Returns-Based Style Analysis

An allocator is evaluating a mutual fund that discloses only monthly returns, not its actual holdings, common for hedge funds and older-style disclosure regimes. The allocator still wants to know: does this fund actually behave like a large-cap growth fund, a small-cap value fund, or some blend? William Sharpe's returns-based style analysis answers this without ever seeing a single position, purely by comparing the fund's return pattern to a set of known style benchmarks and finding the mix that best reproduces it.

Guessing a recipe by taste alone

A chef tastes an unfamiliar dish and, without seeing the recipe, tries to guess the proportions: does it taste 60% like a tomato-based sauce, 30% like a cream-based one, and 10% like something spicier? They're not reading a recipe card, they're inferring composition purely from how the finished dish behaves against dishes of known composition. Returns-based style analysis does exactly this for a fund: it never reads the fund's holdings, only tastes its return pattern against benchmarks of known style, and infers the blend.

The regression

Given a fund's return series $r_t$ and a set of $k$ style benchmark indices (e.g., large-value, large-growth, small-value, small-growth), Sharpe's method solves a constrained regression:

$r_t = \sum_{j=1}^{k} w_j \, I_{j,t} + e_t, \quad \text{subject to } w_j \ge 0, \sum_j w_j = 1.$

In words: find the portfolio blend of style indices, weights that are non-negative and sum to 100%, exactly like a real portfolio would, that best reproduces the fund's actual monthly returns. The non-negativity and sum-to-one constraints are the key difference from an ordinary regression: they force the answer to be interpretable as an actual holdable portfolio, not an arbitrary set of positive-and-negative loadings that happen to fit the data. The fitted weights $w_j$ are the fund's implied style, and the leftover $e_t$, the part the style blend can't explain, is the fund's selection return, its performance beyond what pure style exposure would predict.

Worked example

A fund's monthly returns are regressed against four style indices. The fitted weights come out to: large-growth 45%, large-value 10%, small-growth 35%, small-value 10% (summing to 100%, all non-negative, as required). The regression's $R^2$ is 0.92, meaning the style blend explains 92% of the fund's month-to-month variance.

An allocator who expected, from the fund's name ("Large Cap Value Fund"), a large-value-dominated exposure now has concrete evidence the fund actually behaves 80% like a growth-tilted, midcap-leaning portfolio (45% large-growth + 35% small-growth), a significant style drift from its stated mandate, discovered entirely from public return data with no holdings ever disclosed.

The remaining 8% of variance ($1 - R^2$) plus the mean of the residual $e_t$ over time gives the selection return: if the fund averaged 0.15% per month of residual return, positive and not explained by any style tilt, that's evidence of genuine stock-picking skill on top of its (drifted) style exposure.

What this means in practice

• Detects style drift before it shows up in an official prospectus update, useful for allocators monitoring mandate adherence.
• $R^2$ tells you how trustworthy the style read is. A low $R^2$ (say, below 0.7) means the fund's returns aren't well explained by any combination of the chosen style indices, either because the style index set is wrong, or because the fund genuinely does something the standard style boxes don't capture (a multi-strategy or macro fund, for instance).
• Rolling-window analysis (re-running the regression on trailing 24- or 36-month windows) reveals whether a style tilt is stable or has been drifting gradually, which a single full-history regression would average away and hide.
• Related but distinct from factor-based attribution (see Factor-Based Performance Attribution), which typically uses a firm's own known holdings and a proprietary risk model's factors rather than public style indices and a fund's return series alone.

Practice

1. Why are the non-negativity and sum-to-one constraints essential to interpreting the fitted weights as "the fund's effective holdings," rather than just running an unconstrained regression?
2. A fund shows stable style weights over five years but a declining rolling $R^2$. What does that suggest is happening to the fund's strategy?
3. Two style indices used in the regression, small-growth and small-value, are themselves 70% correlated with each other. What problem does this create for interpreting the individual fitted weights, and what's it called? (See Multicollinearity.)