A trader should care more about sign preservation here than the full multifractal machinery. On synthetic data, the plain q=2 signed model records a 10-period 99% VaR of 5.925, versus 6.193 for the fully multifractal absolute-value model. ES is 6.095 against 6.433. Keeping the direction of local co-movements contributes more than aggregation across fluctuation orders. Kakinaka and Umeno make the same point in the paper.
Inside the objective
Take two return series, demean them, and cumulate each into a profile. Each profile is divided into non-overlapping segments of length s. The division is repeated from the far end to retain all observations, producing 2N_s segments. Within every segment, fit a degree-2 polynomial and keep the residuals. Their product, averaged over the segment, gives a local detrended cross-covariance.
The aggregation step separates the two methods. MFDCCA takes absolute values of the residuals before multiplication, leaving every segment with a non-negative contribution. MFCCA raises the local covariance magnitude to q/2 and restores its sign. Co-moving segments lift the aggregate; counter-moving segments reduce it. Indexed by scale s and order q, this signed quantity becomes element (i,j) of the matrix used in place of covariance. The optimizer is a long-only quadratic program with budget constraint and required-return floor r_e.
The information content is plausible. Two assets may co-move over five days and decouple over a hundred, while large moves may couple differently from small ones. Absolute values retain coupling strength while throwing away direction, making a counter-moving pair look as risky as a co-moving pair. The synthetic panel gives that choice a measurable cost. MMFD has an ES of 6.433, compared with 6.095 for signed MC. At q=2, the quadratic form equals the DCCA fluctuation function of the portfolio series itself, recovering mean-variance as a scale-dependent limit.
A separate portfolio for every (q,s) pair still leaves an allocation problem, so Kakinaka and Umeno average the weights. Their empirical run assigns equal preference weights across S={5,30,55,80,105,130} and Q={1,1.5,2,2.5,3,3.5,4}. They describe equal weighting as the naive allocation in the fractal setting. MMFC is the aggregated sign-preserving portfolio, MC is its q=2-only version, and MMFD and MD are the corresponding absolute-value portfolios.
Every synthetic column
The synthetic exercise uses 50 seeded realizations, each of length 2^13. One pair is generated by a two-component ARFIMA process with known long-range cross-correlation. A Markov-switching multifractal process generates the other, forming volatility through a hierarchy of randomly renewed multipliers so small and large fluctuations scale differently. For these tests, S={10,100,1000} and Q={0.5,1,1.5,...,4.5,5}. The return constraint remains slack, leaving the budget and no-short-sale constraints binding.
Rank the models by 10-period 99% VaR, expressed as a positive loss: MV 6.677, MD 6.342, MMFD 6.193, MC 5.925, MMFC 5.923. The 97.5% ES sequence is 6.742, 6.603, 6.433, 6.095, 6.048.
Across the 50 realizations, MC and MMFC have standard deviations of 0.614 and 0.607. Their 0.002 VaR gap carries no weight. MV finishes worst and varies most among the five, with 0.722 on VaR and 0.733 on ES. Wilcoxon tests with Holm correction give the same reading for MC and MMFC: every pairwise VaR difference is significant except that comparison. ES separates every pair, including MMFC over MC at p<10^-5. Realized total returns show no difference across the five, with p>0.3 for all pairs, so lower synthetic risk does not come at the expense of return.
The ordering matters more than first place. MV ranks last on both measures. Both absolute-value portfolios fall between MV and the signed pair, which makes sign preservation the main source of the shift. Additional fluctuation orders lower ES from 6.095 to 6.048. That 0.047 gap is small beside realization standard deviations of 0.612 and 0.594, although the paired tests still report p<10^-5.
How narrow is the out-of-sample result?
The empirical test covers daily log returns for four assets: the Nikkei index, the S&P 500, WTI crude futures and gold spot. Its period runs from January 2011 to November 2023, with N=3330 common dates. The underlying series use one-minute dealer quotes. Trading days run between consecutive 17:00 Eastern-time pauses, and the paper presents the series as tradable proxies rather than instruments.
Weights are estimated from a rolling 520-day window, advanced one month at a time across 131 windows. The first ends in January 2013 and the last in November 2023. In-window sample means supply expected returns, and each resulting weight vector stays fixed through the following month.
MMFC produces the lowest average monthly drawdown and lowest 10-day 99% VaR out of sample at every r_e from 0% to 3%. Its ES is essentially level with MC, while total return exceeds MV and MC. The drawdown advantage over the MFDCCA family is clear, reaching p<10^-4 at r_e=1%. Relative to mean-variance, the sign remains favorable at every r_e, yet the difference cannot be statistically resolved by the block-bootstrap tests for VaR and ES or the paired monthly drawdown tests.
The empirical section says the sample may be too short to establish uniform statistical dominance over mean-variance. It also presents the exercise as a methodological demonstration rather than a fully specified trading strategy. The empirical claim ultimately rests on the MFDCCA comparison. In the conclusion, the authors describe unambiguous gains in negative tail risk control relative to the MFDCCA-based benchmarks, corresponding to the p<10^-4 drawdown result. This is where the abstract overstates the case.
The abstract says the criterion lowers drawdown, VaR and ES against mean-variance at every required return, in and out of sample, without sacrificing realized return. The averages support that ordering. The conclusion limits its out-of-sample list to drawdown and VaR. Reading the abstract as measured superiority over mean-variance goes beyond the concession in the body. The in-sample panel adds another complication: MFDCCA portfolios show slightly lower VaR and higher returns alongside visibly larger drawdowns. The evidence describes a trade between two fractal families rather than dominance by either family.
Scale is decisive. Across all five portfolios, the out-of-sample figure axes cover roughly 6.5% to 6.9% for monthly drawdown and 5.1% to 5.5% for both tail measures.
Every candidate falls within a 0.4-point band.
Significant in-sample findings comprise the drawdown gain over MV and the MFDCCA portfolios, at p≈0.04 for r_e=1%, plus the ES gain over MC at p≈0.01. Other pairwise differences miss significance at 5%. The in-sample panels spread farther than their out-of-sample counterparts, with monthly drawdown ranging from 7.50% to 8.50% across the five portfolios. The 131 windows advance one month while using a 520-day lookback, so neighboring windows share about 500 observations. HAC standard errors account for dependence; they cannot manufacture information.
Risk interpretation beyond q=2
Kakinaka and Umeno acknowledge the interpretation problem in both the methodology and conclusion. Once q≠2, the segment-level power transformation destroys bilinearity. The resulting quadratic form therefore ceases to equal the q-th order fluctuation function of the portfolio series. Instead, it aggregates signed pairwise dependence for a selected (q,s). The authors apply the same warning to the mean-MFDCCA criteria used as benchmarks.
Away from q=2, the objective acts as a heuristic score for pair relationships, with no guarantee of positive semidefiniteness. In the empirical run, 0.95% of MMFC matrices were non-PSD. Every case occurred at q=1, and the matrices were symmetrized and eigenvalue-clipped before optimization. The smallest scale, s=5, also lies below the recommended 20 ≤ s ≤ N/5 range for exponent estimation. Dropping s=5 leaves the results and portfolio ranking essentially unchanged, which settles that concern.
Boundary conditions
Four assets. One sample period. Long-only portfolios. Monthly rebalancing applies to a weight vector averaged over the 42 (q,s) combinations in the empirical exercise. The paper "does not explicitly account for transaction costs or liquidity constraints". It reports no Sharpe ratio anywhere because the authors argue that variance-based indicators assume conditions relaxed by the fractal framework. The rationale is defensible, though it removes the one number most readers would use for comparison with other strategies.
We did not find turnover figures or a table giving empirical values by required return. Performance must be read from the figures. Every portfolio, including MV, uses in-window sample means for expected returns. Holding that estimate constant gives clean attribution because only the risk matrix changes. The return input that destabilizes mean-variance weights remains untouched.
We could not test this ourselves. The signal depends on four dealer-quote series from histdata.com, and we do not hold price histories for those instruments. Replacing them with ETFs would test another cross-correlation structure under different trading sessions. We have previously covered covariance-replacement allocators in MINGLE's factor-graph covariance, where improved conditioning supplied the gain instead of a new signal. Here, the four-asset out-of-sample band matters more to me than the synthetic p-values.
Two results would change my view. The signed construction needs a run on thirty or more liquid futures with turnover and costs included. It also needs comparison against shrinkage covariance and a CVaR optimizer instead of sample mean-variance. Until then, the durable finding is narrower than the abstract and more useful: rectifying local detrended covariances behaves like a return tilt, while preserving their sign behaves like drawdown control.