Pith. sign in

REVIEW 1 major objections 4 minor 38 references

Residual Algebra for Representation-Preserving Learning

T0 review · 1 major / 4 minor · reviewed 2026-08-10 · deepseek-v4-flash

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

Pith's one-line read Typing residuals by their generating representation and composing them in a fixed order—relax, aggregate, close—lifts net return from 13.52% to 19.10% on a frozen 2023–2026 Chinese A-share test.

desk verdict A thoughtful architecture study with clean math and careful controls, but its headline gains are compromised by architectural choices made on the evaluation window. read the letter →

arxiv 2608.07349 v1 pith:SGE5PDVG submitted 2026-08-07 cs.LG

Yao Wu This is my paper · ORCID
classification cs.LG
keywords residualalgebratypedresidualsfactorfieldspoint-in-timeevaluationcross-sectionalreturnpredictioncontrolvariatesgradientboostingrepresentation-preservinglearning
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

Conventional learning from heterogeneous representations flattens them into one feature vector, and in doing so loses the identity of the representation that caused each error. The paper argues that the residual should be typed: each representation owns its own coordinate system and the residual it leaves unresolved, and learning should be an ordered composition of operators that preserve or deliberately erase that type. The evaluated instance, FPRC-PQ, instantiates the algebra as relax–aggregate–close on three $10\times10$ rank-grid fields, and on 3.67 million Chinese A-share stock-day observations (2023–2026) under a frozen point-in-time protocol it raises net-of-cost return from 13.52% to 19.10% and Sharpe from 1.42 to 2.09. Matched-capacity, unified-residual, identity-free two-stage, and pairwise-only controls all trail it, which the paper takes as evidence that the gain comes from residual ownership and composition order, not from more features or more trees.

What carries the argument

The central object is a typed residual: a residual is owned by the representation that generated it and can be corrected only in that representation's own coordinates. Fold realizes representations as $10\times10$ point-in-time conditional-mean grids; FPRC-PQ realizes the algebra as relax–aggregate–close. The load-bearing identities are the telescoping decomposition $\hat S=\bar F+\hat\rho+\widehat{(\rho-\hat\rho)}$, which partitions capacity across stages, and the control-variate form $B_{PQ}=\bar F+\hat\rho$, whose population analogue weakly reduces residual second moment. A further identity fixes the analytical rumination gain $\lambda^\star=\langle e_0,q\rangle/\langle q,q\rangle$ by orthogonal projection, making the reflective extension parameter-free.

What would settle it

A preregistered replication on a different market that fixes all architectural choices using only data before the evaluation window and then runs the same matched controls—direct, unified-residual, identity-free two-stage, pairwise-only—would settle the claim, since the central claim predicts FPRC-PQ should beat those controls out of sample and a null or reversal would falsify it.

Watch

Extended reading notes

Core claim

Formally, a Fold field $F_m$ is the point-in-time conditional-mean map over a $10\times10$ grid of daily rank deciles, and the population residual $r-F_m$ is what that representation leaves unresolved. The algebra composes three typed operators: relaxation corrects each field in its own coordinates, aggregation $A$ takes the equal-mean interface and is the only step that erases representation identity, and closure fits the aggregate's fresh residual without rereading local state. The composition telescopes to $\bar F+\hat\rho+\widehat{(\rho-\hat\rho)}$, so capacity is partitioned across stages, and the aggregate $B_{PQ}=\bar F+\hat\rho$ is a learned control-variate interface. The central discovery is that this ordering, not raw capacity, is what separates the architecture from pooling: the same features, the same tree budget, and the same pairwise interaction class all fail to match FPRC-PQ, which asserts that representation-preserving residual composition is the cause of the reported 5.58 pp net-return improvement and the 0.67 Sharpe improvement.

Load-bearing premise

Everything rests on the claim that no 2023–2026 label or outcome influenced any architectural choice, because if the audits used evaluation-window data to choose the structure, the compared systems are no longer independent of the test set.

Editorial extensions

If this is right

  • Each correction stage should fit only the fresh residual left by its predecessor; letting the shared closer reread local states or the aggregate's own rank is significantly harmful.
  • Capacity placement matters more than capacity size: a matched-capacity direct learner with exactly the combined tree budget trails FPRC-PQ by 2.13 pp net and loses in all four years.
  • Preserving representation identity through local correction is necessary: collapsing the three fields into one concatenated representation and fitting a unified residual trails by 2.27 pp in 4/4 years, and an identity-free two-stage booster trails by 1.69 pp.
  • The equal-mean aggregation interface is already sufficient: learned coefficient reweighting moves net return by less than 0.05 pp within noise, while doubling the anchor harms all three chains.
  • Side information cannot repair a mistyped residual target: eight meta-information channels all fail, several significantly, and an error-correlated drift scalar destroys 1.13 pp when exposed after identity erasure.

Reading between the lines

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

  • A testable transfer consequence: the same relax–aggregate–close routing could be applied to heterogeneous representations outside finance—for example, different model families or feature blocks in a tabular benchmark—provided each operator is fitted only to its own residual and the identity-erasure boundary is fixed; the paper leaves this extension implicit.
  • Reflective rumination is an unevaluated theoretical member of the algebra; a natural next experiment is to fit the reconstruction $T$ and the analytical gain $\lambda^\star$ on data strictly before the evaluation window and compare the $\lambda^\star$ member against the base backbone, which would tell whether the algebra has a second working mechanism beyond the evaluated relax–aggregate–close ins
  • The reported margin could depend on the field-selection rule in Appendix A, which orders candidate triples by pre-2023 residual-error standard deviation; a replication that randomizes field selection while fixing the algebra would separate the contribution of the selected fields from the contribution of the composition itself.
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

1 major / 4 minor

Summary. The paper introduces 'residual algebra' as a representation-learning principle: each Fold field is a 10x10 point-in-time conditional-mean table that owns both a coordinate system and the residual it leaves; FPRC-PQ composes a typed relaxation, an equal-mean aggregation boundary, and a fresh-residual shared closer. The authors derive an exact telescoping decomposition, a control-variate interpretation, a Neyman-style coupled-refit orthogonality, and an analytical 'reflective rumination' extension (explicitly not evaluated). On 3.67M Chinese A-share stock-day rows (2023-2026) under a claimed frozen protocol, FPRC-PQ earns 19.10% net return and Sharpe 2.09 versus 13.52% and 1.42 for the pooled-field baseline, with matched-capacity, unified-residual, identity-free, pairwise-only, and side-information controls trailing. The central empirical claim is that the gain is caused by explicit residual ownership and ordered composition, not by more features or more trees.

Significance. Conditional on the frozen-protocol claim being true, the paper is a valuable and unusually disciplined contribution: the algebraic identities are clean and exactly stated; the matched-capacity direct control, the unified-residual identity control, and the eight side-information channels are the right kinds of falsifiable contrasts; paired 21-day block-bootstrap inference is appropriate; the reflective rumination is honestly scoped as not part of the empirical headline; and the bit-exact reproduction audit is a strong reproducibility promise. The main significance risk is that the architecture-selection audits in §5.2 are reported on the 2023-2026 evaluation window, which, if they were used to choose the architecture, makes the headline comparisons post-selection and undermines the causal attribution. The contribution's value therefore depends on resolving that contradiction.

major comments (1)
  1. [§5.2, Table 2; §3.2; §4] The manuscript asserts in §4 that 'no 2023–2026 label influenced any choice' and in §3.2 that five preregistered audits selected (9)–(11), but Table 2 reports the audit contrasts as paired block-bootstrap deltas on the strict 2023–2026 book (e.g., A1: typed local-PQ vs. pooled fields +4.78 pp; A4: local-only vs. dual-layer Q +0.30 pp; A5: self-state vs. empty −0.45 pp). These realized-window outcomes are used to select the architecture, so the headline comparison FPRC-PQ versus pooled fields and versus matched-capacity direct is post-selection. The bootstrap CIs in Tables 1 and 9 do not condition on this selection and therefore overstate confidence. This directly undermines the paper's strongest claim in the Abstract and §5.1 that the gain is due to residual ownership rather than features or capacity. Please either (i) show that every A1–A5 decision was made from data through 2022-12-31 alone, with the evaluation-window tables provided only as retrospective illustrations, or (ii) treat the audits as exploratory and re-estimate the headline comparisons under a genuinely fixed architecture with selection-adjusted inference.
minor comments (4)
  1. [Table 1] Several numeric columns run together in the printed table (for example, '5.012.0921.67' and '2.5719.10'), making the row values difficult to read; the table should be reformatted with explicit column spacing.
  2. [§2, Eq. (2)] The notation bE_PIT and the purge-gap label availability 'through t−h−1' are used before being defined; please define these terms at first use in Section 2 rather than only in the appendix.
  3. [Appendix A] The term 'strict-OOF modular-FPRC experts' is introduced without definition; please clarify the out-of-fold construction used to score all 1,330 candidate triples.
  4. [§5.2] The reported Pr[Δ>0] values are one-sided probabilities; please state explicitly which direction each probability refers to so that a reader cannot mistake them for two-sided p-values.

Circularity Check

0 steps flagged · score 0.0 of 10

No material circularity: the algebraic identities are definitional and the reported gains are measured, not derived from the algebra.

full rationale

The paper's load-bearing empirical claims are the net-of-cost return and Sharpe improvements (Table 1) measured under a frozen protocol, not quantities derived from the residual algebra. The exact telescoping (Eq. 13), the control-variate inequality (Eq. 16), and the reflective-rumination gain (Prop. 3.1) are identities or standard projection results that hold by construction; the paper explicitly states that reflective rumination is a theoretical extension and that all reported returns use the lambda=0 member (Sections 3.1 and 4). No self-citation is load-bearing: references to stagewise boosting, control variates, and bootstrap methods are standard external results, and the paper explicitly acknowledges its relationship to them rather than presenting them as novel. The only notable concern is that Section 5.2's audits use 2023-2026 outcomes to select architecture (9)-(11), which is a post-selection or statistical-validity issue rather than a circular derivation: the headline comparison to the pooled-field baseline is measured on the same window, not algebraically implied by the fit. That concern belongs in a correctness assessment, not in a finding of circularity. Limitations (one market, one asset class, two-dimensional fields) are stated in Section 7 and do not constitute circularity.

Assumptions & free parameters 7 free parameters · 6 assumptions · 3 invented entities

The central claim relies on several hand-chosen parameters (field committee, Q axes, hyperparameters, grid size) that are frozen before the evaluation window, plus standard mathematical assumptions about projections and control variates. The conceptual entities are not independently verified.

free parameters (7)
  • Field committee (F1, F2, F3) = momentum x reversal, value x momentum, breakout x volume
    Hand-selected from candidate factor pairs; selection rule in App A ranks 1330 triples by pre-2023 field-error STD. A model selection choice frozen before evaluation.
  • Purified Q-axis set = volatility 20d, netprofit yoy, ocf to profit, profit stability
    Selected by pre-2023 residual IC/ICIR and sign-consistency screening (App C, Table 7). Adds four coordinates to each local expert.
  • Fold bin count (10x10) = 10 deciles per axis
    Hand-chosen grid resolution; changed grid would alter residual granularity.
  • Number of fields M = 3
    Chosen by design; affects aggregation and typed residual populations.
  • XGBoost hyperparameters = depth 5, lr 0.035, max_trees 600, min_child_weight 2000, lambda 10
    Frozen across all models; not tuned on evaluation window.
  • Time-decay half-life = 252 trading days
    Used in sample weights for PIT conditional means; hand-chosen.
  • Cost model = blended 10/15 bp
    Assumed transaction cost; affects net return and Sharpe.
assumptions (6)
  • standard math L2 projection properties: the PIT conditional mean is the orthogonal projection onto the sigma-field generated by the grid
    Section 2, projection view. Standard Hilbert space facts.
  • standard math Control variate identity: for a zero-mean control variate, the variance-reduced estimator has lower or equal second moment
    Section 3, Eq. 16. Standard control variate theory (Nelson 1987).
  • standard math Neyman orthogonality of the coupled refit path
    Eq. 17 and surrounding text. The proof is a direct computation.
  • domain assumption Point-in-time data availability: all inputs (factors, ranks, purified axes) are computable from data observable at deployment time
    Section 4 protocol. If PIT assumptions fail (e.g., lookahead in vendor data), all results degrade.
  • ad hoc to paper The equal-mean aggregation captures all useful information across corrected fields
    Audit A2 supports it empirically on the evaluation window, but this is a design choice not derived from first principles.
  • ad hoc to paper The shared closer should not reread local state or the aggregate's own rank
    Audit A5 rejects self-state on the eval window; this is an architectural constraint, not a theorem.
invented entities (3)
  • Factor field F_m (10x10 PIT conditional-mean table)
    purpose: Represents an interaction as a typed object owning a residual
    A mathematical construct defined by the paper; no out-of-sample falsifiable handle beyond the reported backtest.
  • Residual algebra operators (relax, aggregate, close)
    purpose: Formalizes staged composition of residual learning
    Conceptual framework; the operators' efficacy is only evidenced by the paper's own experiments.
  • Reflective rumination operator
    purpose: Analytically fixes a second-stage correction gain via orthogonal projection
    Explicitly untested in the evaluation (Section 3.1); no independent evidence.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Residual Algebra for Representation-Preserving Learning." pith.science (2026). https://pith.science/paper/SGE5PDVG

@misc{pith2026260807349,
  author       = {Pith},
  title        = {Pith review of: Residual Algebra for Representation-Preserving Learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/SGE5PDVG}},
  note         = {Machine review of arXiv:2608.07349}
}
read the original abstract

Learning from heterogeneous representations is usually reduced to feature concatenation, which erases which representation produced an error. We instead algebraize the residual: a representation is a typed object that owns both a coordinate system and the residual it leaves unresolved, and learning is an ordered composition of operators that preserve or deliberately erase that type. Fold realizes the objects as point-in-time conditional-mean fields on 10x10 rank grids. FPRC-PQ realizes the algebra as relax-aggregate-close: each field is relaxed by a correction fitted to its own residual in its own coordinates; corrected fields meet at a fixed mean that is the sole identity-erasure boundary; and a shared learner closes only the aggregate's fresh residual. The composition telescopes exactly into representation, local residual estimate, and residual-of-residual estimate. Its aggregate is a learned control-variate interface with population variance reduction, while refitting the closer along perturbations of the backbone yields first-order coupled-path mean orthogonality. As an analytical extension, a reflective rumination operator reads the displacement of a global reconstruction from the aggregate anchor, reflects it, and fixes its gain by a unique orthogonal projection rather than return-tuned grid search. On 3.67M Chinese A-share stock-day observations (2023-2026) under a frozen point-in-time protocol, the evaluated base algebra raises net-of-cost return from 13.52% to 19.10% and Sharpe from 1.42 to 2.09. Matched-capacity, unified-residual, identity-free two-stage, and pairwise-only controls all trail it. The gain is therefore not explained by more features or more trees, but by making residual ownership and composition explicit while representation identity is still available.

Figures

Figures reproduced from arXiv: 2608.07349 by the authors.

Figure 1
Figure 1. Fold is a double sort read as a one-hot encoding. (a) A daily cross-section in continuous factor space; one stock highlighted. (b) Both factors are ranked within the day and cut into deciles; the rank pair selects exactly one of 10×10 interaction states (bands: the stock’s column decile a=6 and row decile b=3). (c) The induced indicator es ∈ {0, 1} 100 has a single active entry at s = 10a + b. (d) The field is the l… view at source ↗
Figure 2
Figure 2. Residual structure is routed through typed operators. The residual is sequentially consumed by three coordinate systems: field grids sm (top), typed local axes (middle), and shared parents Xpar (bottom), following Eq. 14. Fold removes grid-level structure; RPQ corrects each field in its own coordinates; A aggregates the corrected fields; and C operates only on the remaining ag￾gregate residual. The staircase is sche… view at source ↗
Figure 3
Figure 3. (a) Cumulative net-of-cost return, strict book. Curves separate where the residual algebra predicts: typed local correction over pooling, then fresh closure over either stage alone. (b) The same architecture ladder with annualized Sharpe [PITH_FULL_IMAGE:figures/full_fig_p009_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Eight failed shortcuts. Paired deltas versus frozen FPRC-PQ with 95% block-bootstrap intervals, strict book. Amber denotes local channels and red shared channels; the right column gives Pr[∆ > 0]. Error-correlated metadata does not improve the model merely by being app…
Figure 5
Figure 5. Figure 5: Capacity placement is an either/or. Net return (%) in two frozen 2×2 audits. (a) A4: opening Q locally is the best cell; adding shared Q on top of local Q reduces net (interaction −0.31 pp, CI [−0.35, −0.01]). (b) A3: reopening both cross-level edges is never the best …

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

38 extracted references · 33 canonical work pages

  1. [1]

    Rishabh Agarwal, Levi Melnick, Nicholas Frosst, Xuezhou Zhang, Ben Lengerich, Rich Caru- ana, and Geoffrey E. Hinton. Neural additive models: Interpretable machine learning with neural nets. InAdvances in Neural Information Processing Systems, volume 34, 2021

    work page 2021

  2. [2]

    Arık and Tomas Pfister

    Sercan ¨O. Arık and Tomas Pfister. TabNet: Attentive interpretable tabular learning. InProceed- ings of the AAAI Conference on Artificial Intelligence, volume 35, pages 6679–6687, 2021

    work page 2021

  3. [3]

    Bailey, Jonathan Borwein, Marcos L´opez de Prado, and Qiji Jim Zhu

    David H. Bailey, Jonathan Borwein, Marcos L´opez de Prado, and Qiji Jim Zhu. The probability of backtest overfitting.Journal of Computational Finance, 20(4):39–69, 2017

    work page 2017

  4. [4]

    Deep neural networks and tabular data: A survey.IEEE Transactions on Neural Networks and Learning Systems, 35(6):7499–7519, 2024

    Vadim Borisov, Tobias Leemann, Kathrin Seßler, Johannes Haug, Martin Pawelczyk, and Gjergji Kasneci. Deep neural networks and tabular data: A survey.IEEE Transactions on Neural Networks and Learning Systems, 35(6):7499–7519, 2024

    2024

  5. [5]

    Intelligible models for healthcare: Predicting pneumonia risk and hospital 30-day readmission

    Rich Caruana, Yin Lou, Johannes Gehrke, Paul Koch, Marc Sturm, and Noemie Elhadad. Intelligible models for healthcare: Predicting pneumonia risk and hospital 30-day readmission. InProceedings of the 21st ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, pages 1721–1730, 2015

    work page 2015

  6. [6]

    Factor fields: A unified framework for neural fields and beyond.arXiv preprint arXiv:2302.01226, 2023

    Anpei Chen, Zexiang Xu, Xinyue Wei, Siyu Tang, Hao Su, and Andreas Geiger. Factor fields: A unified framework for neural fields and beyond.arXiv preprint arXiv:2302.01226, 2023

    arXiv 2023

  7. [7]

    XGBoost: A scalable tree boosting system

    Tianqi Chen and Carlos Guestrin. XGBoost: A scalable tree boosting system. InProceed- ings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, pages 785–794, 2016

    work page 2016

  8. [8]

    Double/debiased machine learning for treatment and struc- tural parameters.The Econometrics Journal, 21(1):C1–C68, 2018

    Victor Chernozhukov, Denis Chetverikov, Mert Demirer, Esther Duflo, Christian Hansen, Whitney Newey, and James Robins. Double/debiased machine learning for treatment and struc- tural parameters.The Econometrics Journal, 21(1):C1–C68, 2018. doi: 10.1111/ectj.12097

    doi:10.1111/ectj.12097 2018

Show all 38 references
  1. [9]

    Empirical properties of asset returns: Stylized facts and statistical issues.Quanti- tative Finance, 1(2):223–236, 2001

    Rama Cont. Empirical properties of asset returns: Stylized facts and statistical issues.Quanti- tative Finance, 1(2):223–236, 2001

    2001

  2. [10]

    Friedman

    Jerome H. Friedman. Greedy function approximation: A gradient boosting machine.Annals of Statistics, 29(5):1189–1232, 2001

    2001

  3. [11]

    A survey on concept drift adaptation.ACM Computing Surveys, 46(4):1–37, 2014

    Jo ˜ao Gama, Indr˙e ˇZliobait˙e, Albert Bifet, Mykola Pechenizkiy, and Abdelhamid Bouchachia. A survey on concept drift adaptation.ACM Computing Surveys, 46(4):1–37, 2014

    2014

  4. [12]

    Revisiting deep learning models for tabular data

    Yury Gorishniy, Ivan Rubachev, Valentin Khrulkov, and Artem Babenko. Revisiting deep learning models for tabular data. InAdvances in Neural Information Processing Systems, volume 34, 2021

    2021

  5. [13]

    On embeddings for numerical features in tabular deep learning

    Yury Gorishniy, Ivan Rubachev, and Artem Babenko. On embeddings for numerical features in tabular deep learning. InAdvances in Neural Information Processing Systems, volume 35, 2022

    2022

  6. [14]

    L ´eo Grinsztajn, Edouard Oyallon, and Ga ¨el Varoquaux. Why do tree-based models still out- perform deep learning on typical tabular data? InAdvances in Neural Information Processing Systems, Datasets and Benchmarks Track, volume 35, 2022

    2022

  7. [15]

    Empirical asset pricing via machine learning

    Shihao Gu, Bryan Kelly, and Dacheng Xiu. Empirical asset pricing via machine learning. Review of Financial Studies, 33(5):2223–2273, 2020

    2020

  8. [16]

    Harvey, Yan Liu, and Heqing Zhu

    Campbell R. Harvey, Yan Liu, and Heqing Zhu. . . . and the cross-section of expected returns. Review of Financial Studies, 29(1):5–68, 2016

    2016

  9. [17]

    Accurate predictions on small data with a tabular foundation model.Nature, 637:319–326, 2025

    Noah Hollmann, Samuel M ¨uller, Lennart Purucker, Arjun Krishnakumar, Max K¨orfer, Shi Bin Hoo, Robin Tibor Schirrmeister, and Frank Hutter. Accurate predictions on small data with a tabular foundation model.Nature, 637:319–326, 2025

    2025

  10. [18]

    Jacobs, Michael I

    Robert A. Jacobs, Michael I. Jordan, Steven J. Nowlan, and Geoffrey E. Hinton. Adaptive mixtures of local experts.Neural Computation, 3(1):79–87, 1991

    1991

  11. [19]

    Jordan and Robert A

    Michael I. Jordan and Robert A. Jacobs. Hierarchical mixtures of experts and the EM algo- rithm.Neural Computation, 6(2):181–214, 1994

    1994

  12. [20]

    LightGBM: A highly efficient gradient boosting decision tree

    Guolin Ke, Qi Meng, Thomas Finley, Taifeng Wang, Wei Chen, Weidong Ma, Qiwei Ye, and Tie-Yan Liu. LightGBM: A highly efficient gradient boosting decision tree. InAdvances in Neural Information Processing Systems, volume 30, 2017. 12

    2017

  13. [21]

    Kelly, Seth Pruitt, and Yinan Su

    Bryan T. Kelly, Seth Pruitt, and Yinan Su. Characteristics are covariances: A unified model of risk and return.Journal of Financial Economics, 134(3):501–524, 2019

    2019

  14. [22]

    K ¨unsch

    Hans R. K ¨unsch. The jackknife and the bootstrap for general stationary observations.Annals of Statistics, 17(3):1217–1241, 1989

    1989

  15. [23]

    Machine learning in the Chinese stock mar- ket.Journal of Financial Economics, 145(2):64–82, 2022

    Markus Leippold, Qian Wang, and Wenyu Zhou. Machine learning in the Chinese stock mar- ket.Journal of Financial Economics, 145(2):64–82, 2022

    2022

  16. [24]

    John Wiley & Sons, 2018

    Marcos L ´opez de Prado.Advances in Financial Machine Learning. John Wiley & Sons, 2018

    2018

  17. [25]

    Accurate intelligible models with pairwise interactions

    Yin Lou, Rich Caruana, Johannes Gehrke, and Giles Hooker. Accurate intelligible models with pairwise interactions. InProceedings of the 19th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, pages 623–631, 2013

    2013

  18. [26]

    When do neural nets outperform boosted trees on tabular data? InAdvances in Neural Information Processing Systems, Datasets and Bench- marks Track, volume 36, 2023

    Duncan McElfresh, Sujay Khandagale, Jonathan Valverde, Vishak Prasad C, Ganesh Ramakr- ishnan, Micah Goldblum, and Colin White. When do neural nets outperform boosted trees on tabular data? InAdvances in Neural Information Processing Systems, Datasets and Bench- marks Track, v...

    2023

  19. [27]

    Barry L. Nelson. On control variate estimators.Computers & Operations Research, 14(3): 219–225, 1987. doi: 10.1016/0305-0548(87)90024-4

    1987 doi

  20. [28]

    Politis and Joseph P

    Dimitris N. Politis and Joseph P. Romano. The stationary bootstrap.Journal of the American Statistical Association, 89(428):1303–1313, 1994

    1994

  21. [29]

    CatBoost: Unbiased boosting with categorical features

    Liudmila Prokhorenkova, Gleb Gusev, Aleksandr V orobev, Anna Veronika Dorogush, and An- drey Gulin. CatBoost: Unbiased boosting with categorical features. InAdvances in Neural Information Processing Systems, volume 31, 2018

    2018

  22. [30]

    Outrageously large neural networks: The sparsely-gated mixture-of-experts layer

    Noam Shazeer, Azalia Mirhoseini, Krzysztof Maziarz, Andy Davis, Quoc Le, Geoffrey Hinton, and Jeff Dean. Outrageously large neural networks: The sparsely-gated mixture-of-experts layer. InInternational Conference on Learning Representations, 2017

    2017

  23. [31]

    Tabular data: Deep learning is not all you need.Infor- mation Fusion, 81:84–90, 2022

    Ravid Shwartz-Ziv and Amitai Armon. Tabular data: Deep learning is not all you need.Infor- mation Fusion, 81:84–90, 2022

    2022

  24. [32]

    David H. Wolpert. Stacked generalization.Neural Networks, 5(2):241–259, 1992. 13 Appendix A Choosing the fields: a frozen selection rule Nothing in §2–§3 sayswhichfactor pairs should become fields. This appendix states the rule we use and the evidence that it orders candidate ...

    1992

  25. [33]

    Fix a scoring window strictly before evaluation (here 2020–2022, inside the 2022-12-31 freeze of §4)

    2020

  26. [34]

    For every candidate triple, restrict to the rows on which all three fields aresimultaneously observable, so that triples are scored on a common sample

  27. [35]

    On that aligned sample, compute the standard deviation of the triple’s field error,σ STD = sd r− ¯F

  28. [36]

    Lower is better

    Rank triples byσ STD ascending and take the low end. Lower is better

  29. [37]

    No triple is refitted, reweighted, or re-scored after the evaluation window opens

    Freeze the choice. No triple is refitted, reweighted, or re-scored after the evaluation window opens. The statistic is deliberately crude: one number per triple, no learner, no tuning constant, and no forward return after 2022. Its only job is toordercandidates. Validation des...

    2022

  30. [2026]

    Both the STD and the single-field MSE columns are offset by a constant to expose their true spread—the entire population of triples spans0.43bps of STD on a base of615bps. selector, 2020–2022 realized, 2023–2026 Decile STD−615single MSE−378kcomplementarity Sharpe Net% ICIR (bp...

    2020

Pith tools

Reviewed August 10, 2026 · model on record in the stance chip above.