Fundamental vs Statistical Factor Models

Every equity risk model needs to answer the same question: what are the handful of shared drivers behind thousands of stocks moving together, and how exposed is each stock to each driver? There are two fundamentally different ways to answer it, and they start from opposite ends of the same equation, $r_i = \sum_j \beta_{ij} f_j + \epsilon_i$ (a stock's return equals its exposures $\beta$ times factor returns $f$, plus a leftover). A fundamental model decides what the factors mean first, and then estimates the exposures. A statistical model decides the exposures first, by finding whatever moves stocks together in the data, and figures out what to call the factors afterward, if it can at all.

Two ways to sort a box of unlabeled parts

Imagine you're handed a box of unlabeled mechanical parts and told to sort them into a small number of bins that explain most of what's going on. One approach: you already know a fair amount about mechanical parts, so you sort by category you can name in advance, screws, bolts, springs, gears, and measure each part against those known categories. The other approach: you don't presume to know the categories, so you look for whatever groupings the parts naturally fall into by size, weight, and shape, cluster them statistically, and only afterward try to guess what each cluster "is." Both can produce a workable sorting system. The first is interpretable by design, built from things you already understand; the second can capture patterns you never thought to name, at the cost of a bin you might only be able to describe as "cluster 4."

The fundamental approach: factors first

A fundamental (also called cross-sectional) model starts by specifying factors that are believed, on economic grounds, to matter: value, size, momentum, quality, an industry classification, and so on. For each stock and each factor, you compute an observable characteristic, book-to-market ratio for value, market cap for size, trailing 12-month return for momentum. These characteristics are the exposures $\beta_{ij}$, known and computed directly from data you already have, no estimation needed for that half of the model.

What's unknown is the factor return $f_j$, how much each factor actually paid off this period. That is recovered by running a cross-sectional regression, across all stocks, of returns on the known characteristics, for each time period:

$r_{i,t} = \sum_j \beta_{ij,t} f_{j,t} + \epsilon_{i,t} \quad \text{(regression across stocks } i \text{, at each date } t\text{)}.$

In words: at each point in time, given how exposed every stock is to each factor, find the factor payoffs that best explain that period's cross-section of returns. This is exactly the machinery behind Fama-MacBeth regression, see Fama-MacBeth Regression, and it is what commercial models like Barra run every day; see Barra-Style Equity Risk Models.

The statistical approach: exposures first

A statistical model has no prior beliefs about what matters. It takes a long panel of historical returns for many stocks and applies a dimension-reduction technique, most commonly Principal Component Analysis (PCA), to find a small number of return patterns that explain the most variance in the whole panel. In words: find the small number of "typical daily moves" such that knowing a stock's sensitivity to each of them explains most of how that stock actually moved, historically, without ever being told what those moves mean economically. The first statistical factor in equities is almost always something that looks like "the market," because market-wide co-movement is the largest source of shared variance, but factor two onward often defies clean labeling, an uncomfortable mix of sector and momentum and something else that doesn't map to a single story.

statistical historical return panel PCA extracts factors + exposures jointly try to label factors afterward

Worked example: same stock, two models disagree

A regional bank stock has a fundamental-model exposure profile of value $\beta = 0.8$, size $\beta = 1.1$ (small-cap), financials-industry $\beta = 1.0$. Its fundamental-model predicted variance, using factor volatilities of 12%, 15%, and 18% respectively and assuming (for simplicity) zero cross-factor correlation, is

$\sigma^2 \approx (0.8 \times 0.12)^2 + (1.1 \times 0.15)^2 + (1.0 \times 0.18)^2 = 0.0092 + 0.0272 + 0.0324 = 0.0688,$

giving an annualized systematic volatility of about 26.2%.

A statistical model built with 10 principal components on the same stock might find its exposure concentrated in statistical factor 2 (which correlates 0.7 with the fundamental "size" factor but is not identical to it) and statistical factor 5 (which has no clean fundamental label at all, it's picking up something real, perhaps a regional-economic-exposure pattern the fundamental model's fixed factor list never included). Its predicted systematic volatility might come out at 29%, a genuine difference: the statistical model is capturing a real source of co-movement the fundamental model's factor list simply never asked about.

Worked example: risk decomposition after an oil shock

An oil-price shock hits. A fundamental model with an explicit energy-sector factor immediately shows every energy stock's shock-day loss attributed cleanly to that one named factor, easy to report, easy to hedge against directly. A statistical model, if it was estimated on a calm pre-shock history that never contained a distinct "oil" pattern with enough variance to earn its own principal component, may instead spread the same loss thinly across several unlabeled components and a large idiosyncratic residual, technically capturing the co-movement (stocks that always moved with oil, even without anyone calling it that, still load on some combination of components) but making the shock much harder to explain in a risk report the next morning.

What this means in practice

• Most serious commercial risk models are hybrids. Barra and similar vendors use fundamental factors as the primary structure but often add statistical or "specific risk" adjustments to catch what the named factors miss.
• Statistical models adapt faster to new, unlabeled sources of risk, useful in fast-moving markets (crypto, new asset classes) where nobody has agreed on the right named factors yet.
• Fundamental models are what portfolio managers can act on. A statistical factor return of "principal component 6" doesn't tell a PM what to hedge; "value factor return" does.
• Both need a covariance matrix over the factors themselves to turn exposures into portfolio risk; see Estimating the Factor Covariance Matrix for how that's built and why it's the hardest remaining step either way.

Practice

1. A statistical model's third principal component correlates 0.6 with momentum and 0.5 with a reversal signal. Is this factor "momentum," "reversal," both, or neither? What does your answer say about the limits of naming statistical factors?
2. Why does a fundamental model's exposure to a brand-new factor (say, "AI exposure") require the model builder to add a new characteristic, while a statistical model would, in principle, pick it up automatically, if the data shows enough co-movement?
3. Explain why a fundamental model's factor payoffs $f_{j,t}$ need to be re-estimated every single period, while its exposures $\beta_{ij}$ usually do not.