Pith. sign in

REVIEW 3 major objections 5 minor 2 cited by

Supervised Similarity for High-Yield Corporate Bonds with Quantum Cognition Machine Learning

T0 review · 3 major / 5 minor · reviewed 2026-08-09 · deepseek-v4-flash

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

Pith's one-line read The paper claims that a supervised distance metric derived from quantum-state representations, one minus the squared overlap of learned quantum states, makes k-nearest-neighbor yield prediction more accurate than random-forest or…

desk verdict Plausible empirical claim that QCML fidelity proximity beats RF GAP on HY bonds, but the comparison has an unaddressed ensembling asymmetry and no significance tests; worth refereeing but not yet convincing. read the letter →

arxiv 2502.01495 v1 pith:HSMGKL5F submitted 2025-02-03 q-fin.ST q-fin.CPq-fin.RMq-fin.TRstat.ML

Joshua Rosaler , Luca Candelori , Vahagn Kirakosyan , Kharen Musaelian , Ryan Samson , Martin T. Wells , Dhagash Mehta , Stefano Pasquali This is my paper · ORCID
classification q-fin.STq-fin.CPq-fin.RMq-fin.TRstat.ML
keywords quantumcognitionmachinelearningsupervisedsimilaritycorporatebondshigh-yieldrandomforestproximityk-nearestneighborsfidelitydistancemetric
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper tries to establish that a supervised distance metric drawn from quantum cognition machine learning (QCML) is better than the current tree-based standard for finding similar corporate bonds. In QCML, each bond is encoded as a quantum state, the encoding is trained to predict the bond's yield or spread, and the distance between two bonds is defined as one minus the squared overlap of their quantum states. Using that distance as the metric in a k-nearest-neighbors regressor, the paper reports lower yield-prediction error than random-forest GAP proximity and Euclidean distance on high-yield bonds, and comparable or better performance on investment-grade bonds. This matters because corporate bonds trade sparsely, so a similarity measure that truly reflects the target variable can identify tradable substitutes, price bonds with few recent quotes, and explain model predictions. The paper's explanation is that QCML keeps sparse, outlier-heavy data in a compact representation, while random-forest proximity spreads most pairs to near-maximum distance.

What carries the argument

The central object is the QCML distance $d_Q(x_t,x_{t'}) = 1 - |\langle \psi_t | \psi_{t'} \rangle|^2$, where each bond $x_t$ is mapped to the ground state $|\psi_t\rangle$ of the error Hamiltonian $H(x_t) = \frac{1}{2}\sum_k (A_k - x_{t,k} I)^2$ built from learned Hermitian feature observables $A_k$ and a target observable $B$ trained by minimizing mean absolute error. The quantum fidelity $|\langle \psi_t | \psi_{t'} \rangle|^2$ between two ground states is the similarity measure; it does the work of making closeness in feature space track closeness in yield or spread. The comparison baseline is random forest GAP proximity, which is the exact weight each training point contributes to a random forest prediction, and ordinary Euclidean distance.

What would settle it

Re-run the HYG comparison with matched resources: one QCML initialization versus one random forest, with both unweighted and proximity-weighted neighbor averaging on the same 10 held-out splits. If random-forest GAP achieves equal or lower mean absolute percentage error across most values of $k$, the paper's claim that QCML proximity is better for high-yield bonds collapses.

Watch

Extended reading notes

Core claim

The central claim is that QCML proximity is a genuinely supervised similarity measure: the quantum states used to define it are learned to predict the target, so fidelity between states encodes target-relevant similarity. On the HYG high-yield cohort, k-nearest-neighbors regression with this metric achieves lower mean absolute percentage error than with random-forest GAP or Euclidean distances for both unweighted and proximity-weighted averaging, and on the IGSB investment-grade cohort it wins under unweighted averaging while being comparable or slightly worse when proximity-weighted. The paper attributes the edge to geometry: QCML produces a compact, coherent representation of a sparse manifold with many one-hot features and near-default outliers, whereas random-forest proximity places most pairs at maximum distance, so its neighbors-based predictions rest on very few points. The same pattern appears on public datasets: where QCML regression matches or beats random forests, the QCML metric matches or beats random-forest GAP in unweighted neighbor averaging.

Load-bearing premise

The evaluation assumes that testing each distance metric by averaging the yields of nearby training bonds is a fair and unbiased measure of practical similarity quality, even though the two methods are compared with different averaging rules, ensemble sizes, and numbers of random starts.

Editorial extensions

If this is right

  • For high-yield portfolios, the QCML metric can directly identify tradable substitutes, because neighbors under it have yields much closer to a target bond than neighbors under random-forest GAP or Euclidean distance.
  • For investment-grade bonds, the QCML metric remains competitive: the paper finds it better under unweighted neighbor averaging and comparable or slightly worse under proximity-weighted averaging.
  • Where QCML regression matches or beats random forests on public datasets, the QCML metric also matches or beats random-forest GAP in unweighted neighbor averaging, suggesting the effect is not unique to bonds.
  • The compact representation property means QCML neighbor predictions draw on a broad set of nearby points, while random-forest proximity concentrates weight on a few near-maximum-distance points, which explains the high-yield advantage.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • A decisive test of the proposed mechanism is to compare single-initialization QCML with an un-ensembled random forest using both GAP and out-of-bag proximities; the paper's own comparison averages three QCML distance matrices but no random-forest ensembles.
  • Because the learned distance is target-specific, the same QCML pipeline could be trained on a multi-target objective such as yield, spread, duration, and rating to produce a similarity that balances several dimensions of bond risk, which a single-target metric cannot.
  • A practical extension is to measure realized trading costs rather than neighbor-averaging error: find the nearest substitute under each metric, attempt to trade it, and compare execution quality, since lower yield-prediction error does not by itself prove a better tradable alternative.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. The paper introduces a supervised distance metric for corporate bonds derived from Quantum Cognition Machine Learning (QCML), where each bond is mapped to a quantum state and distance is defined through quantum fidelity. The authors compare this QCML distance with random forest GAP proximity and Euclidean distance by evaluating the MAPE/MAE of a k-nearest-neighbors regressor on hold-out splits of HYG (high-yield) and IGSB (investment-grade) bond data, as well as on three public datasets. They report that QCML proximity outperforms the other metrics for HYG in both unweighted and proximity-weighted KNN, and performs comparably or slightly worse for IGSB in the weighted case. The paper also includes MDS visualizations of the learned distances and a comparison of QCML versus random forest regression performance.

Significance. If the comparison were fully controlled, the paper would make a useful contribution to supervised similarity learning for illiquid corporate bonds: it addresses a concrete practical problem (finding tradeable substitutes for illiquid bonds), benchmarks against the current state-of-the-art (RF GAP proximity), and reports public-dataset controls. The authors are transparent about one known confound, the proximity-weighting scheme favoring RF. However, the central empirical claim of a HYG advantage rests on an asymmetric ensemble setup and on mean differences without significance testing; these issues must be resolved before the claim can be considered established.

major comments (3)
  1. [§5.2, Figure 4(a)-(b)] Section 5.2 states that the QCML distance matrix is "ensembled 3 different distance matrices... by taking the average of each entry," while the Random Forest GAP proximities are computed from a single forest. This asymmetry is not controlled in Figure 4(a)-(b). Because averaging fidelities over independent initializations reduces variance and can smooth the distance distribution, the HYG advantage attributed to QCML may instead reflect ensembling rather than the learned metric itself. Please report results for a single QCML initialization and for RF GAP proximities averaged over multiple forests (e.g., three or more seeds), so that the comparison controls for ensembling.
  2. [§6.2, Figure 4, Table 2] The paper claims that QCML "outperforms" the other metrics, but no statistical significance tests are reported. In Table 2 the HYG metrics for QCML and RF overlap heavily at the level of one standard deviation (e.g., MAE .79±.06 versus .93±.08, RMSE 1.73±.17 versus 1.80±.19), and the bands in Figure 4 are standard errors of the mean only. Please report paired tests over the 10 splits (e.g., paired t-test or Wilcoxon signed-rank) for each value of k, with appropriate multiple-comparison correction, or otherwise quantify the effect size and its uncertainty.
  3. [§5.2, §6.2] The manuscript acknowledges that "proximity-weighting inherently favors the RF-based proximities" because RF distances are near the maximum for most pairs, while QCML distances are broadly distributed. This admission implies that the weighted KNN results in Figure 4(b,d) do not provide a neutral comparison of metric quality: the weighting function is chosen in a model-dependent way, and for RF GAP it coincides with the model's own prediction weights, an interpretation not available for the other metrics. The unweighted KNN results are therefore the primary evidence for metric quality. Please either analyze weighted results under a model-independent weighting function, or clearly state that the weighted comparison is an application-specific evaluation and not a fair head-to-head metric test.
minor comments (5)
  1. [§1] The introduction's organization paragraph states "Section 8, the conclusion" but the manuscript contains a Section 7 on MDS visualization; the outline should be updated to include Section 7.
  2. [§4] There is a typo: "one-hot-econding" should be "one-hot encoding."
  3. [Figure 3] The y-axis is logarithmic for the random forest panels but linear for the QCML panels; this makes the visual comparison of the distance distributions misleading. Please use a consistent scale or add a note explaining the different scales.
  4. [Table 3] The caption for the Student Performance rows explains why MAPE diverges (target values include zero), but the main text in §6.1 does not mention this. Consider adding one sentence to the text to avoid confusion.
  5. [References] Reference [15] (Lin and Jeon) and reference [20] (Rhodes, Cutler, Moon) are both cited in connection with the GAP proximity definition; please verify that the GAP formula in Eq. (4) is attributed to the correct source.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the QCML distance is a learned representation evaluated out-of-sample against independent baselines.

full rationale

The paper's central derivation is self-contained and does not reduce to its inputs. The QCML distance d_Q(x_t, x_t') = 1 - |<ψ_t|ψ_t'>|^2 (Eq. 6) is defined from ground states of an error Hamiltonian (Eq. 5) whose observables are trained by gradient descent to minimize MAE on the regression target (Section 3); the KNN evaluation (Section 5.2) then measures out-of-sample MAPE on held-out splits for QCML, RF-GAP, and Euclidean distances, reporting averages over 10 splits (Figure 4). No equation is defined in terms of the quantity it is said to predict: the distance is not constructed directly from the KNN target values, and the QCML training loss is not the KNN MAPE. The comparison to RF-GAP is a supervised-vs-supervised baseline, and the paper includes public-dataset benchmarks (Table 3, Figure 5) that externally anchor the regressor and metric behavior. The paper itself acknowledges limitations in Section 5.2, stating that 'proximity-weighting inherently favors the RF-based proximities' and that the absence of a model-independent weighting scheme 'obstructs a fair, like-for-like comparison'; Section 6.1 also notes that the regressor's performance 'directly impacts the quality of the metric extracted from that model.' These are validity and fairness concerns, not circularity, because they do not make the reported advantage equivalent to an input by construction. The self-citations to prior QCML papers [7,16,23] supply background and motivation, but the empirical claims are supported by the paper's own out-of-sample experiments. The noted asymmetry—QCML distances averaged over 3 initializations while RF GAP uses a single forest—is a methodological imbalance that could affect the conclusions, but it does not make the derivation circular. Therefore no significant circularity is present.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The central claim depends on the QCML model architecture with fidelity proximity and on the evaluation protocol. No new physical entities are postulated; the quantum observables are learned model parameters rather than invented entities.

free parameters (3)
  • Hilbert space dimension N = N=7 (HYG), N=12 (IGSB)
    Selected by 3-fold cross-validation minimizing MSE on the training set (Section 5.1); controls model capacity and therefore the shape of the learned proximity.
  • Number of QCML models ensembled = 3
    The distance matrix is an average over 3 QCML models with different initializations (Section 5.2); no justification for this number is given.
  • KNN proximity weights = proximity for QCML/RF, 1-d normalized for Euclidean
    The weighting scheme is acknowledged to be subjective and model-dependent (Section 5.2); it affects the weighted KNN comparison.
assumptions (4)
  • domain assumption Quantum fidelity between ground states defines a meaningful supervised similarity after training on the target variable.
    Section 3.1 introduces d_Q = 1 - |<ψ_t|ψ_t'>|^2; this presumes the learned state representation reflects target-relevant structure.
  • domain assumption The error Hamiltonian H(x_t) has a unique ground state for the data points encountered, so the map x_t -> |ψ_t> is well-defined.
    Section 3 notes singular points at degeneracy but does not analyze whether they occur; if degeneracy is common, the fidelity distance is unstable.
  • domain assumption KNN-MAPE is a valid proxy for the utility of a supervised similarity metric in identifying tradable substitutes.
    Section 5.2 uses this evaluation; the paper argues a good metric brings bonds of similar yield closer, but the link to trading costs is not quantified.
  • standard math Random forest GAP proximity is the appropriate state-of-the-art baseline, computed as in refs [19,20].
    The paper relies on the GAP implementation for comparison; no code is provided to verify the computation.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Supervised Similarity for High-Yield Corporate Bonds with Quantum Cognition Machine Learning." pith.science (2026). https://pith.science/paper/HSMGKL5F

@misc{pith2026250201495,
  author       = {Pith},
  title        = {Pith review of: Supervised Similarity for High-Yield Corporate Bonds with Quantum Cognition Machine Learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/HSMGKL5F}},
  note         = {Machine review of arXiv:2502.01495}
}
read the original abstract

We investigate the application of quantum cognition machine learning (QCML), a novel paradigm for both supervised and unsupervised learning tasks rooted in the mathematical formalism of quantum theory, to distance metric learning in corporate bond markets. Compared to equities, corporate bonds are relatively illiquid and both trade and quote data in these securities are relatively sparse. Thus, a measure of distance/similarity among corporate bonds is particularly useful for a variety of practical applications in the trading of illiquid bonds, including the identification of similar tradable alternatives, pricing securities with relatively few recent quotes or trades, and explaining the predictions and performance of ML models based on their training data. Previous research has explored supervised similarity learning based on classical tree-based models in this context; here, we explore the application of the QCML paradigm for supervised distance metric learning in the same context, showing that it outperforms classical tree-based models in high-yield (HY) markets, while giving comparable or better performance (depending on the evaluation metric) in investment grade (IG) markets.

Figures

Figures reproduced from arXiv: 2502.01495 by the authors.

Figure 1
Figure 1. Distributions of bond price, yield and spread for HYG and IGSB [PITH_FULL_IMAGE:figures/full_fig_p006_1.png] view at source ↗
Figure 2
Figure 2. Distribution of some of the data features for both HYG and IGSB We will demonstrate that these substantial differences in the distributions of the data features lead to significant differences in predictive performance when comparing QCML and Random Forest. In particular, we will show below that QCML holds an advantage on HYG, due to the sparse characteristics of the data. 5 Methods We now describe how to train and … view at source ↗
Figure 3
Figure 3. Average distance from a reference bond to its neighbors for both HYG and IGSB. The average is taken over all bonds in the data set. For Random Forest (GAP proximity), almost all neighbors are at a maximum distance from the reference bond. Note that for Random Forest the y-axis is in logarithmic scale. 6 Results Our main goal in this section is to evaluate the relative quality of the similarity matrices extracted fro… view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Average MAPE for K-Nearest-Neighbors (KNN) regression over 10 train/test splits of the data, for both HYG (upper plots) and IGSB (lower plots). The bands represent standard error in estimation of the mean for each k. In the plots on the left, a) and c), the KNN predict…
Figure 5
Figure 5. Figure 5: Average out-of-sample MAPE or MAE as a function of k for k-nearest-neighbors regression over 10 train/test splits of the data, for three well-known public datasets. Each row corresponds to a distinct dataset; plots on the left correspond to unweighted KNN and plots on …
Figure 6
Figure 6. Figure 6: Multi-dimensional scaling visualization of both QCML (top) and Random Forest GAP proximities (bottom), for both (a) HYG and (b) IGSB. Darker colors represent higher yields. In all plots, a reference bond is displayed in red and the 10 closest neighbors are displayed in…

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 2 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Quantum Geometry of Data

    cs.LG 2025-07 conditional novelty 5.0 of 10

    QCML learns matrix configurations whose quantum geometry reproduces known manifolds and reveals structure in real datasets.

  2. Quantum Cognition Machine Learning for Forecasting Chromosomal Instability

    q-bio.QM 2025-06 reject novelty 4.0 of 10

    QCML, a quantum-inspired ML method, predicts LST status from CTC morphology with balanced accuracy 70%, only marginally better than classical models in a 227-cell study.

Reference graph

Works this paper leans on

27 extracted references · 26 canonical work pages · cited by 2 Pith papers

  1. [1]

    Applications of quantum statistics in psychological studies of decision processes

    Diedrik Aerts and Sven Aerts. Applications of quantum statistics in psychological studies of decision processes. Foundations of Science, 1(1):85–97, 1995

    work page 1995

  2. [2]

    Random forests

    L Brieman and A Cutler. Random forests. https://www.stat.berkeley.edu/~breiman/RandomForests/ cc_home.htm, 2001

    work page 2001

  3. [3]

    Jerome Busemeyer and Peter D. Bruza. Quantum Models of Cognition and Decision. Cambridge University Press, 2012

    work page 2012

  4. [4]

    Can an unsupervised clustering algorithm reproduce a categorization system?

    Nathalia Castellanos, Dhruv Desai, Sebastian Frank, Stefano Pasquali, and Dhagash Mehta. Can an unsupervised clustering algorithm reproduce a categorization system? arXiv preprint arXiv:2408.10340 , 2024

    work page Pith review arXiv 2024

  5. [5]

    Quantifying Outlierness of Funds from their Categories using Supervised Similarity

    Dhruv Desai, Ashmita Dhiman, Tushar Sharma, Deepika Sharma, Dhagash Mehta, and Stefano Pasquali. Quantifying outlierness of funds from their categories using supervised similarity. arXiv preprint arXiv:2308.06882, 2023

    work page Pith review arXiv 2023

  6. [6]

    rfproximity: A python package for rf proximity analysis,

    Dhruv Desai, Dhagash Mehta, and Julio Urquidi. rfproximity: A python package for rf proximity analysis,

    work page

  7. [7]

    Robust estimation of the intrinsic dimension of data sets with quantum cognition machine learning, 2024

    Luca Candelori et al. Robust estimation of the intrinsic dimension of data sets with quantum cognition machine learning, 2024. https://arxiv.org/abs/2409.12805

    arXiv 2024

  8. [8]

    Open Set Recognition for Random Forest

    Guanchao Feng, Dhruv Desai, Stefano Pasquali, and Dhagash Mehta. Open set recognition for random forest. arXiv preprint arXiv:2408.02684 , 2024

    work page Pith review arXiv 2024

Show all 27 references
  1. [9]

    Random forests

    Trevor Hastie, Robert Tibshirani, Jerome Friedman, Trevor Hastie, Robert Tibshirani, and Jerome Fried- man. Random forests. The elements of statistical learning: Data mining, inference, and prediction , pages 587–604, 2009

    2009

  2. [10]

    Matrix geometry and coherent states

    Goro Ishiki. Matrix geometry and coherent states. Physical Review D , 92(4), August 2015. 14

    2015

  3. [11]

    Supervised similarity learning for corporate bonds using random forest proximities

    Jerinsh Jeyapaulraj, Dhruv Desai, Dhagash Mehta, Peter Chu, Stefano Pasquali, and Philip Sommer. Supervised similarity learning for corporate bonds using random forest proximities. In Proceedings of the Third ACM International Conference on AI in Finance , pages 411–419, 2022

    2022

  4. [12]

    interference of minds

    Andrei Khrennikov. Quantum-like brain: “interference of minds”. Biosystems, 84(3):225–241, 2006

    2006

  5. [13]

    Quantile regression using random forest proximities

    Mingshu Li, Bhaskarjit Sarmah, Dhruv Desai, Joshua Rosaler, Snigdha Bhagat, Philip Sommer, and Dha- gash Mehta. Quantile regression using random forest proximities. arXiv preprint arXiv:2408.02355 , 2024

    2024 arXiv

  6. [14]

    Classification and regression by randomforest

    Andy Liaw, Matthew Wiener, et al. Classification and regression by randomforest. R news , 2(3):18–22, 2002

    2002

  7. [15]

    Random forests and adaptive nearest neighbors

    Yi Lin and Yongho Jeon. Random forests and adaptive nearest neighbors. Journal of the American Statistical Association, 101(474):578–590, 2006

    2006

  8. [16]

    Quantum cognition machine learning: Ai needs quantum, 2024

    Kharen Musaelian et al. Quantum cognition machine learning: Ai needs quantum, 2024. https://www. qognitive.io/QCML%20-%20Qognitive,%20Inc.pdf

    2024

  9. [17]

    Nielsen and Isaac L

    Michael A. Nielsen and Isaac L. Chuang. Quantum Computation and Quantum Information . Cambridge University Press, 2000

    2000

  10. [18]

    Pothos and Jerome R

    Emmanuel M. Pothos and Jerome R. Busemeyer. Quantum cognition. Annual Review of Psychology , 73(1):749–778, 2022

    2022

  11. [19]

    Rf-gap python: Python code for random forest-geometry- and accuracy-preserving proximi- ties, 2024

    Jake Rhodes. Rf-gap python: Python code for random forest-geometry- and accuracy-preserving proximi- ties, 2024. https://github.com/jakerhodes/RF-GAP-Python

    2024

  12. [20]

    Rhodes, Adele Cutler, and Kevin R

    Jake S. Rhodes, Adele Cutler, and Kevin R. Moon. Geometry- and accuracy-preserving random forest proximities, 2023

    2023

  13. [21]

    Enhanced local explainability and trust scores with random forest proximities

    Joshua Rosaler, Dhruv Desai, Bhaskarjit Sarmah, Dimitrios Vamvourellis, Deran Onay, Stefano Pasquali, and Dhagash Mehta. Enhanced local explainability and trust scores with random forest proximities. In Proceedings of the 5th ACM International Conference on AI in Finance , pag...

    2024

  14. [22]

    Machine learning-based relative valuation of municipal bonds

    Preetha Saha, Jingrao Lyu, Dhruv Desai, Rishab Chauhan, Jerinsh Jeyapaulraj, Philip Sommer, and Dhagash Mehta. Machine learning-based relative valuation of municipal bonds. arXiv preprint arXiv:2408.02273, 2024

    2024 arXiv

  15. [23]

    Quantum cognition machine learning: financial forecasting, 2024

    Ryan Samson et al. Quantum cognition machine learning: financial forecasting, 2024. Risk.net

    2024

  16. [24]

    Steinacker

    Harold C. Steinacker. Quantum (matrix) geometry and quasi-coherent states. Journal of Physics A: Mathematical and Theoretical, 54(5):055401, January 2021

    2021

  17. [25]

    Steinacker

    Harold C. Steinacker. Quantum Geometry, Matrix Theory, and Gravity. Cambridge University Press, 2024

    2024

  18. [26]

    Case-based ex- plainability for random forest: Prototypes, critics, counter-factuals and semi-factuals

    Gregory Yampolsky, Dhruv Desai, Mingshu Li, Stefano Pasquali, and Dhagash Mehta. Case-based ex- plainability for random forest: Prototypes, critics, counter-factuals and semi-factuals. arXiv preprint arXiv:2408.06679, 2024. 15

    2024 arXiv

  19. [2024]

    https://pypi.org/project/rfproximity

Pith tools

Reviewed August 9, 2026 · model on record in the stance chip above.