Theorem 3.3 reduces the Maximum Likelihood Value proposed by Kanno, Ito, Yoshimura and Shibano to a minimum-variance residual. Start with the cross-section of log price returns in any numéraire. Subtract the minimum-variance portfolio's return from each one. The residual is the MLV return vector.
The paper says as much. Its abstract states that MLV "enforces a zero real return on the minimum-variance portfolio." Theorem 3.3 then shows that the closed-form estimate and the min-var weights coincide, up to a sign: Σ⁻¹1 divided by 1'Σ⁻¹1. Estimating latent real value and imposing zero return on the min-var portfolio are the same operation in the authors' construction.
The figures need one boundary around them. The paper backtests S&P 500 high-dividend names from June 2019 to January 2026, a universe we would not trade. An implementable version would instead use point-in-time US equity constituents plus liquid US ETFs. None of the paper's numbers would carry over. The discussion below does not present a failed replication of those results. The paper never says whether it used point-in-time constituent lists, leaving index-membership bias in its figures unmeasured.
The construction
Observed prices satisfy p_ij = v_j / v_i. With N assets, only N−1 prices are independent, leaving one degree of freedom in the latent real values unidentified. The paper assumes elliptical log real returns and maximizes their joint density over that free dimension. This produces the closed form above.
Estimating Σ requires a fixed point. The procedure starts at the identity, then alternates an MLV update with a Tyler M-estimator scatter update across a rolling window of 3N periods.
The synthetic experiment uses N = 500 multivariate-t series, each of length 10N, with a random correlation matrix. The authors conduct 10 runs at each n from 50 to 500. RMSE against the ground-truth correlation drops quickly, then levels off beyond roughly n = 200. They report n ≈ 500 as sufficient in practice, with each update costing O(n³).
The live backtest covers June 2019 through January 2026. It selects high-dividend S&P 500 names using trailing 12-month yield, estimates the MLV series from about 500 names and rebalances every five-week interval. The objective maximizes estimated expected dividend return after an L1 weight-change penalty, C = 0.01. Weekly CVaR at 95% may be no worse than −0.04, estimated over the past 100 weeks.
Only the risk space differs between the runs. In USD terms, replacing the USD risk space with MLV raises annualized return from 0.117 to 0.156 and Sharpe from 0.554 to 0.699. Turnover falls from 0.529 to 0.133. The tradeoff is higher volatility, from 0.211 to 0.223, and a worse maximum drawdown, from −0.376 to −0.466. In MLV terms, the same portfolios return 0.037 and 0.053. Sharpe rises from 0.281 to 0.461, while drawdowns are −0.220 and −0.148.
The inversion problem
Theorem 3.3 applies pointwise, making the min-var portfolio's MLV return exactly zero at every t. Within an estimation window, Σ_t is one matrix and w one vector. The realized MLV return of that portfolio is therefore identically zero across all 3N observations. Any scatter estimate formed from those returns is singular along w. Yet equation (5) requires an inverse of precisely that matrix, Σ_t.
Footnote 19 deals with scale instead. Tyler's estimator identifies scatter only up to a positive scale factor, while MLV is invariant to positive rescaling of Σ_t. Scale indeterminacy causes no problem. Singularity in the w* direction is a separate issue, which the paper does not address.
The authors already describe the Appendix A estimator as provisional. Their conclusion leaves more accurate and real-time methods for estimating Σ_t to future work, although every result in Section 4.2 uses the provisional method. The appendix instructs the reader to iterate "until convergence." We did not find an analysis showing that the fixed point exists or remains stable.
Implementation choices will govern the output: tolerance, iteration cap, and whichever regularization or pseudo-inverse is used. If the singularity argument above is right, different regularization choices will produce different Σ_t estimates and different weights.
Our hypothesis is that this geometry also drives the headline turnover result. Zero variance along the min-var direction makes an MLV-space risk cap slack exactly where the optimizer wants to move, allowing the solution to remain unchanged between rebalances. Turnover falls by almost the same amount under both evaluations, from 0.529 to 0.133 and from 0.527 to 0.137. The constraint geometry determines turnover; changing the evaluation unit leaves it untouched. We reached a similar view about the source of the gain in our note on MINGLE, where better covariance conditioning carried the Sharpe improvement.
Choices left to the implementer
- Problem (11) contains a budget constraint and a risk cap, with no sign constraint. The implementer must choose between long-only and long-short, while unconstrained Σ⁻¹1 routinely assigns short positions to individual names.
- Selection uses trailing 12-month dividend yield. The paper gives neither a name count nor a yield cutoff.
- Appendix B averages dividend returns over a lookback window L_t, but we did not find its length specified.
- Appendix A fixes the estimation window at 3N. With N ≈ 500, that means 1,500 observations. Section 4.2 does not identify the frequency of the underlying data, although the CVaR constraint and rebalancing interval are both weekly. At daily frequency, 1,500 observations cover about six years against a 6.6-year sample. Weekly data would require roughly 29 years, far beyond the sample.
The CVaR description adds another ambiguity. The text says the runs differ by "computing CVaR with covariance structures derived from either USD-based or MLV-based returns," while the footnote describes sorting the past 100 weekly portfolio returns and averaging the worst 5%. Those procedures estimate CVaR differently.
Does the residual measure real value?
Footnote 7 interprets Hovanov et al. as closing the missing degree of freedom by forcing the real log returns of all N assets to sum to zero at every t. MLV makes the same structural choice with min-var weights in place of equal weights. A normalization closes the free dimension, leaving the resulting unit definitionally blind to any real movement shared by the whole min-var basket.
Purchasing power is outside the authors' chosen standard. Section 1 rejects it deliberately because region-specific market baskets vary across jurisdictions and over time, making them ill-suited to cross-border transfer. The relevant question is whether min-var normalization better proxies unobservable real value. We do not think the paper establishes that claim. A basket identifies what remains constant. The min-var version holds constant whatever the estimated covariance marks as quietest in that window.
Evidence from a single configuration
The evidence remains as thin as 6.6 years from one configuration would suggest. There is one universe, one rebalancing frequency, one CVaR level, one cost coefficient and one sample period. The paper supplies no parameter sensitivity and no second window.
Roughly 68 five-week rebalances.
Under USD evaluation, the Sharpe gap is 0.145, moving from 0.554 to 0.699. Under MLV evaluation, it is 0.180, from 0.281 to 0.461. The paper emphasizes that the percentage increases are larger in MLV terms. Neither gap comes with t-statistics or standard errors.
The MLV-space columns use a unit constructed from the same asset set, so they do not provide an independent benchmark.
We found only the C = 0.01 penalty inside the objective as a cost treatment. The 4x turnover reduction is never translated into basis points of net return. That calculation would turn the result into something a trader could assess.
Two of the four authors work at VLUE Inc., and the paper discloses patent applications covering MLV and its portfolio application.
One result would change our view. Price the turnover gap using a real commission and spread, keep the objective fixed, and compare both risk spaces across several disjoint windows and a second universe. Continued net outperformance from the MLV risk space would show that covariance conditioning is doing real work. The current evidence is a gross win from one 2019 to 2026 sample.
The authors explain the low MLV-measured returns, 3.7% and 5.3% versus 11.7% and 15.6% in USD terms, through USD depreciation over the period. They say the difference "likely reflects" that depreciation. Until the result survives elsewhere, it reads as a constraint artifact with a good story attached.