Pith. sign in

REVIEW 2 major objections 6 minor 29 references

One screening-off condition on drivers yields a closed-form projected Markowitz portfolio and a diagonal-plus-low-rank risk structure.

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

2026-07-11 07:18 UTC pith:DKOIBBFJ

load-bearing objection Clean static theory: one CI condition yields diagonal-plus-low-rank risk, projected Markowitz, uniqueness/invariance, and first-order ε-bounds; novelty is real but bounded by the classical factor limit the paper itself states. the 2 major comments →

arxiv 2607.05320 v1 pith:DKOIBBFJ submitted 2026-07-06 q-fin.PM

Causal Separation, Conditional Risk, and Projected Markowitz Portfolios

Alejandro Rodriguez Dominguez This is my paper
classification q-fin.PM MSC 91G1062P0590C20
keywords causal separationprojected Markowitzconditional riskscreening-offdiagonal-plus-low-rank covarianceHansen-Jagannathan boundportfolio selectionapproximate separation
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.

Mean-variance portfolio choice is solvable but fragile because estimation error in the mean and the inverse covariance is amplified. This paper answers a prior question: conditional on what information should those moments be computed? It formalizes causal separation: if a declared set of drivers screens off the assets so that, given the drivers' realized path through the investment horizon, asset returns are mutually independent, then the entire static mean-variance theory follows. The condition forces a diagonal-plus-low-rank conditional covariance via an exact tower decomposition, with the low-rank block identified as the assets' response to driver innovations. The constrained mean-variance problem then admits a closed-form projected Markowitz solution in which the classical information matrix is replaced by its projection onto the constraint-compatible subspace. The paper also proves uniqueness of the minimal sufficient separator as an information set, invariance under reparametrization, a conditional efficient frontier, an exact Hansen-Jagannathan gap priced by shadow prices of the constraints, regularization by the idiosyncratic variance floor, and first-order sensitivity bounds under approximate separation. The causal content is proved under an explicit structural margin rather than assumed: common causes form a separator, observational data identify their realized information only, and interventions distinguish true causes from equally fitting correlational proxies. Seven reproducible experiments check the identities at machine precision and measure estimation risk, robustness, intervention invariance, and scaling to thousands of assets.

Core claim

A single structural condition—causal separation, mutual independence of asset returns given the drivers' realized closed-horizon path, plus screening of the asset past—induces the complete static portfolio theory that follows from it: a diagonal-plus-low-rank decision-node covariance via tower decomposition, a unique closed-form projected Markowitz solution, uniqueness of the minimal sufficient separator as a sigma-algebra, invariance of all derived objects, a conditional efficient frontier, an exact Hansen-Jagannathan gap, regularization by the idiosyncratic floor, and first-order sensitivity under approximate separation at a certified tolerance.

What carries the argument

Causal separation (screening-off on the horizon-closed driver path together with asset-past screening S0), which forces the tower decomposition of the decision-node covariance into a diagonal idiosyncratic block plus a low-rank response-to-innovations block, and thereby the projected Markowitz solution in which the information matrix is replaced by its orthogonal projection onto the constraint-compatible subspace.

Load-bearing premise

That every common cause of the assets' horizon returns is already among the declared drivers and that returns do not cause each other or feed back into the drivers inside the investment horizon; if fire-sale contagion or latent common causes operate inside the window, separation and its certificates fail.

What would settle it

On a controlled panel where the true common drivers are known, inject residual cross-correlation of size epsilon into the residual block (or intervene by destroying a proxy's link to a true cause while preserving its marginal law) and check whether the first-order displacement formulas, the dependence floor, and the invariance of the causal separator hold at the predicted magnitudes while a correlational alternative collapses.

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

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

2 major / 6 minor

Summary. The paper takes a single structural condition—causal separation: mutual conditional independence of asset returns given a declared driver set’s realized path through the investment horizon, plus asset-past screening (S0)—and derives the static mean–variance theory it induces. Separation yields, via an exact tower decomposition, a diagonal-plus-low-rank decision-node covariance; the constrained program admits a closed-form projected Markowitz solution in which Q⁻¹ is replaced by its projection onto the constraint-compatible subspace. The paper proves uniqueness of the minimal sufficient separator as a σ-algebra, invariance under separator equivalence and driver reparametrization, a conditional efficient frontier, an exact Hansen–Jagannathan gap in the shadow prices of the geometry, regularization by the idiosyncratic floor, an O(n(m+p)²) two-stage solver, and first-order sensitivity under approximate separation. Causal semantics (soundness, identification, interventional invariance, dependence floor of latent confounding) are stated under an explicit structural margin A4 and delimited as optional for the portfolio theory. Seven reproducible synthetic experiments check identities at machine precision and quantify estimation risk, robustness, intervention invariance, and scaling.

Significance. If the results hold, the paper supplies a clean, self-contained static theory that answers a prior question the covariance-regularization literature leaves open: conditional on what information should moments be computed? The selection theory (existence, uniqueness as information set, invariance), the exact tower identification of the low-rank block as response to driver innovations, the projected solution with explicit HJ gap and conditioning floor, and the certified-tolerance sensitivity bounds are genuine contributions. Strengths that should count in the assessment include: (i) machine-precision verification of five structural identities (Table 3); (ii) a fully reproducible package with fixed seeds and a pre-specified field protocol; (iii) honest delimitation of the causal package under A4 and of the synthetic scope; (iv) an exact linear-cost solver that never forms the n×n matrix. The classical-factor-model limit is acknowledged rather than hidden. The work is complementary to, and narrower than, the author’s prior sensitivity/PDE-control line, isolating the conditioning condition itself.

major comments (2)
  1. [Abstract; §2.4, Assumption 2.16] The abstract and title lead with “causal separation,” but the portfolio theory of Sections 3–5 requires only the observational screening-off property (Def. 2.4) and regularity A1–A3; the causal package (Props. 2.17–2.20, Thm. 2.19) needs the structural margin A4, especially sink-node returns with no within-horizon contagion (A4(ii)) and completeness of the declared universe for common causes (A4(iv)). The body already delimits this carefully (Remark 2.21). The abstract and introduction should state the same split up front—observational portfolio theory vs. optional causal semantics under A4—so that readers do not over-read the causal claim when fire-sale or feedback channels are present at the horizon.
  2. [§1.1; §6.3, Tables 1–2; §6.7, Table 4 and Fig. 5] When the factor count is correctly specified, PCA-m is numerically indistinguishable from the structured estimator (Table 1; classical-limit discussion in §1.1). The practical content of the framework therefore rests on three points the paper already has but under-emphasizes: (a) selection without choosing a factor count, with minimality restored by backward elimination (Remark 2.9, Table 4); (b) the dependence-floor diagnostic that detects latent confounding rather than absorbing it (Prop. 2.20, Table 4); (c) interventional distinguishability of causal vs. correlational separators of equal fit (Thm. 2.19, E6/Fig. 5). Elevate these three as the empirical value proposition early in §6, and state explicitly that agreement with correctly-specified PCA is the intended classical limit, not a null result.
minor comments (6)
  1. [§6.6, Fig. 4; Remark 5.3] Fig. 4 (E5): the worst-case bound exceeds realized weight displacement by a factor of ~355 at ε=0.1. A short sentence on when the first-order formula (Prop. 5.1) should be preferred as the working diagnostic over Cor. 5.2 would help practitioners; the paper already notes the conservatism but could be more directive.
  2. [Table 2] Table 2: the rebalanced oracle Sharpe of 11.8 is flagged as a simulator artifact and ratios are scale-free, but a one-line note in the table caption that absolute Sharpe levels are not interpretable would reduce misreading.
  3. [§6.2, Fig. 1] E1 uses Fisher-z tests on generated regressors (η̂). The caveat is stated; a pointer that the generalised covariance measure of Shah–Peters is the formal replacement (already cited) could be moved from the text into the figure caption for visibility.
  4. [§3.1, Eq. (8)] Notation: G^{+h}_t and G_t are introduced in Def. 2.11; a one-line reminder at the start of §3 that the tower is over G_t ⊆ G^{+h}_t would help readers who skip the selection section.
  5. [Title page; Def. 2.4] Typographical: “Rodríguez Domínguez” appears with varying accent encoding in the author line and self-citations; unify. Also “Sept(A; δ, H)” vs. “Sep_t(A)”—pick one subscript convention.
  6. [Appendix A, Table 5] Section A Table 5 is useful; add the dependence functional used in each experiment (max residual correlation is the working choice of §5) so the configuration is fully self-contained without the code.

Circularity Check

2 steps flagged

No significant circularity in the derivation chain: portfolio objects follow from the CI separation condition plus regularity via standard tower/KKT arguments; self-citations are contextual and experiments are openly confirmatory on a matching DGP.

specific steps
  1. self citation load bearing [§1.1 Related literature (pp. 3–4)]
    "The closest line of work is the causal, sensitivity-based portfolio optimization framework of Rodríguez Domínguez. In that framework, drivers of asset and portfolio dynamics are defined and optimally selected through the commonality principle... The present paper is complementary to this framework and narrower in scope: it isolates the probabilistic condition that underlies conditioning on common causal drivers..."

    The paper positions its separation condition as the primitive underlying the author’s own prior causal-portfolio series ([25]–[28]). While the static derivations themselves are self-contained and do not invoke those papers as theorems, the framing of the ‘closest line’ and the claim that the present work supplies ‘portfolio-level mathematics for that agenda’ rests on self-citation for motivation and continuity; this is minor and non-load-bearing for the proved claims.

  2. fitted input called prediction [§6 Experiments (esp. E3 Table 3; E1–E7 design)]
    "Seven experiments validate the theory, with five structural identities holding at machine precision... The data-generating process is the normal form (10)... Table 3 verifies, on randomly generated instances, the exact statements of Section 4... Residuals of order 10^{-11} to 10^{-13} are floating-point zeros."

    The DGP is constructed to satisfy the linear-Gaussian normal form (10) and the separation assumptions by design; consequently the structural identities (HJ gap, two-stage = dense KKT, (P1)–(P4), gauge invariance, conditioning floor) hold exactly and the reported machine-precision residuals are confirmatory code checks rather than independent empirical tests. The paper labels them as such, so the circularity is mild and acknowledged.

full rationale

The load-bearing claims (Thm 2.14 uniqueness of minimal sufficient separator as a σ-algebra; Prop 3.1/3.5 factorization and diagonal-plus-low-rank Qt via the two-window tower; Thm 4.1 projected Markowitz with M the projected information matrix; Prop 5.1 first-order sensitivity) are proved from the primitives (Defs 2.4/2.7, A1–A3, graphoid intersection) using only conditional-independence calculus, the tower property, and KKT/Woodbury algebra. A4 is explicitly scoped to the optional causal-soundness package (Props 2.17–2.20) and is not used for the portfolio theory. The classical factor-model limit is acknowledged rather than rebranded. Self-citations to the author’s prior causal-portfolio line ([25]–[28]) appear in the related-work discussion of the broader dynamic/sensitivity framework; they supply motivation and context but are not invoked as external uniqueness theorems that force the present static results. Experiments E1–E7 are synthetic by design on the normal form (10) that satisfies the assumptions, so structural identities hold at machine precision by construction; the paper correctly labels them mathematical verification and code checks, not independent market discovery. No equation reduces to a fitted parameter renamed as a prediction, and no uniqueness is imported from prior self-work. Score 2 reflects only the minor, non-load-bearing self-citation pattern plus openly confirmatory experiments.

Axiom & Free-Parameter Ledger

4 free parameters · 6 axioms · 2 invented entities

The theory is conditional on a declared finite driver universe and on screening-off (plus S0). Mathematical graphoid/intersection and standard KKT/convexity are standard. Causal claims rest on structural margin A4 (no contagion, rich universe, no colliders). Regularity A3 (C¹ responses, full rank B, idiosyncratic floor ς²>0, innovation expansion) is domain modeling. Free parameters are experiment/DGP knobs and the choice of dependence functional, not fitted constants inside the theorems. No new physical entities; ‘separator’ and ‘operative σ-algebra’ are definitional information sets.

free parameters (4)
  • dependence functional dep (and ε threshold)
    Choice among residual correlation, GCM, HSIC, CMI normalizes εt(D); theory is relative to the declared test class (Def. 2.7). Working experiments use max |residual correlation|.
  • penalty κ in penalized selection
    Optional ˆεt(D)+κ|D| breaks ties; field protocol sets κ=0.01. Affects which minimal attainer is returned, not the exact-separation theorems.
  • DGP/experiment knobs (ϕ, Bij scale, ςi range, γ, premium θ)
    Table 5: ϕ=0.95, Bij~N(0,0.04²), ςi~U[0.8%,2%], γ=0.05/3, premium scales 0.12/0.25/0.035. Control synthetic validation only; not parameters of the theorems.
  • inadmissible-motion subspace Ut (geometry constraints Ct)
    Modeling declaration of untrusted driver directions (§3.3); q can be zero. Changes feasible set and shadow prices but is user-chosen geometry, not estimated from the central claim.
axioms (6)
  • domain assumption A1: candidate driver set {1..M} finite and declared ex ante; optimality relative to it
    Assumption 2.2; all selection statements are relative to a pre-declared universe.
  • standard math A2: intersection property (graphoid) for the windowed conditional independences used
    Assumption 2.3; used for uniqueness of minimal sufficient separator (Thm. 2.14). Sufficient condition: positive joint density on finite-dimensional window functionals.
  • domain assumption A3: C¹ response maps, rank Bt=m, mini ςi²≥ς²>0, positive driver innovation cov Λt, first-order innovation expansion of closed-window means
    Assumption 3.4; needed for Qt structure (9), conditioning floor Prop. 4.5, and first-order statements.
  • domain assumption A4 structural margin: SCM Markov to acyclic graph; returns are sinks (no contagion/feedback); asset-specific exogenous noise; all common causes in declared universe; no collider drivers in C
    Assumption 2.16; load-bearing only for causal soundness/identification/intervention results in §2.4, not for pure observational portfolio math in §§3–5.
  • ad hoc to paper Screening-off Def. 2.4: mutual CI of returns given F^D_[t-δ,t+h] plus S0 (asset past informationally inert given driver path)
    The paper’s single structural condition; everything portfolio-theoretic is derived from it. Motivated by Reichenbach but not assumed as a market axiom—work is at achievable ε*.
  • standard math Q≻0, rank A=1+q; portfolios Gt-measurable; W convex closed budget set
    Standard KKT/strict convexity setup for Thm. 4.1.
invented entities (2)
  • Separator / ε-separator and operative σ-algebras Gt, G+h_t independent evidence
    purpose: Name the information sets that screen off residual dependence and on which decisions vs residual risk are conditioned
    Definitional packaging of conditional independence; not a new physical object. Independent evidence is the CI tests and intervention experiment on synthetic SCMs.
  • Projected Markowitz operator M and frontier potential Δt independent evidence
    purpose: Closed-form solution and efficient-frontier potential under linear geometry constraints
    Standard projection of Q⁻¹ onto ker(A Q^{-1/2}); classical object under a new selection of Q. Not an invented force or particle.

reviewed 2026-07-11 · how reviews work

0 comments
Cite this review

Pith. "Pith review of Causal Separation, Conditional Risk, and Projected Markowitz Portfolios." pith.science (2026). https://pith.science/paper/DKOIBBFJ

@misc{pith2026260705320,
  author       = {Pith},
  title        = {Pith review of: Causal Separation, Conditional Risk, and Projected Markowitz Portfolios},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/DKOIBBFJ}},
  note         = {Machine review of arXiv:2607.05320}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

We formalize a single structural condition on a portfolio problem, causal separation: conditional on the realized path of a declared set of drivers through the investment horizon, asset returns are mutually independent. From this condition we derive the complete static portfolio theory it induces. Separation forces a diagonal-plus-low-rank conditional covariance through an exact tower decomposition, with the low-rank block identified as the response to driver innovations, and the constrained mean-variance problem admits a closed-form projected Markowitz solution in which the classical information matrix is replaced by its projection onto the constraint-compatible subspace. We prove uniqueness of the minimal sufficient separator as an information set; invariance of all derived objects under separator equivalence and reparametrization of the driver state; a conditional efficient-frontier theorem; a Hansen-Jagannathan bound whose gap to the unconstrained bound is an exact quadratic form in the shadow prices of the constraint geometry; a conditioning bound showing that separation regularizes estimation through the idiosyncratic variance floor; and exact first-order sensitivity bounds under approximate separation at a certified tolerance. The causal content is stated and proved rather than assumed: under an explicit structural margin the common causes form a separator, observational data identify their realized information and no more, and causal and correlational separators of equal fit are distinguished by interventions on non-parents. Seven reproducible experiments validate the theory, with structural identities holding at machine precision, and quantify its practical content against sample, shrinkage and principal-component covariances, including robustness, intervention invariance, and scaling to thousands of assets.

Figures

Figures reproduced from arXiv: 2607.05320 by Alejandro Rodriguez Dominguez.

Figure 1
Figure 1. Figure 1: Experiment E1: conditional factorization and the two windows. Left: residual correlation [PITH_FULL_IMAGE:figures/full_fig_p018_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Experiment E2: estimation risk across n at T = 252. Left: OOS annualized GMV volatility (log scale). Right: variance calibration; the structured estimator stays on the dashed line [PITH_FULL_IMAGE:figures/full_fig_p021_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Experiment E4: predicted versus realized conditional Sharpe ratio (daily rebalancing, [PITH_FULL_IMAGE:figures/full_fig_p022_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Experiment E5: approximate separation. Actual displacement (dots), exact first-order [PITH_FULL_IMAGE:figures/full_fig_p023_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Experiment E6: intervention on a non-parent. OOS annualized Sharpe ratio (left) [PITH_FULL_IMAGE:figures/full_fig_p024_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Experiment E7: exact two-stage scaling. Seconds per solve (log–log scale). Maximum [PITH_FULL_IMAGE:figures/full_fig_p024_6.png] view at source ↗

Review history (2 revisions) →

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

29 extracted references · 2 linked inside Pith

  1. [1]

    Best, R.R

    M.J. Best, R.R. Grauer. On the sensitivity of mean-variance-efficient portfolios to changes in asset means.Review of Financial Studies4(2):315–342, 1991

    1991

  2. [2]

    S. Boyd, L. Vandenberghe.Convex Optimization. Cambridge University Press, 2004

    2004

  3. [3]

    Chamberlain, M

    G. Chamberlain, M. Rothschild. Arbitrage, factor structure, and mean-variance analysis on large asset markets.Econometrica51(5):1281–1304, 1983

    1983

  4. [4]

    DeMiguel, L

    V. DeMiguel, L. Garlappi, R. Uppal. Optimal versus naive diversification: How inefficient is the 1 /N portfolio strategy?Review of Financial Studies22(5):1915–1953, 2009

    1915

  5. [5]

    R.F. Engle. Dynamic conditional correlation: A simple class of multivariate generalized autoregressive conditional heteroskedasticity models.Journal of Business & Economic Statistics20(3):339–350, 2002. 26

    2002

  6. [6]

    J. Fan, Y. Fan, J. Lv. High dimensional covariance matrix estimation using a factor model.Journal of Econometrics147(1):186–197, 2008

    2008

  7. [7]

    Hansen, R

    L.P. Hansen, R. Jagannathan. Implications of security market data for models of dynamic economies. Journal of Political Economy99(2):225–262, 1991

    1991

  8. [8]

    Hansen, S.F

    L.P. Hansen, S.F. Richard. The role of conditioning information in deducing testable restrictions implied by dynamic asset pricing models.Econometrica55(3):587–613, 1987

    1987

  9. [9]

    Jagannathan, T

    R. Jagannathan, T. Ma. Risk reduction in large portfolios: Why imposing the wrong constraints helps. Journal of Finance58(4):1651–1683, 2003

    2003

  10. [10]

    Garlappi, R

    L. Garlappi, R. Uppal, T. Wang. Portfolio selection with parameter and model uncertainty: A multi-prior approach.Review of Financial Studies20(1):41–81, 2007

    2007

  11. [11]

    Kallenberg.Foundations of Modern Probability, 2nd ed

    O. Kallenberg.Foundations of Modern Probability, 2nd ed. Springer, 2002

    2002

  12. [12]

    R. Kan, G. Zhou. Optimal portfolio choice with parameter uncertainty.Journal of Financial and Quantitative Analysis42(3):621–656, 2007

    2007

  13. [13]

    Kelly, S

    B. Kelly, S. Pruitt, Y. Su. Characteristics are covariances: A unified model of risk and return.Journal of Financial Economics134(3):501–524, 2019

    2019

  14. [14]

    Ledoit, M

    O. Ledoit, M. Wolf. A well-conditioned estimator for large-dimensional covariance matrices.Journal of Multivariate Analysis88(2):365–411, 2004

    2004

  15. [15]

    Markowitz

    H. Markowitz. Portfolio selection.Journal of Finance7(1):77–91, 1952

    1952

  16. [16]

    R.C. Merton. An intertemporal capital asset pricing model.Econometrica41(5):867–887, 1973

    1973

  17. [17]

    R.O. Michaud. The Markowitz optimization enigma: Is ‘optimized’ optimal?Financial Analysts Journal45(1):31–42, 1989

    1989

  18. [18]

    L´ opez de Prado.Causal Factor Investing: Can Factor Investing Become Scientific?Cambridge Elements in Quantitative Finance, Cambridge University Press, 2023

    M. L´ opez de Prado.Causal Factor Investing: Can Factor Investing Become Scientific?Cambridge Elements in Quantitative Finance, Cambridge University Press, 2023

    2023

  19. [19]

    A. Pagan. Econometric issues in the analysis of regressions with generated regressors.International Economic Review25(1):221–247, 1984

    1984

  20. [20]

    Pearl.Probabilistic Reasoning in Intelligent Systems

    J. Pearl.Probabilistic Reasoning in Intelligent Systems. Morgan Kaufmann, 1988

    1988

  21. [21]

    Pearl.Causality: Models, Reasoning, and Inference, 2nd ed

    J. Pearl.Causality: Models, Reasoning, and Inference, 2nd ed. Cambridge University Press, 2009

    2009

  22. [22]

    Peters, P

    J. Peters, P. B¨ uhlmann, N. Meinshausen. Causal inference by using invariant prediction: identification and confidence intervals.Journal of the Royal Statistical Society B78(5):947–1012, 2016

    2016

  23. [23]

    Peters, D

    J. Peters, D. Janzing, B. Sch¨ olkopf.Elements of Causal Inference. MIT Press, 2017

    2017

  24. [24]

    Reichenbach.The Direction of Time

    H. Reichenbach.The Direction of Time. University of California Press, 1956

    1956

  25. [25]

    Rodr´ ıguez Dom´ ınguez

    A. Rodr´ ıguez Dom´ ınguez. Portfolio optimization based on neural networks sensitivities from assets dynamics respect common drivers.Machine Learning with Applications11:100447, 2023

    2023

  26. [26]

    Rodr´ ıguez Dom´ ınguez

    A. Rodr´ ıguez Dom´ ınguez. A portfolio’s common causal conditional risk-neutral PDE. In M. Corazza, F. Gannon, F. Legros, C. Pizzi, V. Touz´ e (eds.),Mathematical and Statistical Methods for Actuarial Sciences and Finance (MAF 2024). Springer, Cham, 2024

    2024

  27. [27]

    Rodr´ ıguez Dom´ ınguez

    A. Rodr´ ıguez Dom´ ınguez. Causal portfolio optimization: principles and sensitivity-based solutions. arXiv:2504.05743, 2025. 27

    Pith/arXiv arXiv 2025

  28. [28]

    Rodr´ ıguez Dom´ ınguez

    A. Rodr´ ıguez Dom´ ınguez. Causal PDE-control for adaptive portfolio optimization under partial information. In A. Itkin, O. Kondratyev (eds.),Reviews in Modern Quantitative Finance, vol. 2. World Scientific, forthcoming; arXiv:2509.09585, 2025

    Pith/arXiv arXiv 2025

  29. [29]

    R.D. Shah, J. Peters. The hardness of conditional independence testing and the generalised covariance measure.Annals of Statistics48(3):1514–1538, 2020. 28

    2020

This paper was first reviewed by grok-4.5 on July 11, 2026.