Plain OAS gives up too much to heavy tails. Lewis measures the damage under Student-t data with five degrees of freedom: with 30 observations on 120 variables, Oracle Approximating Shrinkage (OAS) has normalized Frobenius loss of 1.1907. His correction brings it to 0.6426. That repairs a badly misspecified OAS estimate and leaves the corrected estimator within 2% of Ledoit-Wolf's 0.6308.
How the fourth moment changes OAS
OAS, from Chen and co-authors, shrinks the sample covariance toward a scaled identity. The scale comes from the sample's own average eigenvalue, Tr(Ŝ)/p, so the target itself is random. A fixed-point iteration for the best blend weight reduces to a closed form. Chen et al. derived the two expectations it needs, E[Tr(Ŝ²)] and E[Tr²(Ŝ)], under Gaussian data.
Lewis derives both expectations without assuming Gaussian returns. Each gains an extra α/n. Here α sums the fourth-order joint cumulants κ(X_i, X_i, X_j, X_j) across all pairs; for Gaussian data, α = 0. The resulting intensity, after the same fixed-point step, is min(1, (1+β)/((n+1−2/p)φ̂)), where β is proportional to α. Positive α raises the shrinkage applied to fat-tailed data. Estimating α takes the sample variance of squared observation norms, then subtracts 2Tr(Ŝ²), the Gaussian value for covariance Ŝ. It costs O(np). An appendix establishes that 1+β remains non-negative, keeping ρ in [0, 1] without clamping.
Lewis tests the correction in two settings. The Monte Carlo has p = 120, AR(1) covariance with Σ_ij = 0.5^|i−j|, n/p from 0.25 to 5, and 500 trials per cell. Its draws are Gaussian, t(10) and t(5). The equity study constructs global minimum-variance (GMV) portfolios from 50 large-cap US stocks drawn from the S&P 500. Daily data run from January 2010 to December 2023, with 252-day estimation windows and 21-day holdout months, yielding 155 windows.
Does OAS-Extended beat Ledoit-Wolf?
Under t(5), OAS-Extended reduces loss against OAS by 46 to 48% when n/p ≤ 0.5. The gain falls to about 36% at n/p = 1, about 24% at n/p = 2 and about 16% at n/p = 5. Larger samples narrow the gap; greater kurtosis widens it. Both directions follow the formula.
Ledoit-Wolf (LW) remains the harder comparison. Its loss equals or betters OAS-Extended's in every non-Gaussian cell of Lewis's table: 0.6046 against 0.6120 for t(10) at n/p = 0.25, then 0.5621 against 0.5627 for t(5) at n/p = 1, with both at 0.4116 when n/p = 5. Lewis acknowledges the result. His abstract calls OAS-Extended "performing comparably across all tested settings" and presents it as "a principled correction for OAS practitioners facing heavy-tailed data."
The case for OAS began with its small-sample advantage over LW under Gaussian data. At n/p = 0.25, the losses here are 0.5866 and 0.5884. OAS-Extended gives up that edge: it reaches 0.5935 in the same Gaussian cell, which Lewis attributes to noise in α̂. Across the simulated losses (p = 120, 500 trials), a corrected OAS user is therefore slightly behind LW on Gaussian data at n/p ≤ 0.5. On fat-tailed data the user is within about 2%, with ties from n/p = 1.
Two issues bear on the novelty claim. Lewis argues for greater generality than the Ollila line of elliptical LW extensions because his result requires only finite fourth moments. Yet every simulated distribution is elliptical (Gaussian or Student-t). The conclusion also limits the study to Student-t distributions and one AR(1) covariance. We did not find Ollila's estimators among the benchmarks. The paper proves generality, while the simulations leave it untested; the equity returns are real, though nobody checks whether they are non-elliptical.
There is a further question about α̂. Its consistency argument uses the law of large numbers at fixed p. In the paper's expansion, E[Tr(Ŝ²)] contains a Tr²(Σ)/n term. At p = 120 and n = 30, that term is appreciable, yet α̂ subtracts 2Tr(Ŝ²) without removing it. On our reading of equations (7), (8) and (19), α̂ is consequently biased downward under Gaussian data by roughly 2Tr²(Σ)/n. A negative α̂ lowers shrinkage intensity, whereas the paper says noise inflates it. Bias could explain the Gaussian shortfall at n/p = 0.25. Lewis identifies behaviour at n/p ≪ 1 as future work.
The GMV result, and the VIX split
Across the equity test, mean annualized realized GMV volatility is:
- sample covariance: 13.972%
- OAS: 13.552%
- RBLW (Rao-Blackwell Ledoit-Wolf): 13.551%
- OAS-Extended: 13.282%
- Ledoit-Wolf: 13.276%
OAS-Extended's advantage over OAS has t = 3.83; LW's has t = 3.85. Those statistics say little about the size of the gain. RBLW, too, clears p < 0.001 against OAS, with t = 3.88, although its mean variance reduction is below 10^-7. Lewis calls that "statistically detectable but economically negligible." The mean variance reduction for OAS-Extended exceeds RBLW's by more than 30 times. Its bootstrap 95% interval is [2.0, 5.0]×10^-6.
Lewis treats the VIX split as empirical confirmation. A mean in-sample VIX of at least 20 defines a stressed estimation window: 56 qualify, alongside 99 calm windows. Against OAS, OAS-Extended lowers volatility by 1.0% in calm windows, 13.649% against 13.785%. The cut reaches 4.0% in stress windows, 12.607% against 13.131%. LW records 1.0% and 4.1%.
One detail is odd.
OAS's out-of-sample volatility after stressed estimation windows is 13.131%, below the 13.785% after calm windows. High-VIX estimation periods may often precede calmer months; the "stress" label applies to training data rather than the trading month. Weighting those regime rows by their window counts yields about 13.273% for OAS-Extended, slightly different from Table 2's 13.282%.
Lewis does not explain how the 50 stocks were selected. Choosing names with knowledge of which large caps survived through 2023 would introduce survivorship.
What did the portfolios measure?
The paper evaluates unconstrained, cost-free GMV portfolios by realized volatility. We found no turnover or cost figures, nor any return or Sharpe numbers. Volatility is a reasonable stopping point for a covariance paper, but it confines the trading claim to the 2.0% volatility cut. Net P&L lies outside the test. With no reported turnover, the effect of costs remains unmeasured.
Our automated trading run
We also built a traded version. Our period runs from 2020 to mid-2024, with metrics starting on 31 January 2020. The universe is the top 50 non-ADR US stocks by capitalization, refreshed annually; Lewis instead uses 50 S&P 500 names over 2010 to 2023. At each month-end, we fit OAS-Extended, OAS and analytical LW using 252 centred daily log returns, then solve for GMV weights. We trade OAS-Extended after projecting its weights onto a fully invested portfolio with a 10% absolute position cap and gross leverage up to 4×. Rebalancing occurs at the close. OAS and LW remain shadow books.
From 31 January 2020 to mid-2024, our capped, levered OAS-Extended portfolio returned 40.11%. Its Sharpe was 0.53 (Sortino 0.67), annualized volatility was 17.10%, and maximum drawdown was -29.86% (Calmar 0.27). The 0.53 Sharpe alongside a -29.86% drawdown is weak. Our 17.10% volatility exceeds the paper's 13.282%, but the figures describe different objects. Ours is one levered, capped, cost-bearing portfolio over 2020 to 2024. Theirs averages 155 per-window annualized volatilities for an unconstrained, cost-free portfolio over 2010 to 2023. Neither figure ranks the other. Our returns include commissions of $0.004 a share with a $1 minimum, and our measurement window opens on the 2020 crash.
Lewis's claim would be tested by the shadow-book gap between OAS-Extended and OAS, preferably separated by VIX regime. We did not compute that gap in this run. Our result therefore can neither confirm nor contradict the paper's 2.0% or 4.0%. We also have not established whether our annual capitalization screen was available on each selection date. Short borrow and margin financing are excluded from our figures, although a 4× gross portfolio would incur them. This was one automated pass, rather than a verdict on the authors' work.
Who should switch?
For a plain OAS user working with daily equity returns, Lewis's GMV result favours switching: volatility falls from 13.552% to 13.282% with OAS-Extended. In the t(5) simulation at n/p = 0.25, OAS's 1.1907 loss is close to double LW's 0.6308; the correction costs O(np). An existing Ledoit-Wolf user gets no comparable reason to move from these tables or portfolios. The GMV figures, 13.282% against 13.276%, are effectively tied. A non-elliptical simulation, or a result at n/p well below 0.25 in which OAS-Extended beats LW, would change that view. The paper has neither yet.
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.