pith. sign in

arxiv: 2606.10448 · v1 · pith:6MPVUBONnew · submitted 2026-06-09 · 💻 cs.LG · cs.AI

Mitigating Bias in Low-SNR Financial Reinforcement Learning via Quantum Representations

Zeyu Liu , Xuanzhi Feng , Sing Kwong Lai , Yuanchen Gao , Xiaoyi Pang , Hualei Zhang , Jingcai Guo , Jie Zhang
show 1 more author
Song Guo
This is my paper

Pith reviewed 2026-06-27 13:52 UTC · model grok-4.3

classification 💻 cs.LG cs.AI
keywords reinforcement learningparameterized quantum circuitSoft Actor-Criticfinancial portfolio managementlow signal-to-noise ratiobias mitigationrepresentation learningoff-policy learning
0
0 comments X

The pith

A compact bounded parameterized quantum circuit placed before the actor and critic in SAC constrains noisy state features and lifts cumulative returns by 66.89 percent over standard SAC in real portfolio tasks.

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

Financial markets present a low signal-to-noise environment that destabilizes off-policy maximum-entropy reinforcement learning. Noisy representations produce unreliable Q-value estimates whose errors are amplified by bootstrapping, creating the failure mode the authors term the Financial Entropy Trap. The paper introduces FPQC-SAC, which inserts a compact and bounded Parameterized Quantum Circuit ahead of the actor and critic networks so that feature propagation is constrained at the representation level rather than through input filtering or post-bootstrapping regularization. Trainable quantum entanglement inside the circuit is claimed to retain flexible cross-asset interactions. On real-world portfolio management tasks the method records a 66.89 percent relative gain in cumulative return over unconstrained SAC and a roughly 27 percent improvement over the strongest continuous-control deep RL baseline.

Core claim

Placing a compact and bounded Parameterized Quantum Circuit before the actor and critic networks constrains feature propagation at the representation level, thereby reducing the impact of extreme market fluctuations on Bellman target estimation while trainable quantum entanglement preserves flexible cross-asset interactions; the resulting FPQC-SAC variant yields substantially higher out-of-sample stability and a 66.89 percent relative gain in cumulative return over standard unconstrained SAC on real portfolio tasks.

What carries the argument

The compact bounded Parameterized Quantum Circuit inserted before the actor and critic networks, which performs constrained feature propagation via quantum entanglement.

If this is right

  • FPQC-SAC records a 66.89 percent relative gain in cumulative return over standard SAC.
  • The method outperforms the strongest continuous-control deep RL baseline by approximately 27 percent.
  • Out-of-sample stability improves because extreme market fluctuations exert less influence on Bellman targets.
  • Quantum entanglement inside the circuit maintains cross-asset interaction flexibility without extra regularization.
  • The architecture functions as a plug-and-play addition to existing SAC implementations.

Where Pith is reading between the lines

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

  • The same bounded-representation idea could be tested in other high-variance continuous-control domains such as robotic control under sensor noise.
  • If the PQC size scales with the number of assets, the approach might extend to larger portfolios without proportional growth in classical network capacity.
  • The method supplies an existence proof that representation-level quantum constraints can substitute for explicit noise-filtering layers in RL.
  • Future comparisons could measure whether classical bounded activations achieve similar gains or whether the quantum entanglement term is necessary.

Load-bearing premise

Inserting a compact bounded PQC will constrain feature propagation enough to limit extreme fluctuation effects on Bellman targets without discarding useful cross-asset information.

What would settle it

An ablation that removes the PQC or removes its boundedness on the same real-world portfolio datasets and measures whether cumulative return and stability revert to or fall below the levels of unconstrained SAC.

Figures

Figures reproduced from arXiv: 2606.10448 by Hualei Zhang, Jie Zhang, Jingcai Guo, Sing Kwong Lai, Song Guo, Xiaoyi Pang, Xuanzhi Feng, Yuanchen Gao, Zeyu Liu.

Figure 1
Figure 1. Figure 1: The system architecture and end-to-end workflow of FPQC-SAC. By forcing high-variance [PITH_FULL_IMAGE:figures/full_fig_p005_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Representation quality for FPQC-SAC, RFF/Fourier features, Wavelet filtering, and Kalman [PITH_FULL_IMAGE:figures/full_fig_p015_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Representation quality for the classical bottleneck controls in Table 1. Lower Total Latent [PITH_FULL_IMAGE:figures/full_fig_p015_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Extended out-of-sample median account value trajectories from 2019 to 2023. The shaded [PITH_FULL_IMAGE:figures/full_fig_p021_4.png] view at source ↗
read the original abstract

The financial market is a typical low signal-to-noise ratio (SNR) setting, which often destabilizes off-policy maximum-entropy methods like Soft Actor-Critic (SAC). Specifically, noisy state representations may produce unreliable Q-value estimates, and bootstrapping amplifies these errors, forming a failure mode we call the "Financial Entropy Trap". In this paper, we propose FPQC-SAC, an efficient and plug-and-play SAC variant that places a compact and bounded Parameterized Quantum Circuit (PQC) before the actor and critic networks to constrain feature propagation at the representation level, rather than filtering raw inputs or regularizing Q-values after bootstrapping. Notably, FPQC-SAC reduces the impact of extreme market fluctuations on Bellman target estimation, while trainable quantum entanglement preserves flexible cross-asset interactions. Empirical evaluations on real-world portfolio management tasks demonstrate that FPQC-SAC substantially enhances out-of-sample stability and cumulative returns by achieving a 66.89% relative gain in cumulative return over standard unconstrained SAC and outperforms the best continuous-control deep reinforcement learning baseline by approximately 27%. Open-source code is available at https://github.com/ZeyuLIU-UST/FPQC-SAC-main.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 0 minor

Summary. The paper proposes FPQC-SAC, a variant of Soft Actor-Critic (SAC) for low-SNR financial portfolio management tasks. It inserts a compact, bounded Parameterized Quantum Circuit (PQC) before the actor and critic networks to constrain feature propagation at the representation level, thereby mitigating the 'Financial Entropy Trap' (error amplification in Bellman targets due to noisy states). The central empirical claim is a 66.89% relative gain in cumulative return over unconstrained SAC and an approximately 27% improvement over the best continuous-control DRL baseline, with improved out-of-sample stability attributed to trainable quantum entanglement preserving cross-asset interactions while bounding extreme fluctuations.

Significance. If the mechanism is shown to be quantum-specific rather than a generic effect of bounded representations, the approach could provide a new regularization strategy for off-policy RL in noisy financial settings. The open-source code link is a positive factor for reproducibility, but the absence of implementation details, ablations, error bars, and dataset descriptions in the current manuscript prevents assessment of whether the reported gains support the claimed quantum advantage.

major comments (2)
  1. [Abstract] Abstract and method description: the central claim attributes the 66.89% gain and reduced Bellman error amplification specifically to the PQC's quantum entanglement providing a qualitatively different regularization effect than classical bounded layers. No ablation is presented against equivalent classical constrained networks (e.g., scaled tanh MLPs with matched parameter count and bounded output), so it is impossible to determine whether the quantum representation is load-bearing for the out-of-sample stability improvements.
  2. [Abstract] Abstract: the reported performance numbers are presented without implementation details, error bars, dataset descriptions, ablation studies, or statistical significance tests. This makes it impossible to evaluate whether the gains are robust or support the claim that the PQC constrains feature propagation in a manner that reliably reduces extreme market fluctuation impact on Q-targets.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback. We address the two major comments below and will revise the manuscript to incorporate additional ablations and experimental details.

read point-by-point responses

  1. Referee: [Abstract] Abstract and method description: the central claim attributes the 66.89% gain and reduced Bellman error amplification specifically to the PQC's quantum entanglement providing a qualitatively different regularization effect than classical bounded layers. No ablation is presented against equivalent classical constrained networks (e.g., scaled tanh MLPs with matched parameter count and bounded output), so it is impossible to determine whether the quantum representation is load-bearing for the out-of-sample stability improvements.

    Authors: We agree that the current manuscript lacks an explicit ablation against classical bounded representations with matched parameter counts. Although the PQC is motivated by the ability of trainable quantum entanglement to preserve cross-asset interactions while enforcing bounds (a mechanism not identically replicable by classical tanh layers), a direct comparison is needed to isolate the quantum contribution. We will add this ablation study to the revised manuscript. revision: yes

  2. Referee: [Abstract] Abstract: the reported performance numbers are presented without implementation details, error bars, dataset descriptions, ablation studies, or statistical significance tests. This makes it impossible to evaluate whether the gains are robust or support the claim that the PQC constrains feature propagation in a manner that reliably reduces extreme market fluctuation impact on Q-targets.

    Authors: The open-source repository supplies the implementation, yet we acknowledge that the manuscript text itself should contain dataset descriptions, error bars across runs, and statistical tests. We will expand the experimental section accordingly to report means and standard deviations, dataset details, and significance testing for the reported gains. revision: yes

Circularity Check

0 steps flagged

No derivation chain present; results are direct empirical measurements.

full rationale

The paper proposes FPQC-SAC by inserting a bounded PQC layer before actor/critic networks to constrain representations in low-SNR financial RL, then reports measured performance gains (66.89% relative cumulative return improvement) on real-world portfolio tasks. No first-principles derivation, uniqueness theorem, ansatz, or prediction step is claimed or shown; the central mechanism is introduced as a design choice and evaluated empirically without reduction to fitted inputs or self-referential definitions. No self-citations or load-bearing external results are invoked in the abstract or method description. This is a standard empirical ML proposal with no circularity in any derivation chain.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 1 invented entities

The central claim rests on the unverified effectiveness of the quantum circuit layer for representation constraint in low-SNR settings; the abstract supplies no independent evidence or derivation for this mechanism beyond the empirical claim.

free parameters (1)
  • PQC parameters
    Trainable parameters inside the parameterized quantum circuit are optimized jointly with the actor and critic.
axioms (1)
  • domain assumption Standard off-policy maximum-entropy RL assumptions continue to hold when a quantum circuit is inserted at the representation stage.
    The method is described as a plug-and-play variant of SAC.
invented entities (1)
  • Financial Entropy Trap no independent evidence
    purpose: To label the failure mode caused by noisy state representations and bootstrapping in financial RL.
    Introduced in the abstract to describe the problem the method targets.

pith-pipeline@v0.9.1-grok · 5767 in / 1215 out tokens · 19933 ms · 2026-06-27T13:52:45.713240+00:00 · methodology

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

63 extracted references · 25 canonical work pages

  1. [1]

    S. J. Abdulkadir and S.-P. Yong. Unscented kalman filter for noisy multivariate financial time-series data. InMulti-disciplinary Trends in Artificial Intelligence, volume 8271 ofLecture Notes in Computer Science, pages 87–96. Springer, 2013. doi: 10.1007/978-3-642-44949-9_9

    work page doi:10.1007/978-3-642-44949-9_9 2013

  2. [2]

    Acuto, P

    A. Acuto, P. Barillà, L. Bozzolo, M. Conterno, M. Pavese, and A. Policicchio. Variational quantum soft actor-critic for robotic arm control.arXiv preprint arXiv:2212.11681, 2022

    arXiv 2022

  3. [3]

    R. M. Alrumaih and M. A. Al-Fawzan. Time series forecasting using wavelet denoising: An application to saudi stock index.Journal of King Saud University: Engineering Sciences, 14(2): 221–233, 2002. doi: 10.1016/S1018-3639(18)30755-4

    work page doi:10.1016/s1018-3639(18)30755-4 2002

  4. [4]

    J. L. Ba, J. R. Kiros, and G. E. Hinton. Layer normalization.arXiv preprint arXiv:1607.06450, 2016. 11

    Pith/arXiv arXiv 2016

  5. [5]

    D. H. Bailey, J. M. Borwein, M. López de Prado, and Q. J. Zhu. The probability of backtest overfitting.Journal of Computational Finance, 20(4):39–69, 2017. doi: 10.21314/JCF.2016.322

    work page doi:10.21314/jcf.2016.322 2017

  6. [6]

    W. Bao, J. Yue, and Y . Rao. A deep learning framework for financial time series using stacked autoencoders and long-short term memory.PLOS ONE, 12(7):e0180944, 2017. doi: 10.1371/journal.pone.0180944

    work page doi:10.1371/journal.pone.0180944 2017

  7. [7]

    Bhatt, D

    A. Bhatt, D. Palenicek, B. Belousov, M. Argus, A. Amiranashvili, T. Brox, and J. Pe- ters. Crossq: Batch normalization in deep reinforcement learning for greater sam- ple efficiency and simplicity. InInternational Conference on Learning Representations (ICLR), 2024. URL https://proceedings.iclr.cc/paper_files/paper/2024/hash/ f381114cf5aba4e45552869863dea...

    2024

  8. [8]

    Cameron, T

    S. Cameron, T. Cameron, A. Pretorius, and S. Roberts. Robust and scalable sde learning: A functional perspective.arXiv preprint arXiv:2110.05167, 2021

    arXiv 2021

  9. [9]

    M. C. Caro, E. Gil-Fuster, J. J. Meyer, J. Eisert, and R. Sweke. Encoding-dependent generalization bounds for parametrized quantum circuits.Quantum, 5:582, 2021. doi: 10.22331/q-2021-11-17-582

    work page doi:10.22331/q-2021-11-17-582 2021

  10. [10]

    X. Chen, C. Wang, Z. Zhou, and K. Ross. Randomized ensembled double q-learning: Learning fast without a model. InInternational Conference on Learning Representations, 2021. URL https://openreview.net/forum?id=AY8zfZm0tDd

    2021

  11. [11]

    J. Choi, Z. Wang, S. Venkataramani, P. I.-J. Chuang, V . Srinivasan, and K. Gopalakrish- nan. Pact: Parameterized clipping activation for quantized neural networks.arXiv preprint arXiv:1805.06085, 2018

    Pith/arXiv arXiv 2018

  12. [12]

    Duarte, D

    V . Duarte, D. Duarte, and D. H. Silva. Machine learning for continuous-time finance.The Review of Financial Studies, 37(11):3217–3271, 2024. doi: 10.1093/rfs/hhae043

    work page doi:10.1093/rfs/hhae043 2024

  13. [13]

    Fujimoto, H

    S. Fujimoto, H. van Hoof, and D. Meger. Addressing function approximation error in actor-critic methods. InInternational conference on machine learning, pages 1587–1596. PMLR, 2018. URLhttps://proceedings.mlr.press/v80/fujimoto18a.html

    2018

  14. [14]

    S. Gu, B. Kelly, and D. Xiu. Empirical asset pricing via machine learning.The Review of Financial Studies, 33(5):2223–2273, 2020. doi: 10.1093/rfs/hhaa009

    work page doi:10.1093/rfs/hhaa009 2020

  15. [15]

    Haarnoja, A

    T. Haarnoja, A. Zhou, P. Abbeel, and S. Levine. Soft actor-critic: Off-policy maximum entropy deep reinforcement learning with a stochastic actor. InInternational conference on machine learning, pages 1861–1870. PMLR, 2018. URL https://proceedings.mlr.press/v80/ haarnoja18b.html

    2018

  16. [16]

    G. E. Hinton and R. R. Salakhutdinov. Reducing the dimensionality of data with neural networks. science, 313(5786):504–507, 2006. doi: 10.1126/science.1127647

    work page doi:10.1126/science.1127647 2006

  17. [17]

    Hiraoka, T

    T. Hiraoka, T. Imagawa, T. Hashimoto, T. Onishi, and Y . Tsuruoka. Dropout q-functions for dou- bly efficient reinforcement learning. InInternational Conference on Learning Representations (ICLR), 2022. URLhttps://openreview.net/forum?id=xCVJMsPv3RT

    2022

  18. [18]

    Hohenfeld, D

    H. Hohenfeld, D. Heimann, F. Wiebe, and F. Kirchner. Quantum deep reinforcement learning for robot navigation tasks.IEEE Access, 12:87217–87236, 2024. doi: 10.1109/ACCESS.2024. 3417808

    work page doi:10.1109/access.2024 2024

  19. [19]

    Neural Networks , volume =

    K. Hornik, M. Stinchcombe, and H. White. Multilayer feedforward networks are universal approximators.Neural networks, 2(5):359–366, 1989. doi: 10.1016/0893-6080(89)90020-8

    work page doi:10.1016/0893-6080(89)90020-8 1989

  20. [20]

    Jerbi, A

    S. Jerbi, A. Cornelissen, M. Ozols, and V . Dunjko. Quantum policy gradient algorithms. In 18th Conference on the Theory of Quantum Computation, Communication and Cryptography (TQC 2023), volume 266, pages 13:1–13:24. Schloss Dagstuhl-Leibniz-Zentrum für Informatik,

    2023

  21. [21]

    doi: 10.4230/LIPIcs.TQC.2023.13

    work page doi:10.4230/lipics.tqc.2023.13 2023

  22. [22]

    Jiang, D

    Z. Jiang, D. Xu, and J. Liang. A deep reinforcement learning framework for the financial portfolio management problem.arXiv preprint arXiv:1706.10059, 2017. 12

    Pith/arXiv arXiv 2017

  23. [23]

    J. Kong, F. Hudelist, Z. Y . Ou, and W. Zhang. Cancellation of internal quantum noise of an amplifier by quantum correlation.Physical Review Letters, 111(3):033608, July 2013. doi: 10.1103/PhysRevLett.111.033608

    work page doi:10.1103/physrevlett.111.033608 2013

  24. [24]

    Krogh and J

    A. Krogh and J. Hertz. A simple weight decay can improve generalization.Advances in neural information processing systems, 4, 1991. URL https://proceedings.neurips.cc/ paper/1991/hash/8eefcfdf5990e441f0fb6f3fad709e21-Abstract.html

    1991

  25. [25]

    Kuznetsov, P

    A. Kuznetsov, P. Shvechikov, A. Grishin, and D. Vetrov. Controlling overestimation bias with truncated mixture of continuous distributional quantile critics. InProceedings of the 37th International Conference on Machine Learning, volume 119 ofProceedings of Machine Learning Research, pages 5556–5566. PMLR, 2020. doi: 10.48550/arXiv.2005.04269

    work page doi:10.48550/arxiv.2005.04269 2020

  26. [26]

    LeCun, L

    Y . LeCun, L. Bottou, G. B. Orr, and K.-R. Müller. Efficient backprop. InNeural Networks: Tricks of the Trade, volume 7700 ofLecture Notes in Computer Science, pages 9–48. Springer,

  27. [27]

    doi: 10.1007/978-3-642-35289-8_3

    work page doi:10.1007/978-3-642-35289-8_3

  28. [28]

    T. P. Lillicrap, J. J. Hunt, A. Pritzel, N. Heess, T. Erez, Y . Tassa, D. Silver, and D. Wierstra. Continuous control with deep reinforcement learning.arXiv preprint arXiv:1509.02971, 2015

    Pith/arXiv arXiv 2015

  29. [29]

    X.-Y . Liu, H. Yang, Q. Chen, R. Zhang, L. Yang, B. Xiao, and C. D. Wang. Finrl: A deep reinforcement learning library for automated stock trading in quantitative finance. InDeep RL Workshop, NeurIPS 2020, 2020. URLhttps://arxiv.org/abs/2011.09607

    arXiv 2020

  30. [30]

    X.-Y . Liu, Z. Xia, J. Rui, J. Gao, H. Yang, M. Zhu, C. D. Wang, Z. Wang, and J. Guo. Finrl-meta: Market environments and benchmarks for data-driven financial reinforcement learning. InAdvances in Neural Information Process- ing Systems 35 (NeurIPS 2022) Datasets and Benchmarks Track, volume 35,

    2022

  31. [31]

    URL https://papers.neurips.cc/paper_files/paper/2022/hash/ 0bf54b80686d2c4dc0808c2e98d430f7-Abstract-Datasets_and_Benchmarks.html

    2022

  32. [32]

    Liu, Y .-H

    Y .-K. Liu, Y .-H. Pan, P.-F. Lu, Y .-C. Tsai, and S. Y .-C. Chen. Quantum-enhanced reinforcement learning with LSTM forecasting signals for optimizing fintech trading decisions. InProceedings of the IEEE International Conference on Quantum Computing and Engineering (QCE), pages 235–240, 2025. doi: 10.1109/QCE65121.2025.10325

    work page doi:10.1109/qce65121.2025.10325 2025

  33. [33]

    Lokossou, B

    R. Lokossou, B. S. Girma, O. K. Tonguz, and A. Biyabani. Quantum deep reinforcement learning for humanoid robot navigation task.arXiv preprint arXiv:2509.11388, 2025

    arXiv 2025

  34. [34]

    C. Lyle, Z. Zheng, K. Khetarpal, J. Martens, H. van Hasselt, R. Pascanu, and W. Dabney. Normal- ization and effective learning rates in reinforcement learning. InAdvances in Neural Information Processing Systems, volume 37, 2024. URL https://papers.nips.cc/paper_files/ paper/2024/hash/c04d37be05ba74419d2d5705972a9d64-Abstract-Conference. html

    2024

  35. [35]

    Mair and R

    P. Mair and R. Wilcox. Robust statistical methods in r using the wrs2 package.Behavior research methods, 52(2):464–488, 2020. doi: 10.3758/s13428-019-01246-w

    work page doi:10.3758/s13428-019-01246-w 2020

  36. [36]

    Martyniuk, J

    D. Martyniuk, J. Jung, and A. Paschke. Quantum architecture search: A survey. InProceedings of the IEEE International Conference on Quantum Computing and Engineering (QCE), pages 1695–1706, 2024. doi: 10.1109/QCE60285.2024.00198

    work page doi:10.1109/qce60285.2024.00198 2024

  37. [37]

    Miyato, T

    T. Miyato, T. Kataoka, M. Koyama, and Y . Yoshida. Spectral normalization for generative adversarial networks.arXiv preprint arXiv:1802.05957, 2018

    Pith/arXiv arXiv 2018

  38. [38]

    V . Mnih, A. P. Badia, M. Mirza, A. Graves, T. Lillicrap, T. Harley, D. Silver, and K. Kavukcuoglu. Asynchronous methods for deep reinforcement learning. InInternational conference on machine learning, pages 1928–1937. PMLR, 2016. URL https://proceedings.mlr.press/v48/ mniha16.html

    1928

  39. [39]

    K. Pearson. Liii. on lines and planes of closest fit to systems of points in space.The London, Edinburgh, and Dublin philosophical magazine and journal of science, 2(11):559–572, 1901. doi: 10.1080/14786440109462720. 13

    work page doi:10.1080/14786440109462720 1901

  40. [40]

    Pérez-Salinas, A

    A. Pérez-Salinas, A. Cervera-Lierta, E. Gil-Fuster, and J. I. Latorre. Data re-uploading for a universal quantum classifier.Quantum, 4:226, 2020. doi: 10.22331/q-2020-02-06-226

    work page doi:10.22331/q-2020-02-06-226 2020

  41. [41]

    and Turkay, Cagatay , title =

    N. Pippas, E. A. Ludvig, and C. Turkay. The evolution of reinforcement learning in quantitative finance: A survey.ACM Computing Surveys, 57(11):1–51, 2025. doi: 10.1145/3733714

    work page doi:10.1145/3733714 2025

  42. [42]

    Röseler, D

    P. Röseler, D. Willsch, and K. Michielsen. How to find expressible and trainable parameterized quantum circuits?arXiv preprint arXiv:2603.14451, 2026

    arXiv 2026

  43. [43]

    Schulman, F

    J. Schulman, F. Wolski, P. Dhariwal, A. Radford, and O. Klimov. Proximal policy optimization algorithms.arXiv preprint arXiv:1707.06347, 2017

    Pith/arXiv arXiv 2017

  44. [44]

    D. Song, A. M. C. Baek, and N. Kim. Forecasting stock market indices using padding-based fourier transform denoising and time series deep learning models.IEEE Access, 9:83786–83796,

  45. [45]

    doi: 10.1109/ACCESS.2021.3086537

    work page doi:10.1109/access.2021.3086537 2021

  46. [46]

    Sweke, E

    R. Sweke, E. Recio-Armengol, S. Jerbi, E. Gil-Fuster, B. Fuller, J. Eisert, and J. J. Meyer. Potential and limitations of random fourier features for dequantizing quantum machine learning. Quantum, 9:1640, 2025. doi: 10.22331/q-2025-02-20-1640

    work page doi:10.22331/q-2025-02-20-1640 2025

  47. [47]

    T.-T. Tran, T. N. Canh, N. Y . Chong, and X. HoangVan. Hybrid td3: Overestimation bias analysis and stable policy optimization for hybrid action space.arXiv preprint arXiv:2603.01302, 2026

    Pith/arXiv arXiv 2026

  48. [48]

    W. Xu, D. Evans, and Y . Qi. Feature squeezing: Detecting adversarial examples in deep neural networks. InNetwork and Distributed System Security Symposium (NDSS), 2018. doi: 10.14722/ndss.2018.23198. URL https://www.ndss-symposium.org/wp-content/ uploads/2018/02/ndss2018_03A-4_Xu_paper.pdf

    work page doi:10.14722/ndss.2018.23198 2018

  49. [49]

    Xu and V

    Y . Xu and V . Aggarwal. Accelerating quantum reinforcement learning with a quantum natural policy gradient based approach. InInternational Conference on Machine Learning, 2025. URL https://proceedings.mlr.press/v267/xu25a.html

    2025

  50. [50]

    Zhang and L

    L. Zhang and L. Hua. Major issues in high-frequency financial data analysis: A survey of solutions.Mathematics, 13(3):347, 2025. doi: 10.3390/math13030347

    work page doi:10.3390/math13030347 2025

  51. [51]

    Zhang, S

    Z. Zhang, S. Zohren, and S. Roberts. Deep reinforcement learning for trading.The Journal of Financial Data Science, 2(2):25–40, 2020. doi: 10.3905/jfds.2020.1.030. A Representation Quality Evaluation To verify whether PQC fulfills its core design goal of noise suppression and signal structure preserva- tion, we analyze the variance and V AF of the learned...

    work page doi:10.3905/jfds.2020.1.030 2020

  52. [52]

    Pre-Net (Classical):A linear projection with ReLU activation maps the 64-dimensional state to 7 dimensions (64→7)

  53. [53]

    This is followed by 2 layers of BasicEntanglerLayers 15 Table 6: State space, action space, and environment configuration

    PQC Layer (Quantum):The 7-dimensional vector is encoded into 7 qubits via Angle Embedding ( Rx rotations). This is followed by 2 layers of BasicEntanglerLayers 15 Table 6: State space, action space, and environment configuration. Parameter Value / Description State Space Dimension64 State Composition[1Cash,6Holdings,6Prices,8×6Tech Indicators,3Additional ...

  54. [54]

    Post-Net (Classical):A linear projection with ReLU activation expands the 7-dimensional quantum measurement back to a 64-dimensional feature vector (7→64)

  55. [55]

    backprop

    Actor-Critic Heads:Standard SB3 MLP networks (two layers of 256 units: [256,256] ) process the 64-dimensional feature vector to output policies and Q-values Table 7: Quantum Parameterized Quantum Circuit (PQC) settings. Parameter Configuration Number of Qubits (N) 7 Encoding Strategyqml.AngleEmbedding Variational Layers 2 layers ofqml.BasicEntanglerLayers...

    2018

  56. [56]

    Destructive Interference of Isolated Noise (Analogous to Denoising):For independent, isotropic noise lacking intrinsic correlation (whereΣ Z =σ 2I), the joint mode satisfies: (ej ±e k)⊤ΣZ(ej ±e k) = 2σ2,(23) 19 which is larger than the single-coordinate variance σ2. This indicates that unstructured random noise lacking correlation will induce more violent...

  57. [57]

    Constructive Interference of Structural Consensus (Analogous to Signal Amplification): Conversely, effective trading signals are often hidden within the joint distributions or non-linear combinations of multiple indicators (e.g., price breakouts accompanied by specific volume and volatility contractions). When there is a weak drift structure shared across...

  58. [58]

    (27) The output noise power is measured by the trace of the covariance across the representation dimen- sions: Tr(Cov[ϕ(X+ξ)]) = kX j=1 Var(ϕj(X+ξ)).(28) By the bounded Pauli readout established in Proposition 2, each PQC feature dimension satisfies ϕj(X+ξ)∈[−1,1]. Therefore, by Popoviciu’s inequality, Var(ϕj(X+ξ))≤ (1−(−1)) 2 4 = 1.(29) Consequently, Tr(...

    2019

  59. [59]

    CR= VT −V 0 V0 ×100%(33) 22

    Cumulative Return (CR%)Cumulative Return measures the overall percentage increase of the portfolio value from the beginning to the end of the trading period. CR= VT −V 0 V0 ×100%(33) 22

  60. [60]

    Assuming 252 trading days in a year, it is calculated as: AR= (1 +CR) 252 T −1 ×100%(34)

    Annualized Return (AR%)Annualized Return represents the geometric average amount of money earned by the investment each year over a given time period. Assuming 252 trading days in a year, it is calculated as: AR= (1 +CR) 252 T −1 ×100%(34)

  61. [61]

    Sharpe Ratio (SR)The Sharpe Ratio evaluates the risk-adjusted return of the portfolio by penalizing excessive volatility. It is the ratio of the annualized expected excess return to the annualized standard deviation of returns: SR= √ 252·E[R t −R f] σ(Rt) (35) where E[Rt −R f] is the mean of the daily excess returns, and σ(Rt) is the standard deviation of...

  62. [62]

    It only penalizes negative returns: Sortino= √ 252·E[R t −R f] σd (36) whereσ d = p E[min(0, Rt −R f)2]is the target downside deviation

    Sortino Ratio (Sortino)The Sortino Ratio is a variation of the Sharpe ratio that differentiates harmful volatility from total overall volatility by using the asset’s downside deviation. It only penalizes negative returns: Sortino= √ 252·E[R t −R f] σd (36) whereσ d = p E[min(0, Rt −R f)2]is the target downside deviation

  63. [63]

    Calmar Ratio (Calmar)The Calmar Ratio measures the annualized return relative to the Maximum Drawdown (MDD), serving as an indicator of return relative to tail-end downside risk. The MDD represents the largest peak-to-trough drop in the portfolio value: MDD= max τ∈(0,T) maxt∈(0,τ) Vt −V τ maxt∈(0,τ) Vt (37) Calmar= AR MDD (38) H Limitations and Future Wor...