Managing Portfolios Across the Return Distribution

Attila Sarkany; Jozef Barunik; Lukas Janasek

REVIEW 4 major objections 5 minor 26 references

Managing Portfolios Across the Return Distribution

T0 review · 4 major / 5 minor · reviewed 2026-08-04 · deepseek-v4-flash

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

Pith's one-line read The paper establishes that investors who maximize a chosen quantile of discounted returns—low τ for downside, high τ for upside—generate state-contingent policies that de-risk in volatile states, making volatility management an optimal resp

desk verdict Good theory, broken empirics: the contraction/regime example are worth a look, but the actor update has a sign error and the 'out-of-sample' results are actually in-sample — the paper as written shouldn't be used as evidence. read the letter →

arxiv 2510.19271 v2 pith:43MY3D6W submitted 2025-10-22 q-fin.GN

Jozef Barunik , Lukas Janasek , Attila Sarkany This is my paper · ORCID

classification q-fin.GN MSC 91G1090C4068T05

keywords quantilepreferencesdynamicportfoliochoicevolatilitymanagementdistributionalreinforcementlearningtailriskregimeswitchinginverse-variancescalingactor-critic

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 proposes that portfolio choices should be driven by which part of the return distribution an investor wants to improve, encoded in a single number τ. It shows that maximizing the τ-quantile of discounted returns—rather than expected utility—produces policies that are naturally state-dependent: a downside-focused (low-τ) investor cuts risky exposure in high-volatility states, while an upside-focused (high-τ) investor concentrates in high-dispersion assets. The paper demonstrates that this behavior is not a heuristic: volatility management falls out of quantile preferences when volatility evolves in regimes. It also introduces a distributional reinforcement-learning algorithm that learns such policies without specifying a utility function, and reports out-of-sample frontiers where low-τ portfolios deliver left-tail protection while high-τ portfolios earn the highest mean return. If right, the framework gives a preference-based rationale for inverse-variance timing and a way to tailor portfolios to tail preferences.

What carries the argument

The load-bearing machinery is a quantile recursion operator: for a fixed τ, the value function is defined as the τ-quantile of current reward plus the discounted next-period value, conditional on the current state. The paper shows this operator is a contraction on bounded functions, so it has a unique fixed point and iterative learning converges. In portfolio problems, the operator generates nested conditional quantiles when the next state is random (volatility regimes), and this nesting is what turns corner solutions into interior, state-contingent policies. The learning algorithm pairs this quantile value estimator (the critic) with a stochastic policy over portfolio weights (the actor) th

What would settle it

Set the two-period regime model so the high-volatility state is absorbing (zero probability of transitioning to low volatility) and compute the optimal risky share for a τ=0.1 investor. The theory's left-tail-flattening mechanism predicts a corner allocation of all risk-free; any positive interior risky share in that calibration would falsify the mechanism.

Watch

Extended reading notes

Core claim

The paper's central claim is that quantile preferences—maximizing the τ-quantile of discounted returns rather than expected utility—provide a preference-based microfoundation for volatility management. In the static one-period problem, a τ-investor follows an all-or-nothing rule: hold the risky asset only if its τ-quantile return beats the risk-free rate. The result changes once returns are conditioned on a volatility regime process: the next-period regime is random, so the decision involves nested conditional quantiles, and the corner solution breaks. A downside-focused (low-τ) investor in a high-volatility state optimally holds an interior mixed portfolio, because a small risky allocation

Load-bearing premise

The interior, de-risking result depends on volatility following an exogenous, action-independent regime process with known transition probabilities; if the investor's own trades change future volatility, or the regime process is unknown, the proof that low-τ investors should hold a mixed risky position in high-volatility states no longer applies.

Editorial extensions

If this is right

Volatility-managed portfolios—scaling exposure by inverse variance—are not a puzzle but the equilibrium behavior of a downside-focused investor; the same logic predicts when de-risking will be strongest.
The quantile index τ is a complete, utility-free description of risk attitude for dynamic portfolio choice, so one algorithm serves investors with different tail preferences.
Because policies adjust composition as well as scale, quantile targeting gives a principled way to tilt factor or sector exposures toward a client's downside or upside priorities.
Tail-adjusted performance measures (CVaR- or modified-VaR-based ratios) are the right evaluation metrics for these portfolios; rankings based on symmetric volatility can invert the ordering.
Transaction costs have a τ-specific impact, with downside-protective policies most sensitive to turnover costs, so cost-aware mandate design should interact with tail preferences.

Reading between the lines

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

An empirical implication the paper leaves untested: the optimal de-risking response should grow with the persistence of the high-volatility regime and with the probability of a transition to calm; comparing estimated policies across regimes with different transition matrices would test this.
The same quantile-learning machinery could be adapted to other tail-sensitive objectives—expected shortfall, CVaR, or prospect-theory distortions—by changing the loss weights, making τ one point on a broader spectrum of distributional mandates.
If fund-flow patterns reveal investors' τ-mandates, a calibration exercise could recover a distribution of revealed tail preferences across investor clienteles, turning the quantile index into an empirically observable preference parameter.
The theoretical mechanism assumes trades do not affect regime dynamics; with price impact or volatility feedback, the optimal interior allocation would shift, so the de-risking prescription is a partial-equilibrium statement.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Desk Editor's Note

read the letter

Two things you should know before reading: the dynamic-quantile theory and the two-period regime example are genuinely nice, but the empirical section is not what the abstract claims, and Appendix C contradicts the actor update in Section 3.2. The sign error isn't cosmetic.

What's new and good: The paper formulates recursive conditional-quantile preferences for portfolio choice and proves the τ-Bellman operators are β-contractions in sup-norm (Appendix E). That part is correct. The two-period example with Markov volatility regimes shows how nested conditional quantiles break the corner solution and create an interior allocation — a clean intuition for why a downside-focused investor might still hold a risky asset in a high-volatility state. I also think the mapping of τ to risk attitude and the discussion of the static corner rule are clear.

Where it falls apart: First, the actor update in Eq. (26). With θ←θ−η∇Lactor and L=logπ·δ·[(1−τ)1{δ<0}+τ1{δ≥0}], a negative TD error (δ<0) gets positive weight (1−τ), so minimizing L increases logπ for that action. A positive δ gets positive weight τ, so minimizing L decreases logπ. That is the opposite of quantile maximization. The correct gradient in Appendix C, Eq. (63), is E[∇logπ·(τ−1{δ<0})] without the δ factor. So the algorithm as written is not ascending Jτ; the text's interpretation ('minimizing the log-likelihood of actions that result in lower value') contradicts the math. Unless the implementation flips the sign, the learned policies in Section 5 cannot be attributed to τ-quantile optimization.

Second, the abstract promises out-of-sample gains and gains over volatility-managed portfolios, plus fund-flow evidence. The full text says the focus is in-sample evaluation, Algorithm D.1 evaluates on the full training set, no volatility-managed benchmark appears, and I couldn't find the fund-flow analysis. That is a mismatch between promise and delivery.

Third, the main empirical validation of 'preference-consistent learning' uses the critic's own inverse-CDF estimates (Figure H.1, Table F.4). That is circular — the critic was trained to fit those quantiles.

Fourth, there are small inconsistencies between Tables 4 and H.1 for the same cost level, which need reconciling.

Bottom line: the theoretical machinery is worth a referee, but the paper in its current form cannot be accepted as evidence for the empirical claims. I would send it to peer review — not to desk reject — so the authors get a clear path to fix the sign error, do a genuine holdout test, and add the promised benchmarks. A reader working on quantile RL or volatility timing might find the regime example useful, but the empirical results should not be cited.

Referee Report

4 major / 5 minor

Summary. The paper develops a dynamic portfolio-choice framework in which investors maximize a recursive τ-quantile of discounted returns rather than expected utility. The authors prove contraction properties for the quantile Bellman operator, illustrate with a two-period regime-switching example in which low-τ investors de-risk in high-volatility states, propose a distributional actor-critic algorithm (Q-A2C), and report empirical applications on three factor ETFs and ten Fama–French industry portfolios. The central claims are that low-τ policies provide downside protection while high-τ policies earn higher mean returns, forming an ordered frontier, and that this behavior provides a preference-based microfoundation for volatility-managed portfolios.

Significance. If the theoretical and empirical results were valid, the paper would offer a genuinely utility-free way to tailor dynamic portfolios to tail preferences and would rationalize inverse-variance volatility scaling as optimal behavior rather than an ad hoc heuristic. The contraction proofs in Appendix E are correct and self-contained, and the two-period example in Section 2.3.3 is coherent and illustrative. The availability of replication code is also a strength. However, the core algorithmic update appears to have a sign error, and the empirical evaluation is in-sample and circular in its key diagnostic. These issues undermine the central claims as currently presented.

major comments (4)

[§5.1, Algorithm D.1] The actor update as written is not an ascent direction for Jτ(θ). With Lactor = logπ·δ·[(1−τ)1{δ<0}+τ1{δ≥0}] and θ←θ−η∇Lactor, a sample with δ<0 has coefficient (1−τ)δ<0, so minimizing L increases logπ and reinforces that action. A sample with δ>0 has coefficient τδ>0, so minimizing L decreases logπ and suppresses the action. This is the opposite of the intended quantile maximization. The correct gradient from Appendix C is ∇Jτ = E[ψτ(δ)∇logπ] with ψτ(z)=τ−1{z<0}, which is negative for δ<0. The extra factor δ and the displayed weights do not yield an unbiased variance-reduced ascent direction. Consequently Theorem 3.1(iii) does not apply to the implemented algorithm, and the Section 5 policies cannot be interpreted as τ-quantile optimizers.
[§5.1, Algorithm D.1] The abstract states that 'Out of sample, the estimated policies form an ordered frontier,' but §5.1 explicitly says 'our focus is primarily on in-sample evaluation.' Algorithm D.1 performs 'Final evaluation: test πθ on full training set,' and the 15% validation set is used only for early stopping. The performance metrics in Tables 4–6 are therefore in-sample fits, not out-of-sample evidence. This discrepancy is load-bearing for the headline empirical claim and must be resolved either by removing the out-of-sample language or by providing a proper walk-forward evaluation on held-out data.
[Appendix F, Table F.4; Figure H.1] The diagnostic that the learned policies optimize quantiles rather than a mean-variance proxy uses the critic's own inverse-CDF estimates (Table F.4 and Figure H.1). Since the critic is trained jointly with the actor and these quantile estimates are the same targets used in the losses, this is circular: it does not demonstrate that the realized return distributions of the τ-policies have the claimed quantile ordering. The authors should report empirical quantiles of the actual realized portfolio returns (preferably out of sample) instead of the fitted value-function quantiles.
[§2.3.3, Eq. (18)] The microfoundation for volatility management relies on volatility regimes being exogenous, action-independent, and governed by a known Markov chain with known probabilities. The paper elsewhere (Section 2.2, footnote 1; Section 3) emphasizes action-dependent states and model-free learning, but the analytical interior-solution result does not cover endogenous or unknown regime dynamics. The broad conclusions in Sections 1 and 6 that 'volatility management is what a low-τ policy does' are therefore established only for the stylized example. This limitation should be stated explicitly, and the generality claims should be tempered or supported by additional analysis.

minor comments (5)

[Table 6 caption] The caption of Table 6 says 'Summary statistics for minimum volatility, momentum, and quality ETFs,' but the table reports Fama–French industry portfolio results. This is a copy-paste error; the caption should refer to FF10 industry portfolios.
[Figures 4–5] The x-axis starts at 2017 although the sample is stated as June 2016–December 2022. Please relabel to include 2016 or clarify the actual plotted period.
[Introduction] Typo: 'conseqeuently' should be 'consequently' in the introduction.
[References] The reference 'Borkar, V. S. and V. S. Borkar (2008)' duplicates the author name; it should be simply Borkar (2008).
[Algorithm D.1] The final line 'test πθ on full training set' is confusing and inconsistent with any claim of out-of-sample evaluation. If this is intentional, it should be justified; otherwise it should be changed.

Circularity Check

1 steps flagged · score 5.0 of 10

The theoretical microfoundation is self-contained, but the key empirical diagnostic validates the actor using the critic's own fitted quantile estimates, making the 'preference-consistent learning' evidence partly circular.

fitted input called prediction [Section 5.1.3, Appendix H Figure H.1; cf. Eq. (20) and Eq. (26)]
"Finally, Appendix H Figure H.1 shows value distributions that confirm preference-consistent learning. The estimated inverse CDFs of value are ordered across quantiles and exhibit curvature (skewness and heaviness of the tails), indicating that the algorithm is optimising τ-specific objectives rather than a mean-variance proxy across different costs."

The 'estimated inverse CDFs' are outputs of the critic network Vτω(s), the very quantity used to define the actor's training signal: Eq. (20) computes δτ_t = r_{t+1}+βVτω(s_{t+1})−Vτω(s_t), and Eq. (26) uses δτ_t with τ-dependent weights to update the policy. The diagnostic therefore checks whether the learned policy scores well on the critic's own fitted quantile estimates, not on an independent or realized criterion. A policy trained to maximize that same critic-based objective will tend to produce high critic estimates at the targeted τ, so the displayed ordering is a consistency check between two fitted components rather than independent evidence that the true quantile objective is optimized.

full rationale

The theoretical core is largely self-contained: the contraction of the quantile Bellman operators is proven in Appendix E, and the policy-gradient derivation in Appendix C is carried out in the paper rather than imported as a black box. The self-citation to Janasek (2025) is not load-bearing for the central derivation because the gradient is rederived. The two-period volatility-regime example is a numerical consequence of the stated assumptions, not a circular fit. The main circularity concern is empirical: the inverse-CDF diagnostic used to confirm 'preference-consistent learning' relies on the critic's own fitted value estimates, the same object that defines the actor's loss. Additionally, the abstract's 'out of sample' ordered-frontier claim is weakened by Section 5.1's statement that evaluation is primarily in-sample and by Algorithm D.1's final evaluation on the full training set; this is a validity concern adjacent to circularity. A separate correctness issue—the actor update in Eq. (26) appears to have the opposite sign/weighting relative to the derived gradient in Eq. (63)—is outside the circularity pass but further clouds the interpretation of the empirical policies.

Assumptions & free parameters 7 free parameters · 6 assumptions · 0 invented entities

The theory relies on imported quantile-preference axioms and adds an ad hoc exogenous-volatility assumption to generate the central microfoundation. The empirical results depend on several hand-tuned scalings and an unreported transaction-cost setting, and the headline numbers in Tables 3-4 are inconsistent with the robustness tables.

free parameters (7)

σ_H/σ_L ratio = 1.7
Chosen by hand in the two-period numerical example (Section 2.3.3); drives the size of the interior solution (46% allocation to risky asset).
Regime persistence probabilities = 0.7
Chosen by hand for the illustrative two-period example; directly controls the left-tail-flattening effect that generates the interior solution.
Reward scaling = 1221
Hand-tuned preprocessing constant (Table D.1); changes the effective learning signal and therefore the learned portfolio weights.
TD error scaling = 10
Hand-tuned (Table D.1); affects actor-critic loss magnitude and hence exploration/exploitation balance.
Policy Gaussian std (σ) = 0.5
Exploration variance of the Gaussian policy (Table D.1); entropy regularization is disabled, so this controls exploration.
Order loss regularization λ_ω = 5
Monotonicity penalty weight (Table D.1); imposes quantile ordering on the critic.
Main-result transaction cost = unreported
Tables 3-4 report weights/metrics that match none of the cost columns in Table 5/H.1; the cost level used for headline results is not stated.

assumptions (6)

domain assumption Quantile preferences are a valid representation of risk attitudes (Eq. 2-3)
Imported from Manski (1988), Rostek (2010); the paper does not re-derive the axiomatization.
domain assumption Dynamically consistent preferences are given by the recursive conditional-quantile value function (Eq. 5)
From de Castro and Galvao (2019) / de Castro et al. (2025); the paper adopts this without proof.
ad hoc to paper Volatility regimes are exogenous, action-independent, with known Markov transition probabilities (Section 2.3.3)
Modeling assumption for the two-period example that produces the interior solution; not justified empirically.
standard math Returns are bounded or satisfy concentration conditions (Section 2.2)
Needed for the contraction argument; the paper notes empirical returns are unbounded and clips targets.
domain assumption Critic realizability/ergodicity (Assumption 1, Appendix E)
Unproved regularity conditions for convergence of the neural-network critic.
standard math Two-timescale stochastic approximation (Assumption 2, Appendix E)
Standard SA theory; the paper provides only a proof outline.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Managing Portfolios Across the Return Distribution." pith.science (2026). https://pith.science/paper/43MY3D6W

@misc{pith2026251019271,
  author       = {Pith},
  title        = {Pith review of: Managing Portfolios Across the Return Distribution},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/43MY3D6W}},
  note         = {Machine review of arXiv:2510.19271}
}

read the original abstract

We develop a dynamic portfolio-choice framework in which investors target the region of the payoff distribution that the portfolio is designed to improve. Out of sample, the estimated policies form an ordered frontier: the policy focused on the downside delivers the strongest left-tail protection and the highest Sharpe ratio, while the policy focused on the upper quantile earns the highest mean return. The gains over volatility-managed portfolios are concentrated in periods when downside-tail dispersion is high. Evidence from fund flows in income, growth and downside protection products supports the interpretation of the quantile index as a reduced-form mandate measure.

Figures

Figures reproduced from arXiv: 2510.19271 by the authors.

**Figure 1.** Value functions for risk-averse (τ = 0.1), risk-neutral (τ = 0.5), and risk-loving (τ = 0.9) investor with quantile preferences as functions of risky asset portfolio share. Terminal period (t = 1) in red, initial period (t = 0) in blue, high volatility in solid line, and low volatility in dashed line. Crosses indicate optimal allocations. Risk-free rate: Rf = 1.04 (4% net return). Risky asset return follows normal d… view at source ↗

**Figure 2.** Portfolio return distribution when mixing with risky asset in high volatility regime. Top panel compares the 0.1 quantile of the target distribution with risk free asset return in the last period. Bottom panel shows the left tail flattening effect on the target distribution caused by the volatility regime transition in the first period. The volatility regime example illustrates the impact of state dependence on inve… view at source ↗

**Figure 3.** Regime-conditional portfolio weights for [PITH_FULL_IMAGE:figures/full_fig_p021_3.png] view at source ↗

Figures from the paper (4 more)

**Figure 4.** Figure 4: Cumulative returns of optimal τ-quantile portfolios. Cumulative total return indices for portfolios preferred by “risk-averse” (τ = 0.1), “neutral” (τ = 0.5), and “risk-seeking” (τ = 0.9) agents over June 2016 to December 2022 benchmarked to a static equally-weighted (…

**Figure 5.** Figure 5: Effect of transaction costs on cumulative returns of optimal [PITH_FULL_IMAGE:figures/full_fig_p027_5.png]

**Figure 6.** Figure 6: Cumulative returns of optimal τ-quantile portfolios (FF 10 industries). Cumulative total return indices for portfolios preferred by “risk-averse” (τ = 0.1), “neutral” (τ = 0.5), and “risk-seeking” (τ = 0.9), benchmarked to equally weighted (EW) and global minimum-varia…

**Figure 7.** Figure 7: (Average) portfolio weights across τ-quantile preferences (FF 10 industries). Stacked average sector weights for GMV and τ ∈ {0.1, 0.5, 0.9} portfolios formed from the ten Fama–French industry portfolios (weights sum to one). Sectors follow the standard FF10 taxonomy: …

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

26 extracted references · 1 linked inside Pith

[1]

Ang, A. and G. Bekaert (2002). International asset allocation with regime shifts. The review of financial studies\/ 15\/ (4), 1137--1187

2002
[2]

Bellemare, M. G., W. Dabney, and R. Munos (2017). A distributional perspective on reinforcement learning. In International conference on machine learning , pp.\ 449--458. PMLR

2017
[3]

Franzoni, B

Ben-David, I., F. Franzoni, B. Kim, and R. Moussawi (2023). Competition for attention in the etf space. The Review of Financial Studies\/ 36\/ (3), 987--1042

2023
[4]

Borkar, V. S. and V. S. Borkar (2008). Stochastic approximation: a dynamical systems viewpoint , Volume 100. Springer

2008
[5]

Castro, L. d., A. F. Galvao, G. Montes-Rojas, and J. Olmo (2022). Portfolio selection in quantile decision models. Annals of finance\/ 18\/ (2), 133--181

2022
[6]

Chambers, C. P. (2009). An axiomatization of quantiles on the domain of distribution functions. Mathematical Finance: An International Journal of Mathematics, Statistics and Financial Economics\/ 19\/ (2), 335--342

2009
[7]

Rowland, M

Dabney, W., M. Rowland, M. Bellemare, and R. Munos (2018). Distributional reinforcement learning with quantile regression. In Proceedings of the AAAI conference on artificial intelligence , Volume 32

2018
[8]

de Castro, L. and A. F. Galvao (2019). Dynamic quantile models of rational behavior. Econometrica\/ 87\/ (6), 1893--1939

2019

Show all 26 references

[9]

de Castro, L., A. F. Galvao, and D. Nunes (2025). Dynamic economics with quantile preferences. Theoretical Economics\/ 20\/ (1), 353--425

2025
[10]

Mart \' n-utrera, and R

DeMiguel, V., A. Mart \' n-utrera, and R. Uppal (2024). A multifactor perspective on volatility-managed portfolios. The Journal of Finance\/ 79\/ (6), 3859--3891

2024
[11]

Duarte, and D

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

2024
[12]

Guidolin, M. and A. Timmermann (2007). Asset allocation under multivariate regime switching. Journal of Economic Dynamics and Control\/ 31\/ (11), 3503--3544

2007
[13]

Hornik, K. (1991). Approximation capabilities of multilayer feedforward networks. Neural networks\/ 4\/ (2), 251--257

1991
[14]

Song, and H

Huang, S., Y. Song, and H. Xiang (2024). The smart beta mirage. Journal of Financial and Quantitative Analysis\/ 59\/ (6), 2515--2546

2024
[15]

Janasek, L. (2025). Gradient-based reinforcement learning for dynamic quantile. Technical report, Charles University Prague, Faculty of Social Sciences, Institute of Economic

2025
[16]

Peng, and J

Jiang, J., Y. Peng, and J. Hu (2022). Quantile-based policy optimization for reinforcement learning. In 2022 Winter Simulation Conference (WSC) , pp.\ 2712--2723. IEEE

2022
[17]

Sabbatucci, and A

Johansson, A., R. Sabbatucci, and A. Tamoni (2025). Tradable risk factors for institutional and retail investors. Review of Finance\/ 29\/ (1), 103--139

2025
[18]

Ma, X., L. Xia, Z. Zhou, J. Yang, and Q. Zhao (2020). Dsac: Distributional soft actor critic for risk-sensitive reinforcement learning. arXiv preprint arXiv:2004.14547\/

2020 arXiv
[19]

Manski, C. F. (1988). Ordinal utility models of decision making under uncertainty. Theory and Decision\/ 25 , 79--104

1988
[20]

Markowitz, H. (1952). Portfolio selection. The Journal of Finance\/ 7\/ (1), 77--91

1952
[21]

Moreira, A. and T. Muir (2017). Volatility-managed portfolios. The Journal of Finance\/ 72\/ (4), 1611--1644

2017
[22]

Rostek, M. (2010). Quantile maximization in decision theory. The Review of Economic Studies\/ 77\/ (1), 339--371

2010
[23]

Rumelhart, D. E., G. E. Hinton, and R. J. Williams (1988). Neurocomputing: Foundations of research

1988
[24]

Todorov, K. (2024). When passive funds affect prices: evidence from volatility and commodity etfs. Review of Finance\/ 28\/ (3), 831--863

2024
[25]

Tversky, A. and D. Kahneman (1992). Advances in prospect theory: Cumulative representation of uncertainty. Journal of Risk and uncertainty\/ 5 , 297--323

1992
[26]

Wang, S. S. (2000). A class of distortion operators for pricing financial and insurance risks. Journal of risk and insurance\/ , 15--36

2000

Pith tools

Reviewed August 4, 2026 · model on record in the stance chip above.

[1] [1]

Ang, A. and G. Bekaert (2002). International asset allocation with regime shifts. The review of financial studies\/ 15\/ (4), 1137--1187

2002

[2] [2]

Bellemare, M. G., W. Dabney, and R. Munos (2017). A distributional perspective on reinforcement learning. In International conference on machine learning , pp.\ 449--458. PMLR

2017

[3] [3]

Franzoni, B

Ben-David, I., F. Franzoni, B. Kim, and R. Moussawi (2023). Competition for attention in the etf space. The Review of Financial Studies\/ 36\/ (3), 987--1042

2023

[4] [4]

Borkar, V. S. and V. S. Borkar (2008). Stochastic approximation: a dynamical systems viewpoint , Volume 100. Springer

2008

[5] [5]

Castro, L. d., A. F. Galvao, G. Montes-Rojas, and J. Olmo (2022). Portfolio selection in quantile decision models. Annals of finance\/ 18\/ (2), 133--181

2022

[6] [6]

Chambers, C. P. (2009). An axiomatization of quantiles on the domain of distribution functions. Mathematical Finance: An International Journal of Mathematics, Statistics and Financial Economics\/ 19\/ (2), 335--342

2009

[7] [7]

Rowland, M

Dabney, W., M. Rowland, M. Bellemare, and R. Munos (2018). Distributional reinforcement learning with quantile regression. In Proceedings of the AAAI conference on artificial intelligence , Volume 32

2018

[8] [8]

de Castro, L. and A. F. Galvao (2019). Dynamic quantile models of rational behavior. Econometrica\/ 87\/ (6), 1893--1939

2019