Pith. sign in

REVIEW 2 major objections 5 minor 35 references

Quantum kernel ridge regression does not beat classical baselines for stock return prediction once budgets and universes are fairly matched.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-01 10:34 UTC pith:43IQWD5K

load-bearing objection A rare controlled null result for quantum kernels in finance, built with the right controls; the main weakness is an asserted but undemonstrated power claim. the 2 major comments →

arxiv 2607.20168 v1 pith:43IQWD5K submitted 2026-07-22 q-fin.PR quant-phstat.ML

Quantum Kernels and the Cross-Section of Stock Returns: Anatomy of a Vanishing Advantage

Junchi Shen This is my paper
classification q-fin.PR quant-phstat.ML
keywords quantum machine learningkernel methodscross-section of stock returnssurvivorship biasempirical asset pricingChina A-sharesquantum kernel ridge regressionwalk-forward validation
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper asks whether quantum kernels improve cross-sectional stock return prediction, and answers no for the Chinese A-share market. Running 170 walk-forward windows from 2012 to 2025 on a point-in-time universe, with the same subsample, solver, and tuning budget applied to a quantum fidelity kernel, a projected quantum kernel, and a classical radial-basis control, the author finds the fidelity kernel statistically indistinguishable from its classical control (ΔIC = +0.005, p = 0.42), and no model difference survives family-wise correction, with penalized linear regressions topping every ranking. The paper then shows how an apparent quantum advantage can be manufactured unintentionally: a 60-window evaluation on a universe screened with full-sample hindsight makes the same quantum kernel look dominant on stability metrics and significantly better than neural baselines. It also reports that interaction characteristics from the anomalies literature help no model and actively harm the quantum kernel, and that bandwidth tuning finds an interior optimum with generalization concentrated in low-order, classically accessible terms. The contribution is a protocol standard for future claims of quantum advantage in finance: kernel-swap controls, budget equalization, point-in-time data, and enough windows to resolve the claimed effect size.

Core claim

On the paper's own terms, the central discovery is that a quantum fidelity kernel, trained on an eight-qubit entangling feature map and compared against a classical RBF kernel under a kernel-swap control, is statistically indistinguishable for 20-day-ahead return prediction: over 170 walk-forward windows on a point-in-time A-share universe, the quantum kernel's mean rank IC is 0.0254 versus 0.0208 for the RBF control (Δ = +0.005, p = 0.42), and no pairwise difference among the eleven models remains significant after family-wise correction. When the training budget is equalized with a low-rank kernel approximation that lets the quantum kernel see all roughly 38,000 observations per window, th

What carries the argument

The load-bearing mechanism is the paper's kernel-swap control: the quantum fidelity kernel, the projected quantum kernel, and a classical radial-basis kernel share the same training subsample (1,536 observations), the same ridge solver, and the same per-window hyperparameter grid, so that any performance difference is attributable to kernel geometry alone. The companion two-by-two design crosses kernel type with training budget by extending the quantum kernel to the full window via a low-rank kernel approximation, which removes the cubic-scaling computational constraint and thereby separates the kernel's statistical value from its computational cost. The walk-forward protocol itself is also

Load-bearing premise

The conclusion of no quantum advantage assumes those 170 test windows carry enough independent information to rule out a real but modest predictive edge, and no formal power analysis is given for that assumption; if the overlapping 20-day return series are strongly autocorrelated, the effective sample size could be far smaller than 170.

What would settle it

A replication on the same point-in-time universe and matched budgets in which a quantum kernel beats the classical control by at least 0.015 in mean rank IC, with 170 windows and a family-wise adjusted p-value below 0.05, would overturn the paper's central claim. A cheaper check: compute the effective number of independent windows from the autocorrelation of the quantum-minus-classical IC series; if it is far below 170, the null conclusion itself cannot be taken at face value.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • If the paper is right, empirical claims of quantum advantage in finance need kernel-swap controls; changing only the kernel while holding the pipeline fixed is the way to isolate what quantum geometry contributes.
  • Budget equalization matters: the apparent quantum loss to ridge disappears once the quantum kernel is extended to the full training set, and the apparent sample-efficiency advantage also disappears—quantum kernels gain from data as classical models do.
  • The results imply that the heavily screened, approximately linear A-share cross-section is a hostile arena for quantum kernels; penalized linear regressions, with or without explicit interactions, sit atop every point-estimate ranking.
  • The only quantum wins that survive in either study are against neural-network baselines, the weakest models at this data scale, so comparisons against modest networks are unlikely to establish quantum advantage in short samples.
  • A 60-window design cannot resolve effect sizes of ±0.015 in IC or even diagnose which design ingredient manufactured an apparent edge, so future studies need window counts sufficient for the claimed effect size.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • A direct consequence the paper leaves implicit is that the same diagnostics could be used to pre-screen data sets before running quantum kernels: if the geometric difference between quantum and classical kernels does not correlate with out-of-sample gains, the label is likely living in a classically reachable subspace and quantum resources are wasted.
  • The paper's difference-in-differences null result (p = 0.71) leaves open that part of the static-universe flattery may come from sample period or baseline weakness rather than survivorship alone; future work could separate survival per se from trading delisting by building a real-time tradeable universe that includes stocks that later delist.
  • The protocol standards proposed here could be applied to other markets where quantum kernels have been suggested, such as short-history or post-break asset classes; the paper's own evidence suggests those are natural test beds but only if the same controls travel.
  • The bandwidth finding suggests a testable prescription: choose bandwidth by validation IC on an interior grid, and treat pile-up at the grid endpoint as a warning that the classical or near-classical limit is doing the work rather than the quantum structure.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

2 major / 5 minor

Summary. The paper reports a controlled walk-forward horse race on Chinese A-shares (170 windows, 2012–2025) in which a quantum fidelity kernel, a projected quantum kernel, and a classical RBF control share identical training subsamples, solver, and tuning budgets. It finds no quantum advantage: the fidelity kernel is statistically indistinguishable from the RBF control (ΔIC=+0.005, p=0.42); after Holm correction no pairwise difference among eleven models survives; and a 2×2 budget-equalized design shows the apparent quantum deficit to ridge is a training-budget artifact. The paper then reproduces a spurious quantum 'success' using a 60-window evaluation on a hindsight-screened universe, and decomposes it into universe construction, insufficient power/multiplicity, and weak neural baselines. Interaction characteristics do not help, bandwidth diagnostics locate an interior optimum, and the geometric difference does not predict gains. The conclusion is that no quantum advantage exists for cross-sectional return prediction in this market, and the paper proposes protocol standards for future QML-finance claims.

Significance. The contribution is a carefully executed negative result. The design is unusually good: point-in-time universe, kernel-swap controls, budget-equalized 2×2 with Nyström extension, family-wise correction, and honest diagnostics (including a difference-in-differences test that fails to attribute the artifact to universe construction alone). If the result holds, it provides strong empirical support for the theoretical caution about quantum kernels on generic classical data and offers a reusable protocol for evaluating whether a 'quantum advantage' is real. Code and intermediate data are promised, which supports reproducibility, though a permanent public repository would be stronger. The main weaknesses are the unsupported power claim and an incomplete full-budget control cell; both are fixable and should be addressed before publication.

major comments (2)
  1. [§10, §5] The categorical conclusion 'no quantum advantage exists' is stronger than the evidence supports without a formal power analysis. The main comparison is ΔIC=+0.005 with p=0.42 over 170 windows, implying an approximate SE of 0.0062. Detecting a practically meaningful advantage of ΔIC=0.015 with 80% power at α=0.05 would require about 223 independent windows; 170 windows give only ~70% power under independence. Because the IC series may be autocorrelated through persistent factor regimes, the effective sample could be materially smaller. Section 10's assertion of 'statistical power to resolve the effect sizes at stake' is not demonstrated. Please add an autocorrelation-robust effective sample size and power curve, or an equivalence test with a pre-specified boundary. As written, the data establish 'no evidence of advantage' rather than 'no advantage exists.'
  2. [§5.1, Table 4] The full-budget cell of the 2×2 is not a clean kernel-swap comparison. The KRR–RBF control row has '—' at N≈37,800, so at full budget the comparison is between the Nyström QKRR and linear ridge, not between quantum and classical RBF kernels. Moreover, the Nyström QKRR uses 'each window's already-selected bandwidth' from the 1,536-observation tuning, rather than re-tuning on the full sample; the optimal bandwidth can shift with sample size. The conclusion that quantum kernels 'matching, but never beating, equal-budget linear models' at full scale therefore rests on a cell in which the quantum model may be disadvantaged by both a missing classical-kernel control and a reused hyperparameter. Please add a Nyström RBF control at the full budget and re-tune the bandwidth, or justify why the small-sample optimum is appropriate at the full sample.
minor comments (5)
  1. [§6.2] Internal inconsistency: the text says Holm-adjusting the diagnostic study's eight-pair family 'leaves no comparison significant (smallest adjusted p=0.19)', then states 'across both studies, the only quantum win that survives family-wise correction anywhere is the comparison against the MLP'. These cannot both hold if the MLP comparison is in that family; if it is in a different family, define it. Please correct the multiplicity reporting and the 'weak deep baselines' paragraph accordingly.
  2. [§5] Phrases like 'ties its RBF control' and 'ties the three-layer network' overstate what a failure to reject means. The paper should say 'statistically indistinguishable from' and report confidence intervals for the main ΔIC, especially because the paper's contribution is a null result.
  3. [§5.1] 'Both model classes gain comparably from data (quantum +0.018, p=0.07; linear +0.014, p=0.06)' describes insignificant differences as 'gains comparably'. Better: 'point estimates suggest comparable gains, but neither is statistically significant'.
  4. [§6.2] 'Halves the quantum kernel's IC' from 0.0512 to 0.0184 is a 64% reduction, not a halving; please correct.
  5. [Data and code availability] For a paper whose contribution is reproducibility, a permanent public repository would be stronger than 'available from the author on request'.

Circularity Check

0 steps flagged

No circularity: controlled out-of-sample experiment with independent benchmarks; the unpowered-null caveat is a statistical limitation, not a circular step.

full rationale

The paper's central quantitative claims are generated by a walk-forward protocol in which model hyperparameters are selected on a temporal validation split within each window and evaluation is out-of-sample on future rebalancing dates. The kernel-swap control (Section 4.3) holds training subsamples, solver, and tuning budget fixed across the fidelity kernel, projected kernel, and RBF control, so the reported IC differences are not fixed by construction. The top-8 characteristic set is selected by in-window rank IC and shared verbatim by classical and quantum models, so no fitted parameter is later relabeled as a prediction. The bandwidth-tuning analysis, widened-grid diagnostic, geometric-difference correlation, and interaction experiments are all measured against realized out-of-sample IC or used as falsifiable diagnostics; none is defined in terms of the paper's conclusion. Citations are to external literature and no load-bearing argument rests on a self-citation by the author. The reader's concern about missing formal power analysis and possible autocorrelation in the 170-window IC series is a legitimate statistical validity caveat, but accepting a null after a controlled experiment is not circularity. No step in the paper reduces a prediction to its own input by construction.

Axiom & Free-Parameter Ledger

0 free parameters · 4 axioms · 0 invented entities

The central claim is an empirical null result; it rests on standard statistical and domain assumptions rather than on fitted free parameters. The quantum kernels and classical baselines are taken from prior literature, and no new theoretical entity is introduced.

axioms (4)
  • standard math Kernel ridge regression and the representer theorem apply to the quantum fidelity and projected kernels used.
    Section 4.3 solves α=(K+αI)^{-1}y; this is the standard KRR solution.
  • domain assumption Exact statevector simulation of the 8-qubit, 2-repetition IQP circuit faithfully represents the noiseless quantum kernel.
    Section 4.2 computes all states with PennyLane; the paper argues hardware noise could only subtract from this ceiling.
  • domain assumption The 170 walk-forward windows provide approximately independent observations for the paired t-tests and Wilcoxon tests.
    Section 4.1 uses rebalancing every 20 trading days; consecutive windows are non-overlapping, but no explicit test for autocorrelation in cross-sectional IC series is reported.
  • domain assumption The point-in-time universe and announcement-date-aligned characteristic construction are correctly implemented from raw data.
    Section 3.1-3.2 defines the rules; any error would affect all models but could bias the kernel comparison if quantum kernels are more sensitive to feature quality.

pith-pipeline@v1.3.0-alltime-deepseek · 11370 in / 13840 out tokens · 134556 ms · 2026-08-01T10:34:51.295623+00:00 · methodology

0 comments
read the original abstract

Do quantum kernels improve cross-sectional stock return prediction? We run a controlled horse race on the Chinese A-share market in which a quantum fidelity kernel, a projected quantum kernel, and a classical RBF control share identical training subsamples, solver, and tuning budgets, so that only the kernel is exchanged. On the main evaluation -- a point-in-time universe and 170 walk-forward windows (2012-2025) -- no quantum advantage exists: the fidelity kernel is indistinguishable from its RBF control ($\Delta$IC $=+0.005$, $p=0.42$), and a $2\times2$ design crossing kernel type with training budget (a Nystrom extension to the full ~38,000-observation windows) shows quantum kernels matching, but never beating, equal-budget linear models; after family-wise correction no pairwise difference among eleven models is significant, with point estimates favoring penalized linear regressions throughout. We then document how the opposite conclusion arises: a 60-window evaluation on a universe screened with full-sample information makes the same quantum kernel appear dominant on stability criteria and significantly better than neural baselines. Interaction characteristics from the anomalies literature help nothing, quantum or classical; a widened bandwidth grid reveals an interior optimum rather than the near-classical endpoint a coarse grid suggests; and the geometric difference, while large throughout ($g \gg 1$), does not predict out-of-sample gains ($\rho=-0.20$). We propose protocol standards -- kernel-swap controls, budget-equalized comparisons, point-in-time universes, and multiplicity-robust inference -- for empirical claims of quantum advantage in finance.

Figures

Figures reproduced from arXiv: 2607.20168 by Junchi Shen.

Figure 1
Figure 1. Figure 1: Main evaluation, 170 windows. Top: 12-window rolling mean out-of-sample rank [PITH_FULL_IMAGE:figures/full_fig_p009_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Short-sample study, 60 windows. Top: regularized geometric difference [PITH_FULL_IMAGE:figures/full_fig_p013_2.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

35 extracted references · 2 linked inside Pith

  1. [1]

    Asness, C. S. (1997). The interaction of value and momentum strategies. Financial Analysts Journal, 53(2):29--36

    1997

  2. [2]

    S., Moskowitz, T

    Asness, C. S., Moskowitz, T. J., and Pedersen, L. H. (2013). Value and momentum everywhere. Journal of Finance, 68(3):929--985

    2013

  3. [3]

    G., Cakici, N., and Whitelaw, R

    Bali, T. G., Cakici, N., and Whitelaw, R. F. (2011). Maxing out: Stocks as lotteries and the cross-section of expected returns. Journal of Financial Economics, 99(2):427--446

    2011

  4. [4]

    G., Brown, S

    Bali, T. G., Brown, S. J., Murray, S., and Tang, Y. (2017). A lottery-demand-based explanation of the beta anomaly. Journal of Financial and Quantitative Analysis, 52(6):2369--2397

    2017

  5. [5]

    Bergholm, V., Izaac, J., Schuld, M., et al. (2018). PennyLane: Automatic differentiation of hybrid quantum-classical computations. arXiv:1811.04968

    Pith/arXiv arXiv 2018

  6. [6]

    M., and Shaydulin, R

    Canatar, A., Peters, E., Pehlevan, C., Wild, S. M., and Shaydulin, R. (2023). Bandwidth enables generalization in quantum kernel models. Transactions on Machine Learning Research

    2023

  7. [7]

    Chan, L. K. C., Jegadeesh, N., and Lakonishok, J. (1996). Momentum strategies. Journal of Finance, 51(5):1681--1713

    1996

  8. [8]

    and Guestrin, C

    Chen, T. and Guestrin, C. (2016). XGBoost: A scalable tree boosting system. In Proceedings of the 22nd ACM SIGKDD, 785--794

    2016

  9. [9]

    J., Gambella, C., Marecek, J., et al

    Egger, D. J., Gambella, C., Marecek, J., et al. (2020). Quantum computing for finance: State-of-the-art and future prospects. IEEE Transactions on Quantum Engineering, 1:3101724

    2020

  10. [10]

    Fama, E. F. and French, K. R. (2015). A five-factor asset pricing model. Journal of Financial Economics, 116(1):1--22

    2015

  11. [11]

    Francis, J., LaFond, R., Olsson, P., and Schipper, K. (2007). Information uncertainty and post-earnings-announcement-drift. Journal of Business Finance & Accounting, 34(3--4):403--433

    2007

  12. [12]

    Freyberger, J., Neuhierl, A., and Weber, M. (2020). Dissecting characteristics nonparametrically. Review of Financial Studies, 33(5):2326--2377

    2020

  13. [13]

    R., Gujarati, T

    Glick, J. R., Gujarati, T. P., C\' o rcoles, A. D., et al. (2024). Covariant quantum kernels for data with group structure. Nature Physics, 20:479--483

    2024

  14. [14]

    Gu, S., Kelly, B., and Xiu, D. (2020). Empirical asset pricing via machine learning. Review of Financial Studies, 33(5):2223--2273

    2020

  15. [15]

    D., Temme, K., et al

    Havl\' i c ek, V., C\' o rcoles, A. D., Temme, K., et al. (2019). Supervised learning with quantum-enhanced feature spaces. Nature, 567:209--212

    2019

  16. [16]

    Herman, D., Googin, C., Liu, X., et al. (2023). Quantum computing for finance. Nature Reviews Physics, 5:450--465

    2023

  17. [17]

    Hoerl, A. E. and Kennard, R. W. (1970). Ridge regression: Biased estimation for nonorthogonal problems. Technometrics, 12(1):55--67

    1970

  18. [18]

    Huang, H.-Y., Broughton, M., Mohseni, M., et al. (2021). Power of data in quantum machine learning. Nature Communications, 12:2631

    2021

  19. [19]

    Kozak, S., Nagel, S., and Santosh, S. (2020). Shrinking the cross-section. Journal of Financial Economics, 135(2):271--292

    2020

  20. [20]

    u bler, J. M., Buchholz, S., and Sch\

    K\" u bler, J. M., Buchholz, S., and Sch\" o lkopf, B. (2021). The inductive bias of quantum kernels. Advances in Neural Information Processing Systems, 34

    2021

  21. [21]

    Lakonishok, J., Shleifer, A., and Vishny, R. W. (1994). Contrarian investment, extrapolation, and risk. Journal of Finance, 49(5):1541--1578

    1994

  22. [22]

    Leippold, M., Wang, Q., and Zhou, W. (2022). Machine learning in the Chinese stock market. Journal of Financial Economics, 145(2):64--82

    2022

  23. [23]

    F., and Yuan, Y

    Liu, J., Stambaugh, R. F., and Yuan, Y. (2019). Size and value in China. Journal of Financial Economics, 134(1):48--69

    2019

  24. [24]

    V., Kong, D., and Wu, J

    Nartea, G. V., Kong, D., and Wu, J. (2017). Do extreme returns matter in emerging markets? Evidence from the Chinese stock market. Journal of Banking & Finance, 76:189--197

    2017

  25. [25]

    Novy-Marx, R. (2013). The other side of value: The gross profitability premium. Journal of Financial Economics, 108(1):1--28

    2013

  26. [26]

    Or\' u s, R., Mugel, S., and Lizaso, E. (2019). Quantum computing for finance: Overview and prospects. Reviews in Physics, 4:100028

    2019

  27. [27]

    Piotroski, J. D. (2000). Value investing: The use of historical financial statement information to separate winners from losers. Journal of Accounting Research, 38:1--41

    2000

  28. [28]

    Schuld, M. (2021). Supervised quantum machine learning models are kernel methods. arXiv:2101.11020

    Pith/arXiv arXiv 2021

  29. [29]

    and Killoran, N

    Schuld, M. and Killoran, N. (2019). Quantum machine learning in feature Hilbert spaces. Physical Review Letters, 122:040504

    2019

  30. [30]

    and Wild, S

    Shaydulin, R. and Wild, S. M. (2022). Importance of kernel bandwidth in quantum machine learning. Physical Review A, 106:042407

    2022

  31. [31]

    Slattery, L., Shaydulin, R., Chakrabarti, S., Pistoia, M., Khairy, S., and Wild, S. M. (2023). Numerical evidence against advantage with quantum fidelity kernels on classical data. Physical Review A, 107:062417

    2023

  32. [32]

    Sloan, R. G. (1996). Do stock prices fully reflect information in accruals and cash flows about future earnings? The Accounting Review, 71(3):289--315

    1996

  33. [33]

    Thanasilp, S., Wang, S., Cerezo, M., and Holmes, Z. (2024). Exponential concentration in quantum kernel methods. Nature Communications, 15:5200

    2024

  34. [34]

    Williams, C. K. I. and Seeger, M. (2001). Using the Nystr\" o m method to speed up kernel machines. Advances in Neural Information Processing Systems, 13

    2001

  35. [35]

    Zhang, X. F. (2006). Information uncertainty and stock returns. Journal of Finance, 61(1):105--137

    2006