No trader can size a 2.2 percentage point gross return lead drawn from one calendar year, 100 stocks, a three-day holding period and zero transaction costs. That narrow description is the honest version of the paper's headline result. Lopena, Buguks and co-authors report an average total return of 0.236 for GBS Roots and 0.214 for SPONGE in the lossless 100-stock 2020 test. Terminal returns produce a two-tailed Welch p of 0.023. In the same paragraph, however, the Sharpe comparison with SPONGE lands at p = 0.949.
The authors explain the parity through dispersion. GBS Roots retains a competitive risk-adjusted profile while recording less than half the cross-run standard deviation of SPONGE, ±0.293 against ±0.669. Cross-run dispersion tells us something about the sampler. It does not establish an edge in expected risk-adjusted return.
The mechanism matters more than the physics.
From residual returns to clusters
The process begins with daily dividend-adjusted CRSP prices. Market beta is estimated over 60 days against the Fama-French excess market return, after which residual returns are calculated as R minus beta times the market. The authors build a 5-day rolling correlation matrix from those residuals and clean it with Random Matrix Theory. They kill the market mode, nullify eigenvalues in the Marchenko-Pastur bulk and zero the diagonal to form an adjacency matrix. Negative entries are then thresholded to zero.
Clustering follows. Within each cluster, stocks whose cumulative residual return exceeds the cluster mean become previous winners. Those below it become previous losers. The portfolio shorts the winners and buys the losers, using equal weights and remaining dollar neutral. Positions stay open for three days with a 2% stop-win before the process rolls forward. A 252-day year therefore contains roughly 84 clustering applications.
The intended source of profit is cross-sectional mean reversion among stocks that genuinely move together. Every choice made by the sampler comes before that trade. Its job is to decide which names belong together.
Gaussian Boson Sampling is relevant because its detection probabilities are proportional to squared hafnians of submatrices, biasing the output toward dense subgraphs. The adjacency matrix goes in and subgraph samples come out. The algorithm chooses a dense sample as a cluster, removes it and repeats. GBS Boost, adapted from earlier work, selects the single densest sample during each iteration. GBS Roots, introduced by the paper, instead assembles a disjoint union of samples for as long as weighted density continues to rise.
Neither algorithm requires a cluster count K. Spectral and SPONGE, the classical comparators, do. For every window, the paper chooses K through a Tracy-Widom eigenvalue-spacing test with a permutation-calibrated threshold of alpha = 0.05.
No photonic hardware appears in the experiments. Every GBS result comes from a classical simulation in Strawberry Fields, using threshold detectors and a mean photon number equal to the square root of the matrix dimension, so 10 at N = 100. Running the clustering stage itself requires a photonic Gaussian Boson Sampler. We do not have one. Anything available for us to build would reproduce the paper's own classical baseline. We therefore did not test the paper's method, and no result of our own is quoted below.
Costs dominate a thin gross gap
The authors make their assumption explicit: "A key assumption, for simplicity, in our simulations is that the trades executed do not incur price impact and there are zero transaction costs." Zero-cost trading also removes borrow cost and any short-availability constraint.
Compare that assumption with the return lead under discussion. The portfolio is a dollar-neutral, equal-weight book covering 100 names. It is rebalanced every three days, while cluster membership is rebuilt each time from a 5-day correlation window. For the best partition by total return from each algorithm, inter-window Jaccard similarity reaches 0.163 ± 0.024 under GBS Roots and 0.321 ± 0.042 under SPONGE. Across that subset, the quantum partitions are roughly half as persistent from one window to the next.
Lower persistence sends more names across cluster boundaries. A cluster change can flip the labels regardless of what prices have done.
Section III B expects turnover to move the other way. The paper argues that GBS Roots tends to create larger clusters, whose stability should produce steadier signals and reduced turnover between previous winners and losers. No turnover figure appears anywhere in the paper. Either direction remains possible, and the cost-adjusted ranking has yet to be tested.
The authors do not claim otherwise. Their comparison still places two clustering methods beside each other when the relative gross return gap is small and relative turnover goes unmeasured. With no trading charges and no turnover measurement, the paper has not shown that a 2.2 point gross gap survives costs on a portfolio rebalanced every three days.
Density leaves the mechanism unresolved
The paper becomes more interesting here than its abstract suggests. SPONGE records the highest weighted density among the best partitions, 0.336 plus or minus 0.015, along with the lowest cluster value at 2.752. It does not lead on return. GBS Boost has a cluster value of 17.031, compared with 3.945 for GBS Roots. Its summed intra-cluster correlation is more than four times as large, yet total returns are 0.239 and 0.236. Economically, those results are essentially identical.
The authors acknowledge the mismatch. Strictly maximising graph features, they write, "does not guarantee optimal economic performance." Weighted density is described as a useful but incomplete proxy for economic performance. The discussion goes further: under uniform latent factors and idiosyncratic noise, the return-maximising weighted density is 2/3. Once two assets become too correlated, too little remains to revert.
Maximising the paper's stated objective, the max-agree correlation clustering problem, therefore does not guarantee the best economic result. The authors separately concede that existing GBS methods are not natively optimised for that objective. They continue to argue that a higher partition value should raise the likelihood that clustering algorithms produce economic returns. The source of the reported outperformance remains undemonstrated. Their proposed explanation is that GBS samples clusters that vary more dynamically and structurally across windows, potentially creating high-alpha opportunities. Possible.
The paper includes the control required by the noise objection. In the loss study, a random-clustering baseline generally underperforms, leading the authors to conclude that pure stochasticity carries no economic edge. That baseline belongs to the 50-stock loss sweep. It does not appear in the lossless 100-stock 2020 experiment that produces the 2.2 point gap. The narrower objection survives: the headline result has no random comparator, and its quantum averages rely on roughly 20 runs over one year.
The loss study reaches the same tension, which the paper flags. Without displacement, weighted density for GBS clusters jumps above 60% photon loss even as returns worsen. The sampler has collapsed to two-node clusters. Density rises while PnL falls for a purely degenerate reason.
What does volatility add?
The regime tests use N = 50 and cover 2008, 2017, 2020 and 2022. GBS methods remain profitable through the 2008 crash, then underperform during the low-volatility 2017 bull market. GBS Boost and GBS Roots, the quantum methods, deliver their highest returns in the highly dynamic 2020 regime. Their advantage fades in 2022 while they remain competitive. Average VIX rises alongside the Sharpe ratios of the quantum methods.
Those regime-level Sharpe values are shown only in Figure 5, without a table. Each of the four regimes is represented by a single one-year snapshot. The authors themselves call for multi-year tests spanning 2007-2009 and 2017-2018.
High realised volatility helping cross-sectional reversion, with low volatility hurting it, is a familiar shape.
The loss experiment does provide genuine engineering evidence. Without compensation, GBS falls behind Spectral and SPONGE once loss exceeds 70%. When coherent displacement restores mean photon number, the GBS methods beat SPONGE in 85% of cases at loss up to 40%. These are simulation results. The lossy experiments apply the small-squeezing approximation, replacing the adjacency matrix with eta times itself.
The authors report the approximation error directly. Frobenius error reaches its maximum of 0.194 at eta = 0.5. About 98% comes from the single largest squeezing parameter, r_1 of roughly 1.513. The disclosure is clear. The lossy results remain an approximation to a device.
The classical comparator that matters
The adjacency matrix contains only non-negative entries after thresholding. The paper states the consequence: for non-negative input matrices, classical hafnian simulation "is generally believed to be feasible in polynomial time." It cites recent work that samples in proportion to the squared hafnian through a Jerrum-Sinclair Markov chain.
The authors defend thresholding on clustering grounds. If negative edges remain, perfect matchings containing an even number of negative pairs can contribute positively to the hafnian. Anticorrelated subgraphs would then receive perversely higher sampling probability. Residualisation makes the decision consequential. Positive entries account for 99.53% of raw correlations, compared with only 51.36% after market factors are removed.
The defence makes sense. It also places the experiment in the setting where the quantum claim is weakest. The introduction itself identifies spoofing through non-negativity as one of two primary challenges to GBS advantage.
QIC-GBS, the paper's classical quantum-inspired proxy, finishes last on the financial measures. Its total return is 0.195 plus or minus 0.099, with a Sharpe of 1.781 plus or minus 0.893. The result initially favours the quantum methods, although the dispersion complicates it. This classical proxy has a return standard deviation three times the 0.031 recorded by GBS Roots, leaving open whether another classical stochastic clustering method could reproduce the volatile-regime edge. The authors call QIC-GBS "a pragmatic proxy for GBS dynamics" and describe its performance as "highly idiosyncratic," with dominance in 2008 and weakness in 2020.
The universe construction also deserves weight. The samples are GICS-proportional subsets chosen "based on the previous year's top-performing assets." Selection for the main 2020 run uses 2019 information, which preserves the correct temporal direction. The same prior-year rule is applied to 2008 and 2017, keeping the construction consistent. The binding restriction is that the sample covers only a 10-20% subset of the wider stock universe. The authors concede that this misses some market correlations.
A multi-year run could change my view if it charged realistic borrow costs, measured turnover and still showed GBS Roots ahead of SPONGE after costs. The Sharpe difference would also need to clear significance instead of remaining at p = 0.949. Table I puts SPONGE's mean Sortino above that of GBS Roots as well, 3.737 ± 1.282 against 3.708 ± 0.673.
The abstract already limits its claim to quantum clustering producing superior alpha in large stock universes during high-volatility periods. My objection concerns the size of the advantage in precisely that setting and the costs left uncharged. Table I supports comparable risk-adjusted performance, alongside less than half the cross-run dispersion of SPONGE, in the lossless 100-stock 2020 run. Lower dispersion across simulation seeds in a large universe is the sampler's demonstrated benefit.
We have previously examined correlation-spectrum signals that describe market structure cleanly while forecasting it only in-sample (our earlier note). This paper lands nearby. Its graph object is carefully built, while the link between the graph metric and PnL remains the missing piece the authors themselves identify.