Pith. sign in

REVIEW 4 major objections 5 minor 30 references

The paper claims that a reservoir's response contrast between open-chain and periodic-ring dynamics can isolate residual volatility information that linear models miss.

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 04:33 UTC pith:2VTSGKIK

load-bearing objection Original reservoir architecture and unusually careful OOS protocol, but the abstract's significance claim on IWM/XLP likely does not survive the paper's own multiple-comparison correction — the authors admit as much in Section 2.6 and never show adjusted results. the 4 major comments →

arxiv 2607.22491 v2 pith:2VTSGKIK submitted 2026-07-24 cs.LG

Susceptible Reservoir Architectures for Regime-Conditional Volatility Forecasting

Aliaksei Kaliutau This is my paper
classification cs.LG
keywords volatility forecastingreservoir computingstructural susceptibilitymixture of expertsquantum reservoir computingQLIKEHARQGARCH
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 introduces a reservoir-design principle, SUSA, in which the same volatility input is fed to two matched complex-valued reservoirs that differ only by one boundary edge: an open chain and a periodic ring. The signed, normalized difference between their observable responses is treated as a 'susceptibility' feature and passed to a small QLIKE-trained readout that corrects a persistent AR-Ridge anchor. The authors claim these features capture information complementary to GARCH and HARQ-style persistence models: the SUSA models beat AR in all experiments, outperform GARCH with statistically significant QLIKE improvements on two assets (IWM, XLP), and, when stacked with HARQ, improve average QLIKE by 0.0116 with a 75% win rate. A phase-conditioned mixture-of-experts variant interprets the same contrast differently in calm, onset, recovery, and persistent-stress regimes, and quantum-reservoir implementations extend the same principle to qubit dynamics, without yet establishing quantum advantage.

Core claim

The central claim is that structural susceptibility—defined operationally as the response contrast between a complex open-chain reservoir and its periodic-ring counterpart under identical input encoding, coupling, and spectral radius—is a stable out-of-sample source of nonlinear volatility information. The contrast feature Δ = O − R and its normalized form is the only input the readout sees beyond the anchor. The paper reports that this feature improves QLIKE forecasts for small-cap and consumer-staples assets (IWM, XLP) with confidence intervals entirely below zero, and that the forecasts disagree usefully with HARQ, so an equal-weight ensemble outperforms both constituents. The authors con

What carries the argument

The central object is the open-chain/periodic-ring counterfactual pair. Two reservoirs share everything—node count, input masks, random realization, coupling scale, and spectral radius—except that the ring adds a wrap edge connecting the last node to the first. The signed difference of observables (final populations, currents, boundary imbalance, entropy) is the susceptibility representation z = [Δ; normalized Δ]. A low-rank projection and bounded QLIKE readout fit the correction; in the PHASE-RC-MoE variant, four phase-specialized readouts are gated by a causal softmax. The same contrast is lifted to a quantum density-matrix reservoir in the open-system counterpart, where the wrap edge is t

Load-bearing premise

The load-bearing premise is that the difference between open-chain and periodic-ring reservoir responses is a stable, out-of-sample meaningful nonlinear feature of the volatility path; if that contrast is just noise induced by the random masks, or if IWM and XLP emerged from post hoc selection, the central claim loses its support.

What would settle it

Re-run the identical pipeline with pre-registered symbols and random permutations of the reservoir input masks: if the open-minus-ring contrast fails to beat AR-Ridge on IWM and XLP under Holm-Bonferroni correction on the full set, or if the contrast computed from shuffled masks produces the same QLIKE gains, the susceptibility interpretation is falsified. Alternatively, compute the contrast from a reservoir whose open/ring distinction is removed (e.g., two independent chains with the same reservoir) and show no predictive difference.

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

If this is right

  • If the susceptibility contrast is genuinely informative, linear volatility models and fixed reservoirs leave a small but exploitable residual that boundary-topology differences expose.
  • Stacked with HARQ, SUSA forecasts improve QLIKE and win in 75% of symbol/fold units, showing that the extracted features are partly non-redundant with standard volatility features.
  • The anchor-plus-bounded-correction design means a susceptibility branch can be attached to any reliable persistence forecast without replacing it, limiting downside risk.
  • Regime-conditioned interpretation matters: the same reservoir response maps to different corrections depending on the inferred market phase, which helps in calm versus stress transitions.
  • The quantum versions inherit the susceptibility principle but, as the paper states, do not yet demonstrate a quantum advantage; the portable part is the architectural idea, not the quantum hardware.

Where Pith is reading between the lines

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

  • The paper does not test whether the open-ring contrast is specific to the chosen reservoir geometry; a natural extension is permuting the input masks or the wrap edge and checking whether the QLIKE gains vanish.
  • The two significant assets (IWM, XLP) may reflect systematic sector-level regimes—small-cap and consumer staples—rather than a general property; a sector-stratified multi-asset study would clarify this.
  • The susceptibility representation could be computed for any fixed nonlinear feature extractor, suggesting a general 'matched-counterfactual' preprocessing that is not limited to reservoir computing.
  • Because the paper's multiple-comparison correction leaves almost all individual assets indistinguishable from noise, a pre-registered out-of-sample replication focused on IWM and XLP would be the decisive next check.

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

4 major / 5 minor

Summary. The paper introduces Susceptible Architectures (SUSA), a reservoir-computing design principle for volatility forecasting, instantiated as classical complex-valued reservoirs (SC-RC, PHASE-RC-MoE) and open-system quantum reservoirs (SC-OSQRC, PHASE-OSQRC-MoE). The core idea is to feed an input volatility path into two matched reservoirs that differ only by a boundary condition (open chain vs. periodic ring) and use their signed and normalized contrast as features, optionally conditioned on four volatility regimes. All models share an AR-Ridge anchor and learn a bounded residual correction under QLIKE. Evaluation on 16 U.S. assets uses three disjoint chronological folds with purge intervals. The abstract claims statistically significant QLIKE improvements for IWM and XLP and a stacking gain over HARQ. The paper also states that after Holm-Bonferroni correction, performance is not statistically differentiable from random noise on almost all individual stocks, yet still concludes that the core hypothesis is validated.

Significance. If the empirical claims survive rigorous multiple-comparison adjustment, the paper would provide an interesting and falsifiable demonstration that a deliberately constructed open-vs-periodic reservoir contrast can extract residual information beyond linear persistence models. The experimental design is a clear strength: chronological folds, purge intervals, out-of-fold anchor predictions, and fold-specific scalers are handled carefully, and the QLIKE readout is trained and evaluated on disjoint segments with loss-aligned objectives. The quantum implementations are correctly framed as not yet providing a quantum advantage. However, the central significance claim currently rests on two assets whose reported confidence intervals are not shown to be adjusted for the many tests performed, and the conclusion overstates the evidence. The stacking result is more robust but is secondary and also needs clarity on multiplicity.

major comments (4)
  1. [§2.6 and §3 (IWM/XLP)] The abstract's central claim of 'statistically significant QLIKE improvements for specific assets (IWM, XLP)' is not supported by the evidence as reported. Section 2.6 states that after Holm-Bonferroni adjustment, performance is 'not statistically differentiable from random noise on almost all individual stocks,' but the results section reports only unadjusted 95% CIs for IWM and XLP (-0.107 to -0.033 and -0.146 to -0.049) and does not report adjusted CIs, adjusted p-values, or which tests were in the family. If these intervals are not multiplicity-adjusted, the headline two-asset claim may be a false positive. Please report Holm-Bonferroni-adjusted intervals or p-values for all 16 symbols and all model variants, or explicitly restrict the significance claim to the unadjusted exploratory level.
  2. [§4 Conclusion] The conclusion that the 'core hypothesis' is 'validated' goes beyond the evidence presented. The paper's own results show aggregate differences from GARCH that are often statistically inconclusive and only two symbols with entirely negative unadjusted CIs. Even with adjusted significance, validating a general architectural principle on two assets from a set of 16 is not sufficient for the strength of that claim. The conclusion should be reworded to describe the results as partial evidence, not validation, unless substantially stronger adjusted evidence is provided.
  3. [§3, Table 3] The stacking result is presented with Diebold-Mariano p-values ('<0.001', '0.0046') without any indication of whether these are adjusted for the three stacking combinations shown or for the broader set of model comparisons performed elsewhere in the paper. The win share is computed over symbol/fold units, which may be dependent, and the block-bootstrap CI is mentioned in the text but not shown in Table 3. Please clarify the test family and dependence treatment, and report adjusted p-values or confidence intervals for the stack gains.
  4. [§2.3, §2.4, §4] The SUSA principle is defined operationally as the open-minus-ring contrast, but the paper does not provide a theoretical argument or diagnostic analysis showing that this contrast is a stable, out-of-sample meaningful feature rather than a noise realization of the random reservoir. The empirical tests are a legitimate first step, but the claim that 'the mathematical richness of this architecture is now established' is not supported by the evidence. I suggest either adding an ablation that varies random masks/spectral radii and shows the contrast's predictive content is not coincidental, or tempering the language.
minor comments (5)
  1. [§1] Typographical issue: 'quaestio vexata' appears in an unusual form; the standard phrase is 'vexata quaestio'.
  2. [§3, Figure 5] Figure 5 is difficult to read because of overlapping symbol labels and the large number of intervals. A table of all point estimates and CIs would improve accessibility and reproducibility.
  3. [§3, Table 3] The column header 'Stack QLIKE Mean' is ambiguous: it could be the mean QLIKE of the stack or the mean gain. Clarify the units and the definition of 'Gain'.
  4. [§5] Code and data are described as 'will be released'; for reproducibility, provide a working repository URL or code supplement at review time.
  5. [§2.6] The statement 'Failing this gate means that when controlling for data mining bias, the model’s performance is not statistically differentiable from random noise on almost all individual stocks' is a strong and important caveat. It should appear prominently in the Results section, not only in Methods, to help readers interpret the subsequent claims.

Circularity Check

0 steps flagged

No significant circularity: the susceptibility features and readout are independent of the test target, and the reported forecasts are evaluated out-of-sample.

full rationale

The derivation chain is empirical and self-contained: (i) fixed open-chain and periodic reservoirs are constructed with identical node counts, masks, couplings, and spectral radius, differing only by the wrap edge (Eq. 5); (ii) the susceptibility input is the signed and normalized open-minus-ring contrast of the corresponding state summaries (Eqs. 6-8), which depends only on the input volatility window and fixed reservoir dynamics, not on the target; (iii) a bounded readout is fitted to that fixed feature representation under QLIKE using chronological out-of-fold anchor predictions (Eq. 4); (iv) model quality is then measured by QLIKE differentials on disjoint held-out test folds. There is no equation in which a fitted constant is defined in terms of the test target or in which the claimed prediction is equivalent to a training input by construction. The readout being trained under the same QLIKE loss used for evaluation is not circular because the evaluation is on held-out test segments and the anchor correction is trained on out-of-fold predictions, as stated in Eq. (4) and Appendix A. No load-bearing self-citation appears: the cited GARCH, HAR, HARQ, and quantum-reservoir work is external, and the only author-specific component is the new SUSA feature itself, which is tested rather than assumed. The paper's own caveat in Section 2.6—that after controlling for data mining bias the model is not statistically differentiable from random noise on almost all individual stocks—is a legitimate robustness concern about the IWM/XLP significance claim, but it is not circularity. Accordingly, the central forecasting result is not a disguised fit or a renamed input.

Axiom & Free-Parameter Ledger

8 free parameters · 6 axioms · 0 invented entities

The paper's empirical claims rest on a large number of chosen hyperparameters and domain assumptions. Many configuration parameters (clip bound, validation weights, phase background weight, candidate grids) are hand-set without sensitivity analysis; the final selected configurations are not reported. The 'susceptibility' idea is borrowed from physics by analogy and is not derived. The most important assumption is that squared daily returns are a valid RV proxy for HARQ comparisons.

free parameters (8)
  • Validation-selection criterion weights (0.20, 0.10, 2e-5) = 0.20, 0.10, 2e-5
    Coefficients in Eq. (A2) are hand-chosen penalties for temporal instability, calm-period deterioration, and feature dimension; not optimized or justified.
  • Bounded correction clip C = 1.0
    Enforced in Eq. (3); chosen by hand without sensitivity analysis.
  • Phase-expert background weight = 0.15
    Eq. (B8) sets off-phase expert weight to 0.15; hand-chosen.
  • Stress quantile thresholds q_{s,k} = estimated per asset per fold
    Eq. (11) uses training-segment quantile to define calm/onset/recovery/persistent-stress labels; this is a fitted threshold.
  • Low-rank projection dimension d = validation-selected
    Features are projected to a validation-selected rank before readout; the rank is a selected hyperparameter.
  • Classical reservoir hyperparameters (q, D, chi, A) = candidate set in Eq. (B6)
    Configuration chosen by validation score; spectral radius and random masks not specified.
  • Quantum dynamics parameters (eta_y, eta_phi, theta, h, p_forward, p_backward, gamma) = two candidate sets in Eq. (C17)
    Discrete dynamics chosen by validation; exact simulation used.
  • AR-Ridge regularization lambda = not reported
    The anchor ridge penalty is fit but its value/selection is not reported.
axioms (6)
  • domain assumption Squared daily returns are an adequate realized-variance proxy
    Section 2.6 constructs RV proxy from close-to-close squared returns; no intraday data. HARQ/RQ features and the target variance are built on this proxy, which may bias comparisons against models designed for true realized measures.
  • domain assumption Reservoir echo-state/fading-memory property holds after spectral-radius rescaling
    Equation (5) assumes stable reservoir dynamics without formal verification for the specific non-normal complex matrices; standard ESN assumption.
  • ad hoc to paper Susceptibility of physical critical systems transfers to computational contrast features
    The introduction (Sec. 1) borrows 'critical systems amplify' from quantum sensing (ref [4]) and applies it to reservoir contrast without proof that open-vs-ring differences carry forecast-relevant information.
  • domain assumption Regime definitions by static quantile thresholds are meaningful for forecasting
    Eq. (11) defines calm/onset/recovery/persistent-stress using training quantiles; the causal gate (Eq. 12) must approximate future-dependent labels.
  • domain assumption QLIKE is the correct loss and direct optimization with a bounded correction generalizes
    The readout minimizes QLIKE (Eq. 4); direct loss alignment plus ridge is assumed to generalize despite the small correction.
  • domain assumption Chronological folds with purge intervals prevent leakage
    Section 2.6 states purging and disjoint folds; this is an assumption about data integrity that cannot be verified without data.

pith-pipeline@v1.3.0-alltime-deepseek · 8814 in / 15966 out tokens · 159034 ms · 2026-08-01T04:33:32.845609+00:00 · methodology

0 comments
read the original abstract

Volatility forecasting is dominated by persistence and measurement noise, leaving limited residual structure for nonlinear models to exploit. We introduce Susceptible Architectures (SUSA), a reservoir-design principle for volatility forecasting, and its two concrete implementations, based on complex-valued open-chain and periodic reservoirs and regime-conditioned experts to interpret reservoir features across calm, onset, recovery, and persistent-stress states. We also implement open-system $q$-qubit counterparts in Qiskit while retaining a common AR-Ridge anchor and a bounded residual correction trained under QLIKE. We evaluate models on 16 U.S. equity and exchange-traded-fund series using three disjoint chronological training, validation, and test folds, a 12-observation input window, and a five-observation forecast horizon. The proposed models perform competitively with GARCH, achieving statistically significant QLIKE improvements for specific assets (IWM, XLP). Also models' forecasts complement HARQ-style predictions: a stacked ensemble improves mean QLIKE by 0.0116 over its strongest constituent and wins in 75% of test scenarios.

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

30 extracted references · 2 linked inside Pith

  1. [1]

    Journal of Econometrics , volume =

    Bollerslev, Tim , title =. Journal of Econometrics , volume =. 1986 , doi =

    1986

  2. [2]

    Journal of Financial Econometrics , volume =

    Corsi, Fulvio , title =. Journal of Financial Econometrics , volume =. 2009 , doi =

    2009

  3. [3]

    and Quaedvlieg, Rogier , title =

    Bollerslev, Tim and Patton, Andrew J. and Quaedvlieg, Rogier , title =. Journal of Econometrics , volume =. 2016 , doi =

    2016

  4. [4]

    , title =

    Patton, Andrew J. , title =. Journal of Econometrics , volume =. 2011 , doi =

    2011

  5. [5]

    and Mariano, Roberto S

    Diebold, Francis X. and Mariano, Roberto S. , title =. Journal of Business & Economic Statistics , volume =. 1995 , doi =

    1995

  6. [6]

    and Lunde, Asger and Nason, James M

    Hansen, Peter R. and Lunde, Asger and Nason, James M. , title =. Econometrica , volume =. 2011 , doi =

    2011

  7. [7]

    2001 , note =

    Jaeger, Herbert , title =. 2001 , note =

    2001

  8. [8]

    Neural Computation , volume =

    Hochreiter, Sepp and Schmidhuber, J. Neural Computation , volume =. 1997 , doi =

    1997

  9. [9]

    2007 , doi =

    Statistical Applications in Genetics and Molecular Biology , volume =. 2007 , doi =

    2007

  10. [10]

    Physical Review Applied , volume =

    Fujii, Keisuke and Nakajima, Kohei , title =. Physical Review Applied , volume =. 2017 , doi =

    2017

  11. [11]

    and Yamamoto, Naoki , title =

    Chen, Jiayin and Nurdin, Hendra I. and Yamamoto, Naoki , title =. Physical Review Applied , volume =. 2020 , doi =

    2020

  12. [12]

    npj Quantum Information , volume =

    Mujal, Pere and Mart. npj Quantum Information , volume =. 2023 , doi =

    2023

  13. [13]

    Quantum , volume =

    Sannia, Antonio and Mart. Quantum , volume =. 2024 , doi =

    2024

  14. [14]

    and Wu, Xiaodi and Fatemi, Valla and McMahon, Peter L

    Senanian, Alen and Prabhu, Sridhar and Kremenetski, Vladimir and Roy, Saswata and Cao, Yingkang and Kline, Jeremy and Onodera, Tatsuhiro and Wright, Logan G. and Wu, Xiaodi and Fatemi, Valla and McMahon, Peter L. , title =. Nature Communications , volume =. 2024 , doi =

    2024

  15. [15]

    2024 , archivePrefix =

    Kornja. 2024 , archivePrefix =. doi:10.48550/arXiv.2407.02553 , note =. 2407.02553 , primaryClass =

    Pith/arXiv arXiv doi:10.48550/arxiv.2407.02553 2024

  16. [16]

    and Presutti, Federico and Malia, Benjamin K

    Cimini, Valeria and Sohoni, Mandar M. and Presutti, Federico and Malia, Benjamin K. and Ma, Shi-Yuan and Yanagimoto, Ryotatsu and Wang, Tianyu and Onodera, Tatsuhiro and Wright, Logan G. and McMahon, Peter L. , title =. npj Quantum Information , volume =. 2026 , doi =

    2026

  17. [17]

    2025 , archivePrefix =

    Sch. 2025 , archivePrefix =. doi:10.48550/arXiv.2501.15528 , note =. 2501.15528 , primaryClass =

    Pith/arXiv arXiv doi:10.48550/arxiv.2501.15528 2025

  18. [18]

    npj Quantum Information , volume =

    Sannia, Antonio and Ravell Rodr. npj Quantum Information , volume =. 2026 , doi =

    2026

  19. [19]

    Physical Review E , volume =

    Mart. Physical Review E , volume =. 2023 , doi =

    2023

  20. [20]

    Optica Quantum , volume =

    Sannia, Antonio and Giorgi, Gian Luca and Longhi, Stefano and Zambrini, Roberta , title =. Optica Quantum , volume =. 2025 , doi =

    2025

  21. [21]

    Communications Physics , volume =

    Palacios, Ana and Mart. Communications Physics , volume =. 2024 , doi =

    2024

  22. [22]

    Physical Review Research , volume =

    Li, Qingyu and Mukhopadhyay, Chiranjib and Bayat, Abolfazl and Habibnia, Ali , title =. Physical Review Research , volume =. 2026 , doi =

    2026

  23. [23]

    and Minganti, F

    Di Candia, R. and Minganti, F. and Petrovnin, K. V. and Paraoanu, G. S. and Felicetti, S. , title=. npj Quantum Information , year=. doi:10.1038/s41534-023-00690-z , url=

    doi:10.1038/s41534-023-00690-z

  24. [24]

    Scandinavian Journal of Statistics , volume =

    Holm, Sture , title =. Scandinavian Journal of Statistics , volume =

  25. [25]

    , title =

    Patton, Andrew J. , title =. Journal of Econometrics , year =

  26. [26]

    and Jordan, Michael I

    Jacobs, Robert A. and Jordan, Michael I. and Nowlan, Steven J. and Hinton, Geoffrey E. , title =. Neural Computation , year =

  27. [27]

    Physical Review Applied , year =

    Fujii, Keisuke and Nakajima, Kohei , title =. Physical Review Applied , year =

  28. [28]

    Quantum Reservoir Computing in Finite Dimensions , journal =

    Mart. Quantum Reservoir Computing in Finite Dimensions , journal =. 2023 , volume =

    2023

  29. [29]

    Dissipation as a Resource for Quantum Reservoir Computing , journal =

    Sannia, Andrea and Mart. Dissipation as a Resource for Quantum Reservoir Computing , journal =. 2024 , volume =

    2024

  30. [30]

    and Chuang, Isaac L

    Nielsen, Michael A. and Chuang, Isaac L. , title =. 2010 , isbn =

    2010