A portfolio manager can cut a skewed market down to q+2 funds only if one random direction stays independent. Framstad puts the condition in the abstract: returns are an affine combination of the selection variables "plus an elliptical term whose direction alone is independent." That word, alone, carries the contribution. By the abstract's account, the condition suffices for q+2 fund separation while relaxing Simaan's (1993) three-fund assumptions. I grant the relaxation fully. The skew factors V can depend on the magnitude R however the data dictates.
The direction has to stand apart.
The fund count
Framstad begins with an elliptical return vector observed after selecting on q auxiliary variables Z. Selection produces the skewness. A skew-elliptical law conditions on Z falling in a translated orthant; his weighted-selection-elliptical law instead reweights outcomes by any ϱ(Z) with mean 1. Drawn from Rao's weighted distributions, that construction also covers mixtures of conditionings.
The paper is pure theory, with no data or performance figure. Its payoff is a smaller asset menu: q+2 funds can preserve what an investor with nondecreasing utility can achieve. The author supplies the fund counts and model dimensions in this piece. The trading results below are ours, from a strategy built around his idea. There is no alpha source in the theorem.
The additive return form does the work. Excess returns comprise a nonnegative "good" term μ0R0, a linear loading M on q random variables V, and an elliptical shock R·A·U_ℓ. Here R measures the shock's nonnegative size, A sets its scale, and U_ℓ gives its uniformly drawn direction on a sphere. The independence requirement applies to U: it must be independent of R0, R and V jointly. R itself may depend on V without restriction.
Proposition 1 establishes that weighted selection produces this form. Framstad takes jointly elliptical n asset returns W and q selection variables Z, then changes the probability measure using a nonnegative function of Z whose expectation is 1. An indicator weight yields a selection-elliptical law conditional on Z entering a set; a translated orthant yields the familiar skew-elliptical case. Across these cases, the additive representation holds statewise with M = ΣΠ'(Λ+ΠΣΠ')^-1. Its direction remains uniform and independent under the reweighted measure.
The separation theorem follows. Suppose q < n and AA'+MM' is positive definite, as required by condition (12). For any portfolio ξ, a portfolio made from the funds f_j = (AA'+MM')^-1 μ_j, with j = 0 to q, has the same terminal wealth law plus a nonnegative bonus ϖR0. The formula for f0 applies when μ0 lies outside the column space of M; otherwise f0 may be any nonzero vector in ker(M'). This gives weak first-order stochastic dominance for every nondecreasing utility, whether concave or not. It covers an investor trying to maximize the probability of reaching a wealth target. The riskless asset brings the count to q+2 funds. When q = 1, the count is three.
Simaan (1993) required V to be independent of R and assumed concave utility. Framstad removes both requirements. As he puts it, "Also Simaan's assumption of concave utility index is superfluous." For selection-elliptical returns, the funds may instead be the columns of Π' together with Σ^-1μ0. That recovers the 2 + rank(Π) funds of Framstad (2011).
There is a case where Framstad needs more funds. In the main branch of Chamberlain's (1983) Theorem 2, μ0 = 0, q = 1, ℓ = n−1, EV ≠ 0 and there is no riskless asset. Chamberlain obtains two-fund separation through free disposal. Framstad writes that "our method needs a third." He defends the distinction for that setting (q = 1, μ0 = 0, nonrandom Y0, nonzero EV): "each univariate projection is indeed determined by its (finite) mean and (finite) variance." The extra fund stems from the comparison in Chamberlain's Theorem 1, which ranks terminal wealths "when also initial wealth is allowed to vary." More starting wealth can therefore compensate for a worse portfolio.
How independent is the direction?
The paper notes that, under the original measure, R is independent of V only when the underlying elliptical distribution is Gaussian. In a fat-tailed skew-t style model, large moves and the skew factor are dependent by construction. Framstad shows that this dependence does no harm to separation. It is a genuine loosening of Simaan's three-fund setup.
Direction is the constraint that remains. Framstad explicitly leaves aside Frahm's (2004) dependent-direction generalization. In a trading book, U passed through A determines which assets move together inside the elliptical shock. The theorem requires the same distribution of cross-sectional patterns for small and large shocks, and for high and low skew factors. When co-movement changes with shock size, the proof's characteristic-function step fails: it integrates out U conditional on everything else. My inference is limited to the proof. Once U loses independence, that argument no longer guarantees a q+2 fund portfolio with the original portfolio's wealth law.
We found no test of directional independence in the paper, as one would expect from a theory note. It is still the first assumption I would challenge with data. A practitioner can check whether residual directions remain uniform conditional on realized magnitude.
Estimation presents another difficulty. Framstad treats μ0, M, A, Π, Σ and Λ as non-random and discusses no estimation step. The fund f0 = Σ^-1μ0, used when μ0 lies outside the column space of ΣΠ', calls for the full inverse scale matrix Σ^-1. With 40 stocks, Σ has 820 free entries before any skew parameters. Restricting the book to three funds leaves that estimation burden intact and puts its errors into the fund directions. The model is single-period as well. Its multiperiod sketch needs fresh independent draws each period, intermediate consumption or free disposal, and an observable R0.
Our 40-stock book, 2020 to 2024
We fitted a one-selection-factor skew-normal model, with a Gaussian base and selection on Z > 0, to daily returns of the top 40 US stocks by trailing dollar volume in excess of SHY. Each monthly fit used the latest 756 completed sessions; the basket refreshed annually. We maximized expected exponential utility at risk aversion 3 within the span of Π' and Σ^-1μ0. The book had a 0.01 daily standard deviation budget, a 10% per-stock cap and gross leverage of 4.0. Residual weight went into SHY. Every setting was our choice; the paper specifies none.
This skew-normal fit puts the run in the paper's narrowest case. Under its Gaussian base, R is independent of V. Our book therefore never exercised the dependence Framstad allows beyond Simaan. The fitted model also makes the direction uniform by construction.
From 2020-01-31 to 2024-07-01, the restricted book returned 27.09% after commissions of $0.004 a share with a $1 minimum. Sharpe was 0.75, annualized volatility was 8.24%, and maximum drawdown was -21.20%. The 8.24% realized volatility was about half the roughly 15.9% annualized permitted by the daily budget, consistent with the caps or leverage limit binding. Maximum drawdown reached 2.6 times annual volatility.
We did not run an unrestricted 40-stock book on the same fits, so this result cannot establish what the fund restriction costs or saves. Beta to SPY was 0.09, leaving the book close to market-neutral. The 0.75 Sharpe omits borrow fees and margin financing on up to 4.0 gross; both would reduce it. Nor does the theorem's dominance identity govern our realized wealth. We traded on a fitted law rather than the true one, with inequality constraints in the form of 10% per-stock caps and a gross leverage limit. The paper says each linear constraint ξ'a ≤ α adds a fund f_a = (AA'+MM')^-1 a. If a 10% cap binds on any of the 40 stocks, the theory calls for another fund and three cease to suffice.
The theorem leaves the value of a q+2 fund restriction on real data open. A paired comparison would put our three-fund book beside an unrestricted 40-stock book using identical fits, caps and costs, then measure both on net utility and turnover.
One finding would change my view: residual directions passing a uniformity test conditional on magnitude across 2020 and 2022. Until then, Framstad's theorem remains clean mathematics whose decisive assumption has yet to face the market.
Our backtest stops at 2024-07-01, and everything after that date is deliberately left untouched so the same strategy can be checked out of sample later.