. Residualize the sector factor against the market — keep only the part the market did not already explain — and estimate that. Residualize the subsector against both. What remains is the position itself.

This is Frisch–Waugh–Lovell logic put to work: regress each candidate factor on the factors that precede it, keep the unexplained part, and carry that orthogonalized factor into the next stage. Because every layer is orthogonal to the ones before it, the variances still add — now across four terms instead of two.

This is the difference that matters. The result is additive attribution — four shares, no double-counting, summing to the whole. It is not achieved by using more factors. It is achieved by using factors in the right order.

NVIDIA in four layers

The cascade earns its keep only when the four shares are real numbers on a real stock. Run it on NVIDIA. As of 20 May 2026 its market beta is 1.75 — the figure on every terminal. Estimated against the sector and subsector factors in turn, the same decomposition reads:

Each loading governs a share of NVIDIA’s return variance:

The four shares sum to 100% — the additive property, on a live stock. The subsector layer contributes essentially nothing: once the sector is removed, the semiconductor sub-industry adds no further independent information for a company this size, so its orthogonalized share rounds through zero.

Source: RiskModels ERM3 decomposition, as of 20 May 2026. Figures roll as the rolling estimation window updates.

Every position has its own benchmark

The NVIDIA breakdown carries a general point: a position is never simply “a tech stock.” It carries a market component, a sector component, a subsector component, and a piece that is only itself. Each component is a benchmark the position is implicitly trading against.

Roll those layers up across every holding and the portfolio inherits a benchmark of its own — the weighted composite of what its positions are actually exposed to. That is rarely the index named in the prospectus. The gap between the stated mandate and the realized benchmark is precisely what a layered decomposition measures, and precisely what a single beta cannot.

Reading a decomposition well

A decomposition is only as honest as its inputs. Four cautions carry most of the weight:

• Near-zero beta: a market-neutral book can still carry large sector, subsector, basis, liquidity, or security-specific variance.
• Multicollinearity: raw sector and subsector proxies overlap with the market; residualization is required before any layer-specific coefficient can be interpreted.
• Changing regimes: rolling betas and residual shares should be monitored, because the same ticker migrates across factor regimes as fundamentals and market structure change.
• Short windows: noisy estimates make residual attribution unstable; use shrinkage, minimum-history requirements, and clear data lineage for production reporting.

Three questions settle most beta-based risk statements:

1. What share of total variance is explained by the market?
2. What sector and subsector exposures remain after the market is stripped out?
3. How much variance is still residual once every hierarchical layer has been removed?

From the identity to one upload

The arithmetic is simple. Doing it on a real book is not — it means estimation windows, residualization at every layer, GICS-aligned subsector nodes, and clean return data across more than 45,000 stocks, funds, and ETFs. That engineering is what the Analyst Workspace does for you. Describe or import a portfolio and it returns the four-layer decomposition derived here, measured on your own holdings. The method is in this primer; the workspace is the single upload that runs it.