REVIEW 3 major objections 5 minor 15 references
Multiscale Markowitz
T0 review · 3 major / 5 minor · reviewed 2026-08-12 · deepseek-v4-flash
Pith's one-line read A scale-averaged covariance matrix gives higher out-of-sample Sharpe and Sortino ratios and lower drawdowns than daily Markowitz, according to the paper's US sector and factor ETF backtests.
desk verdict A clean but under-supported empirical idea: the multiscale covariance estimator is worth testing rigorously, but the paper's own implementation doesn't yet establish the claimed edge. read the letter →
The pith
A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.
The reading
What carries the argument
The central object is the multiscale covariance estimator $\Sigma^{MS}_{ij} = \langle \Sigma_{ij}(\Delta t)/\Delta t \rangle_{\Delta t}$, the average over a range of time scales $\Delta t$ of covariances estimated from returns aggregated to those scales. The optimization uses this estimator in place of the daily covariance matrix, with the usual minimum-variance or maximum-Sharpe objective; a target Hurst exponent $H_{\rm target}$ can be imposed so that portfolio variance scales as $\sigma^2_{\rm target}(\Delta t) \propto (\Delta t)^{H_{\rm target}}$ across frequencies. This estimator carries the argument because it is what makes the optimization sensitive to how risk builds up from daily to monthly horizons, and the paper derives sensitivity results showing that weights fall as volatility, Hurst exponent, or correlation rise.
What would settle it
Rerun the same 2019-2024 backtest on sector ETF returns that have been randomized within each asset to destroy time-series scaling while preserving cross-sectional correlations and volatility; if the multiscale portfolio still beats daily Markowitz, the apparent advantage is not due to genuine scale-dependent risk.
Extended reading notes
Core claim
The paper's central claim is that optimizing a portfolio with the multiscale covariance matrix $\Sigma^{MS}_{ij} = \langle \Sigma_{ij}(\Delta t)/\Delta t \rangle_{\Delta t}$ outperforms traditional Markowitz based on daily variances and covariances. This estimator averages covariances computed at several return aggregation intervals, capturing how volatility and correlation scale with time; the paper connects this scaling to a standardized Hurst exponent $H := \beta/\alpha$ from a fractional diffusion equation. On 11 S&P 500 sector ETFs and 9 factor ETFs, the multiscale minimum-variance portfolio reports Sharpe ratios around 0.53 in both universes, versus 0.35 and 0.43 for traditional Markowitz, with Sortino ratios and maximum drawdowns also improving. The authors interpret this as evidence that scale-dependent risk structure, including the Epps effect and rough volatility, is economically exploitable by choosing a target Hurst exponent.
Load-bearing premise
The load-bearing premise is that the scale-averaged covariance matrix estimated from a rolling 125-day window, with lower frequencies built from non-overlapping sub-samples, remains a reliable estimate of true risk at those horizons; if those low-frequency estimates are noisy or the normalization across scales is arbitrary, the reported improvement could be an artifact.
Editorial extensions
If this is right
- If the claim is correct, investors can choose a target Hurst exponent to express preferences for short-horizon versus long-horizon risk, rather than fixing only daily variance.
- Multiscale optimization should reduce exposure during crashes, since the scale-averaged covariance penalizes the low-frequency volatility spikes that daily Markowitz misses.
- The same estimator applies to factor rotation and other long-only allocation problems, not just sector rotation.
- The results imply that single-scale Markowitz misallocates to assets with Hurst exponents far from one half, such as illiquid or rough-volatility assets.
- The multiscale covariance can be replaced by an L1-modified version to preserve convergence under fat-tailed return distributions.
Reading between the lines
- The reported Sharpe improvement is measured over a single five-year window with two major drawdowns; testing over multiple decades or many rolling sub-periods would show how much of the gain is regime-specific.
- The low-frequency part of the estimator relies on a small number of non-overlapping sub-samples inside a 125-day window; if those terms are noisy, a different choice of scale set could weaken or strengthen the advantage.
- A direct test would use simulated multifractal data with known Hurst spectra to verify that the multiscale estimator recovers the true scale-dependent risk and that the optimization gains are not an artifact of the averaging scheme.
- The framework could be extended to cross-asset portfolios with bonds and commodities, where the Epps effect and rough volatility are more pronounced, to see whether the performance gap widens.
Editorial analysis
A structured set of objections, weighed in public.
Referee Report
Summary. The paper proposes a 'multiscale Markowitz' framework in which portfolio risk is measured with a covariance estimator averaged over multiple return horizons, Sigma^MS_ij = <Sigma_ij(Delta t)/Delta t>, motivated by anomalous scaling sigma(Delta t) proportional to (Delta t)^H. It presents a sensitivity analysis for minimum-variance weights and an out-of-sample backtest on 11 SPDR sector ETFs and 9 factor ETFs over 2019-2024, reporting higher Sharpe and Sortino ratios and lower drawdowns than daily-covariance Markowitz (Tables 1-2). The abstract and introduction also promise a toy example and a target-Hurst optimization, but the delivered backtest uses only the averaged-covariance option (Section 3), and no toy example appears in the text.
Significance. If the empirical claim were robust, the multiscale covariance estimator would be a simple and potentially useful alternative to single-scale covariance estimation for long-only portfolio optimization. The paper's out-of-sample design has strengths: it compares against equal-weight and traditional Markowitz in two asset universes, and it does not fit parameters to the Sharpe outcome, so the central comparison is not circular. However, the main contribution is empirical, and the evidence is not statistically established: there are no confidence intervals, no sensitivity analysis for the scale grid, and the scale normalization has a load-bearing arbitrariness. The reported Sharpe gaps (0.53 vs 0.35 and 0.53 vs 0.43) are therefore not yet convincing evidence of superiority.
major comments (3)
- [Section 7, Tables 1–2] The paper's central claim, stated in the introduction as 'we evidence on US sector index tracking ETFs the superiority of multifrequency optimization over traditional Markowitz,' rests on a single five-year backtest with no measure of statistical uncertainty. The estimator Sigma^MS_ij = <Sigma_ij(Delta t)/Delta t> is computed on a 125-day lookback with non-overlapping lower-frequency blocks; at Delta t = 5 this gives about 25 weekly observations and at Delta t ≈ 21 about 6 monthly observations. For 11 sector ETFs or 9 factors, the monthly covariance block is rank-deficient, and no shrinkage or regularization is described. The Sharpe differences in Tables 1–2 therefore have no confidence intervals, bootstrap, or subperiod checks, so finite-sample noise cannot be ruled out. Section 7 also states that transaction costs are 'assumed to be negligible' without reporting turnover or rebalancing frequency, which can bias a Sharpe comparison in favor of whichever method trades more. Please add error bars or bootstrap/subperiod splits, report the exact scale grid and the number of independent observations per block, and quantify turnover and transaction-cost sensitivity.
- [Section 4.2] The definition Sigma^MS_ij = <Sigma_ij(Delta t)/Delta t> is not scale-invariant under the paper's own scaling law. If sigma(Delta t) is proportional to (Delta t)^H, then Sigma_ij(Delta t) is proportional to (Delta t)^{2H}, so Sigma_ij(Delta t)/Delta t is proportional to (Delta t)^{2H-1}; unless H = 1/2 for all pairs, the average over scales depends on the choice of scale grid and on the weights in the average. The manuscript does not specify the set of scales used, the weights, or whether the grid is varied in the backtest. Consequently, the reported improvement in Tables 1–2 could be driven by the arbitrary normalization rather than by a genuine multiscale effect. Please report the scale grid and demonstrate robustness to alternative grids and weights, for example by excluding the monthly block or using equal weights across scales.
- [Abstract and Sections 1, 3, 5] The abstract promises both a target-Hurst formulation and a toy example ('We illustrate this concept with a toy example'), and the introduction repeats the toy-example plan. However, Section 3 explicitly drops the target-Hurst implementation ('We consider only the last case since the first does not sufficiently constrain the portfolio'), and no toy example or numerical illustration of the multifractal optimization problem appears in Sections 5–7. The delivered backtest uses only option 3, the averaged covariance matrix. Either add the promised toy example and target-Hurst experiments, or revise the abstract and introduction to state that the delivered contribution is the averaged-covariance estimator and its backtest.
minor comments (5)
- [Section 2.1] The notation for H is overloaded: the abstract uses H in sigma(Delta t) proportional to (Delta t)^H, while Section 2.1 redefines H as beta/alpha via the fractional PDE. Please align the notation throughout.
- [Section 6] The Lagrangian solution w = Sigma^{-1}1/S does not enforce the non-negative weight constraint stated in Section 4.2, so the sensitivity conclusions do not directly apply to the long-only optimization used in the backtest.
- [Section 5.2] The statement 'Empirically from studies of the Epps effect, we find that H_rho ≈ 0.3' is given without a citation or derivation; please provide a reference or a supporting calculation.
- [End of manuscript] The file 'epps1.png' is listed at the end of the manuscript but is never referenced in the text; add a figure with a caption or remove the dangling file.
- [Throughout] There are scattered typographical errors, e.g., 'constra int' in the abstract and 'T V olatility' in Section 1.2; a careful proofread is needed.
Circularity Check
No significant circularity: the multiscale estimator is tested out of sample against a daily-covariance Markowitz baseline, with no parameter fitted to the reported performance.
full rationale
The paper's central empirical claim is the out-of-sample superiority of a minimum-variance portfolio built on the multiscale estimator Sigma_MS_ij = <Sigma_ij(Delta t)/Delta t>_{Delta t} over a conventional daily-covariance Markowitz portfolio (Section 7, Tables 1-2). Nothing in this comparison is circular: Sigma_MS is defined directly from historical return covariances at several scales, the portfolio weights are the standard closed-form min-variance solution w = Sigma^{-1}1/(1^T Sigma^{-1}1) derived in Section 6, and neither the estimator nor the weights are fitted to the Sharpe, Sortino, or drawdown outcomes reported. The scaling law sigma(Delta t) proportional to (Delta t)^H is used as motivation and in the sensitivity derivatives, not as a fitted variable used to produce the backtest. The references are to external work (Mandelbrot, Gatheral et al., Calvet and Fisher, etc.) and no load-bearing claim rests on a self-citation. The main caveats are statistical rather than logical: the 125-day lookback leaves only about six non-overlapping monthly observations, and the paper reports no error bars or sensitivity to the scale grid; such fragility concerns the validity of the evidence, not whether the derivation reduces to its inputs. The target-Hurst optimization advertised in the abstract is not used in the empirical section, but that abandonment is a scope limitation, not a circular step.
Assumptions & free parameters
free parameters (3)
- lookback window =
125 days
- set of time scales and dt normalization =
not specified
- Epps-effect scaling constant H_rho =
approx 0.3
assumptions (6)
- domain assumption Asset return variances and covariances obey the scaling law sigma(dt) proportional to dt^H with a well-defined Hurst exponent.
- ad hoc to paper The fractional diffusion equation d^beta P / dt^beta = -K_alpha (-Delta)^alpha P is an appropriate model for returns and supports the standardized Hurst H = beta/alpha.
- domain assumption Returns are invariant under rescaling, so the return maximization condition is the same at all scales.
- domain assumption Covariance matrices at lower frequencies can be reliably estimated from a 125-day window using non-overlapping sub-samples.
- domain assumption Transaction costs are negligible.
- domain assumption The covariance matrix exists or can be replaced by a robust L1-modified version that does not materially change allocations.
Cite this review
Pith. "Pith review of Multiscale Markowitz." pith.science (2026). https://pith.science/paper/K3PVF5UF
@misc{pith2026241113792,
author = {Pith},
title = {Pith review of: Multiscale Markowitz},
year = {2026},
howpublished = {\url{https://pith.science/paper/K3PVF5UF}},
note = {Machine review of arXiv:2411.13792}
}
abstract
Traditional Markowitz portfolio optimization constrains daily portfolio variance to a target value, optimising returns, Sharpe or variance within this constraint. However, this approach overlooks the relationship between variance at different time scales, typically described by $\sigma(\Delta t) \propto (\Delta t)^{H}$ where $H$ is the Hurst exponent, most of the time assumed to be \(\frac{1}{2}\). This paper introduces a multifrequency optimization framework that allows investors to specify target portfolio variance across a range of frequencies, characterized by a target Hurst exponent $H_{target}$, or optimize the portfolio at multiple time scales. By incorporating this scaling behavior, we enable a more nuanced and comprehensive risk management strategy that aligns with investor preferences at various time scales. This approach effectively manages portfolio risk across multiple frequencies and adapts to different market conditions, providing a robust tool for dynamic asset allocation. This overcomes some of the traditional limitations of Markowitz, when it comes to dealing with crashes, regime changes, volatility clustering or multifractality in markets. We illustrate this concept with a toy example and discuss the practical implementation for assets with varying scaling behaviors.
Reference graph
Works this paper leans on
-
[1]
Benoit B. Mandelbrot and John W. Van Ness, ”Fractional Brownia n motions, fractional noises and applications,” SIAM Review, vol. 10, no. 4, pp. 422–437, 1968
work page 1968
-
[2]
Jim Gatheral, Thibault Jaisson, and Mathieu Rosenbaum, ”Volatility is rough,” Quantitative Fi- nance, vol. 18, no. 6, pp. 933–949, 2018
work page 2018
-
[3]
Laurent E. Calvet and Adlai J. Fisher, ”Multifractality in asset re turns: Theory and evidence,” Review of Economics and Statistics , vol. 84, no. 3, pp. 381–406, 2002
work page 2002
-
[4]
Steven L. Heston, ”A closed-form solution for options with stoc hastic volatility with applications to bond and currency options,” The Review of Financial Studies , vol. 6, no. 2, pp. 327–343, 1993
work page 1993
-
[5]
Gregory Berman and Lawrence Hochberg, ”Multiscale modeling of financial time series and portfolio optimization,” Journal of Investment Strategies , vol. 1, no. 2, pp. 45–67, 2008
work page 2008
-
[6]
Jean-Fran¸ cois Muzy, J´ erˆ ome Delour, and Emmanuel Bacry, ”Modelling fluctuations of financial time series: from cascade process to stochastic volatility model,” The European Physical Journal B-Condensed Matter and Complex Systems , vol. 17, no. 3, pp. 537–548, 2000
work page 2000
-
[7]
Andrew Ang and Allan Timmermann, ”Regime Changes and Financial M arkets,” Annual Review of Financial Economics , vol. 4, no. 1, pp. 313–337, 2012
work page 2012
-
[8]
Edgar E. Peters, Fractal Market Analysis: Applying Chaos Theory to Investme nt and Economics , John Wiley & Sons, 1994. 9
work page 1994
Show all 15 references
-
[9]
Rama Cont and Jean-Philippe Bouchaud, ”Herd behavior and aggr egate fluctuations in financial markets,” Macroeconomic Dynamics, vol. 4, no. 2, pp. 170–196, 2000
2000
-
[10]
Bianchi, Michael E
Robert J. Bianchi, Michael E. Drew, and Jesse H. Fan, ”Multisca le Hedge Fund Performance Per- sistence: Evidence from Wavelet Analysis,” Journal of Financial Econometrics , vol. 13, no. 1, pp. 50–78, 2015
2015
-
[11]
H. E. Hurst, ”Long-term storage capacity of reservoirs,” Transactions of the American Society of Civil Engineers , vol. 116, no. 1, pp. 770–799, 1951
1951
-
[12]
Benoit Mandelbrot, ”The variation of certain speculative prices ,” The Journal of Business , vol. 36, no. 4, pp. 394–419, 1963
1963
-
[13]
Rama Cont, ”Empirical properties of asset returns: stylized f acts and statistical issues,” Quantitative Finance, vol. 1, no. 2, pp. 223–236, 2001
2001
-
[14]
Yakov Amihud and Haim Mendelson, ”Asset pricing and the bid-ask spread,” Journal of Financial Economics, vol. 17, no. 2, pp. 223–249, 1986
1986
-
[15]
epps1.png
Thomas W. Epps, ”Comovements in stock prices in the very shor t run,” Journal of the American Statistical Association, vol. 74, no. 366a, pp. 291–298, 1979. 10 This figure "epps1.png" is available in "png" format from: http://arxiv.org/ps/2411.13792v1
1979 arXiv
Reviewed August 12, 2026 · model on record in the stance chip above.
Discussion (0). Continue with ORCID to comment.