{"id":"b32df214-2e98-4826-92f3-9ef7d685359e","arxiv_id":"1908.03905","paper_version":2,"verdict":"REJECT","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"high","formal_verification":"none","parameter_count":9,"one_line_summary":"The paper derives an approximate optimal stock fraction from a stochastic maximum principle applied to quantile-based return processes with a VaR constraint, but the derivation swaps the median and lower-quantile roles and uses an unjustified diffusion structure.","lead":"An investor's stock allocation strategy is derived in continuous time for heavy-tailed returns, using quantiles of the return distribution to define both the objective and a Value at Risk constraint. The paper gives an approximate closed-form allocation rule and illustrates it on one utility company's daily prices, but the derivation mislabels the quantile roles and contains unsupported mathematical steps.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Equation (7) rests on a misspecified quantile SDE: Lambda is used directly as the diffusion matrix for independent Brownians, so increments have covariance Lambda*Lambda^T, not Lambda; the derived strategy is for a different return process.","rationale":"The reader's weakest assumption is exactly the one I would press hardest: the asymptotic covariance matrix of sample quantiles is used as a diffusion coefficient without a square-root factorization. This is not a matter of calibration or approximation; it is a mismatch between the distributional target and the stochastic process actually solved. The paper's own warning that the BSDE solutions are obtained by setting q=0 and Q=0 is a second concern, but the diffusion misspecification is prior: if the state process is wrong, no amount of adjoint manipulation can rescue Eq. (7). The empirical section does not provide independent support because it simulates the same misspecified SDE. My recommendation therefore agrees with the reader's REJECT; no verdict adjustment is needed.","tokens_in":17866,"tokens_out":5054,"duration_ms":47494,"concrete_test":"Analytically compute the covariance of the increments generated by the paper's SDE (with independent W1,W2): it is Lambda*Lambda^T. Evaluate this at the calibrated values f1=47.63579, f2=68.43975, p1=0.05, p2=0.5 from Section 3 and compare with Lambda. If the two matrices differ, the SDE in Section 2 does not reproduce the quoted asymptotic quantile distribution, so any strategy derived from it, including (7), is not a solution to the stated problem.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing step is the quantile SDE in Section 2. After quoting asymptotic joint normality with covariance Lambda, the paper writes [dX1; dX2] = [0; b2] dt + Lambda [dW1; dW2] with independent W1,W2. For a diffusion with coefficient matrix Lambda, the instantaneous covariance of increments is Lambda*Lambda^T, not Lambda. In particular, Var(dX1) = (p1(1-p1)/f1^2)^2 + (p1(1-p2)/(f1 f2))^2, which is not p1(1-p1)/f1^2, and Cov(dX1,dX2) picks up cross terms. To represent the stated asymptotic distribution one needs a matrix square root A with A*A^T = Lambda. All later objects, including the wealth equations (1)-(2), the Hamiltonian, the adjoint BSDEs, and formula (7), are derived from this misspecified diffusion. The label swap compounds the problem: L(1) is called median wealth but is driven by dX1, the p1=0.05 quantile return, while L(2), the VaR process, is driven by dX2, the median return. Even after fixing the diffusion matrix, the optimization would maximize the 5% quantile while constraining the median. The Ekeland approximation argument cited in the Remark cannot repair this because it concerns smoothing of the indicator constraint, not the state dynamics.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper considers a continuous-time portfolio problem with one risky asset and one risk-free asset when the risky return is heavy-tailed. Since moments may not exist, the authors propose to work with two return quantiles (the 5% and the 50% quantiles), use a multivariate normal approximation for their joint distribution, and model the quantile dynamics by an Itô SDE. They then formulate a constrained stochastic control problem that maximizes discounted power utility of a median-wealth process subject to a Value-at-Risk constraint, solve it by the stochastic maximum principle, and derive a closed-form approximate optimal strategy in equation (7). The model is calibrated nonparametrically with a kernel density estimator on daily Entergy Corporation returns, and the resulting strategy and wealth paths are simulated for various parameter choices. The paper claims that the results accord with financial intuition.","tokens_in":18254,"tokens_out":8227,"duration_ms":93599,"significance":"If the derivation were correct, the paper would offer a genuinely useful construction: a quantile-based, nonparametric dynamic allocation rule that bypasses nonexistent moments under heavy tails. The choice of quantiles and the use of asymptotic normality of sample quantiles are appropriate building blocks, and the empirical setting is relevant. However, the central analytical chain contains load-bearing errors: the quantile SDE is misspecified, the objective and constraint labels are reversed, the adjoint BSDEs are solved by setting the martingale integrands to zero without justification, and the closed-form formula (7) rests on an unvalidated exponential approximation. These are not local presentation issues; they affect the claimed optimal strategy and therefore the entire contribution. The paper provides no machine-checked proofs, no code, and no out-of-sample validation, so the empirical conclusions are also not independently verifiable.","major_comments":[{"comment":"The diffusion matrix in the SDE for (dX1, dX2) is set equal to the asymptotic covariance matrix Lambda, but with independent Brownian motions the instantaneous covariance of the increments is Lambda Lambda^T, not Lambda. For example, Var(dX1) would be (p1(1-p1)/f1^2)^2 + (p1(1-p2)/(f1 f2))^2, not p1(1-p1)/f1^2. A correct diffusion model requires a matrix square root A with A A^T = Lambda. Moreover, the displayed Lambda also omits the 1/N factor from the standard asymptotic covariance of sample quantiles. Since equations (1), (2), the Hamiltonian, the adjoint equations, and formula (7) are all derived from this misspecified SDE, the closed-form strategy is not the optimal strategy for the process the authors intend to model.","section":"Section 2, quantile SDE after the Proposition"},{"comment":"The paper says L(1) is the median wealth, but its dynamics are driven by dX1 = dX(p1) with p1 = 0.05, the lower 5% quantile; it then says L(2) is the lower-quantile (VaR) process, but its dynamics are driven by dX2 = dX(p2) with p2 = 0.5, the median. Consequently the objective in equation (2) maximizes a process driven by the 5% return quantile, while the constraint is applied to a process driven by the median. This reverses the roles of the objective and the risk constraint, so the optimization problem actually solved is not the stated one.","section":"Section 2, wealth processes L(1) and L(2)"},{"comment":"The adjoint BSDEs are not solved; the martingale integrands q(t) and Q(t) are simply set to zero. In the stochastic maximum principle, q and Q are part of the solution and are determined by the BSDE, and q appears in the Hamiltonian through a trace term involving sigma. Because sigma depends on the control pi, dropping q changes the first-order condition used to derive the strategy. No argument is given that q = 0 and Q = 0 are the actual integrands for this problem, so the resulting s(t), S(t), and ultimately equation (7) are not established as the SMP solution.","section":"Section 2, adjoint equations (3) and (4)"},{"comment":"The step from equation (6) to equation (7) replaces the ratio of exponential terms by 1 without any error bound or numerical justification. This approximation is the step that produces the claimed closed-form optimal strategy. Since the resulting formula is the central output of the paper, the approximation must be validated with explicit error estimates or at least a careful numerical check over the relevant parameter range; as written, equation (7) is at best a heuristic.","section":"Section 3, derivation of equation (7) from equation (6)"},{"comment":"The parameters b2, Q0.05, f0.05, f0.5, and the KDE bandwidth are calibrated on the same Entergy Corporation series that is then used to generate the optimal-strategy and wealth curves. The reported regularities, such as higher risk-free rates reducing stock investment, are therefore in-sample properties of the fitted formula rather than out-of-sample evidence. A holdout period or a proper cross-validation scheme is needed before the empirical claims can support the analytical conclusions.","section":"Section 3, calibration and empirical illustration"}],"minor_comments":[{"comment":"The figures are not actually included; the text and the figure list contain placeholders such as 'Figure 1 and Figure 2 should be placed here.' This makes it impossible to verify the quantitative claims about confidence intervals and wealth paths.","section":"Section 3 and List of Figures"},{"comment":"The text refers to negative exponential utility as 'Constant Absolute Relative Risk Aversion (CARA)' and says RRA = eta x; this appears to conflate CARA and CRRA. Negative exponential utility has constant absolute risk aversion, not constant relative risk aversion.","section":"Section 2, utility discussion"},{"comment":"The Remark invokes 'Ekeland's theorem' but cites reference [6], which is Clarke's book, and it does not state the metric or the sense in which the solution is approximate. As written, the remark cannot justify the combined effect of the indicator smoothing, the q = 0 and Q = 0 simplifications, and the exponential approximation.","section":"Section 2, Remark after equation (6)"},{"comment":"For r = 0.0004 the text reports portfolio wealth reaching an order of 10^15 starting from L1 = 1 over a 795-day horizon; such magnitudes suggest possible numerical instability in the discretization of equation (7) and should be diagnosed rather than reported as a financial outcome.","section":"Section 3, simulated wealth magnitudes"}],"recommendation":"reject","confidential_remarks":"The manuscript has a relevant topic and a sensible nonparametric quantile-based starting point, but the central derivation is not sound as written. The diffusion misspecification and the reversed quantile labels invalidate the claimed optimal strategy, and the additional unvalidated approximations prevent a local fix. A revised paper would need to redo the modelling from the SDE specification onward and provide out-of-sample empirical support; that is substantial new work rather than a minor revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe paper attacks a real problem: continuous-time portfolio choice when returns are heavy-tailed and moments don't exist, using quantiles as state variables and a VaR constraint. That is a sensible idea, and the paper correctly notes that sample quantiles are asymptotically joint normal with a known covariance. Formulating the optimization via the stochastic maximum principle with a smoothed indicator constraint is also a reasonable route. So the setup is not without merit.\n\nUnfortunately, the central derivation is not sound. The authors take the asymptotic covariance matrix Lambda of sample quantiles and write dX = mu dt + Lambda dW for independent Brownians. That gives increments with covariance Lambda Lambda^T, not Lambda. To match the stated asymptotic distribution you need a matrix square root A with A A^T = Lambda. This is not a minor typo; every later equation, including the Hamiltonian, the adjoint BSDEs, and the optimal strategy (7), is derived from the wrong dynamics. The stress-test note is correct.\n\nThere is also a label swap that compounds the problem. X1 is the 5% quantile and X2 is the median by the definitions, but L(1), called the median wealth, is driven by dX1, and L(2), called the VaR process, is driven by dX2. So the problem actually maximizes the 5% quantile subject to the median, not the intended median maximization with a VaR constraint. That is a load-bearing inconsistency, not a notation choice.\n\nThe adjoint BSDEs are \"solved\" by simply setting q=0 and Q=0. That is not a derivation; it is an assumption that the martingale integrands vanish, which has to be justified or at least shown to be consistent. As it stands, the proof does not go through.\n\nThe empirical section is in-sample only: parameters are calibrated on the same Entergy series used to illustrate the strategy, with no out-of-sample check, no baselines, no code, and the figures are placeholders. The qualitative claim that higher risk-free rates reduce stock investment is plausible but not validated by the analysis.\n\nWhat the paper does well: the idea of using quantile processes for heavy-tailed returns is worth pursuing, and the paper cites relevant literature, including the authors' own discrete-time work. The nonparametric calibration approach is sensible.\n\nBottom line: as it stands this is not a reliable result. The topic deserves serious work, but this version needs a major rethinking of the state dynamics. If the authors fix the diffusion matrix and the label swap, and then actually solve or convincingly approximate the adjoint equations, the paper could have value. I would not cite it in its current form, but I think a serious referee should see it—not for acceptance, but to give detailed feedback. I would send it to review rather than desk-reject, mainly because the errors are instructive and the topic is significant.","headline":"Good idea for continuous-time quantile-based portfolio choice, but the central SDE is misspecified and the median/VaR labels are swapped, so the derived optimal strategy solves a different problem.","tokens_in":18710,"tokens_out":3024,"would_cite":false,"duration_ms":30836,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["90C39","91G10","91G80"],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper claims to provide a closed-form approximate optimal portfolio strategy for heavy-tailed stock returns by maximizing median wealth subject to a Value-at-Risk constraint, and shows empirically that higher risk-free rates reduce…","keywords":["continuous-time portfolio optimization","Value at Risk","heavy-tailed returns","quantile asymptotics","stochastic maximum principle","Hamiltonian system","nonparametric density estimation","median return"],"falsifier":"Simulate a heavy-tailed return process with known parameters, form rolling 5% and 50% sample quantiles over time, and compare the empirical quadratic covariation of those quantile paths with the value implied by $\\Lambda$; a mismatch means the quantile-SDE assumption is false, so formula (7) would not be the optimum for the simulated process.","tokens_in":79,"feed_emoji":"📈","tokens_out":10980,"duration_ms":230644,"temperature":0.7,"pith_summary":"The paper claims that a continuous-time portfolio problem can be solved even when the risky asset's return distribution is so heavy-tailed that variance does not exist. The proposed solution replaces the return process by the pair of its 5% and 50% sample quantiles, which are asymptotically jointly normal, and maximizes the discounted expected power utility of median wealth subject to the lower quantile wealth staying above a floor with at least 95% probability. Using the stochastic maximum principle with a smooth sigmoid approximation of the Value-at-Risk constraint, the paper derives a closed-form approximate optimal strategy, formula (7), and calibrates the needed densities nonparametrically from real daily prices. This matters because, if the derivation holds, it supplies an implementable allocation rule for heavy-tailed assets without knowing the distribution's functional form.","feed_headline":"Closed-form portfolio rule found for heavy-tailed returns","feed_subtitle":"Maximize median wealth while holding the 5% quantile above a floor; higher risk-free rates cut stock allocations.","key_machinery":"The load-bearing object is the asymptotic covariance matrix $\\Lambda$ of two sample quantiles, with entries $\\Lambda_{ij}=p_i(1-p_j)/(f_if_j)$ for $i,j\\in\\{1,2\\}$, where $p_1=0.05$, $p_2=0.5$, and $f_1,f_2$ are the densities at those quantiles. The paper uses $\\Lambda$ as the diffusion coefficient in the SDE governing the quantile processes, turning a statistical fact about sample estimates into a stochastic control model. The second piece of machinery is the sigmoid approximation of the indicator function in the VaR constraint, which makes the objective smooth; the stochastic maximum principle then supplies adjoint equations whose reduced solution leads to the first-order condition and, after a further approximation, to formula (7).","core_discovery":"On its own terms, the paper's result is a closed-form approximate solution to a constrained stochastic control problem, whose state is the pair of wealth processes associated with the two quantiles: the median process $\\bar L_1$ carries the utility, and the 5% quantile process $\\bar L_2$ carries the Value-at-Risk constraint. After forming the Hamiltonian, solving the two adjoint equations with zero matrix-valued terms, and setting the derivative with respect to the portfolio weight to zero, the paper obtains the optimal portfolio weight $$\\bar\\pi_t=\\frac{$e^{{-\\beta t}}$\\psi_0 \\bar L_1(t)^\\gamma $r^{2}$-\\psi k_1\\bar L_2(t)(b_2-r)(r-\\$\\beta$)}{\\left($e^{{-\\beta t}}$\\psi_0 \\bar L_1(t)^\\gamma-\\psi k_1\\bar L_2(t)\\right)(b_2-r)r}.$$ The paper presents this as the optimal strategy for the approximated problem and, via Ekeland's theorem, as an approximate optimum of the original non-smooth problem; it then supports the formula with simulations showing that higher risk-free rates lower the optimal stock fraction and that changes in $\\gamma$ and $\\beta$ leave the strategy largely unchanged.","pith_inferences":["A testable extension is to replace the 5% quantile process by an expected-shortfall or conditional-VaR functional in the same Hamiltonian machinery, since the calibration only requires quantile asymptotics and a smooth approximation of the constraint.","A cautious reading treats formula (7) as a calibrated heuristic rather than an exact optimum for the original heavy-tailed process, because the derivation replaces the return process with a quantile diffusion and then smooths the constraint; simulated fat-tailed data would show how large the gap is.","The empirical insensitivity to $\\gamma$ and $\\beta$ suggests the allocation is governed mainly by the ratio of median to tail quantile levels and by the risk-free rate; if confirmed, the rule could be condensed into a simple tail-thickness-dependent trading formula."],"forward_implications":["If formula (7) is the optimal strategy, a higher risk-free rate $r$ reduces the fraction held in the risky asset, matching the paper's empirical finding.","The strategy is implementable without specifying the return distribution, because calibration needs only sample quantiles and kernel-density estimates of the quantile densities.","Putting more weight on the utility objective and less on the VaR constraint increases both the stock allocation and the final wealth, consistent with accepting more tail risk for higher median gains.","The framework extends to logarithmic and exponential utility, and other constant-risk-aversion forms, at the cost of algebraic complexity, so the median/VaR formulation is not tied to power utility.","Because the constraint is smoothed, the resulting strategy is an approximate optimum rather than an exact one; Ekeland's theorem is invoked to justify the approximation."],"supporting_citations":[{"why":"Supplies the asymptotic joint normality and the covariance matrix of sample quantiles that becomes the paper's quantile SDE.","marker":"[4]"},{"why":"Provides the stochastic maximum principle theorem and adjoint-equation framework used for the optimality derivation.","marker":"[23]"},{"why":"Invoked through Ekeland's theorem to argue that the simplified solution is an approximate solution of the original non-smooth problem.","marker":"[6]"},{"why":"Provides the Entergy daily closing-price series from which returns and quantile parameters are calibrated.","marker":"[16]"},{"why":"Supplies the kernel density estimation method used to obtain the quantile densities nonparametrically.","marker":"[14]"},{"why":"The discrete-time version of this median/VaR heavy-tail problem that the paper generalizes to continuous time.","marker":"[5]"},{"why":"Motivates converting return dynamics into quantile dynamics, a step the paper adopts for the state process.","marker":"[1]"},{"why":"Offers the nonparametric calibration idea that the paper's estimation strategy builds on.","marker":"[2]"}],"fun_headline_variants":["Closed-form VaR-constrained portfolio rule under heavy tails","Explicit approximate solution for VaR-constrained heavy-tail portfolios","Heavy-tail VaR control: explicit optimal weight formula","Nonparametric VaR constraint yields closed-form approximate allocation","Approximate closed-form rule for heavy-tail VaR portfolios"],"cache_read_input_tokens":20736,"weakest_assumption_plain":"The load-bearing assumption is that the asymptotic covariance matrix $\\Lambda$ of quantile estimates, a statement about how estimates vary across samples, can be used directly as the instantaneous random-shock matrix of the quantile processes' own time-series SDE; if the quantile paths do not follow that SDE, formula (7) has not been derived for the actual heavy-tailed return process.","fun_headline_variants_meta":{"raw":{"variants":["Closed-form VaR-constrained portfolio rule under heavy tails","Explicit approximate solution for VaR-constrained heavy-tail portfolios","Heavy-tail VaR control: explicit optimal weight formula","Nonparametric VaR constraint yields closed-form approximate allocation","Approximate closed-form rule for heavy-tail VaR portfolios"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000761,"raw_usage":{"total_tokens":3368,"prompt_tokens":927,"completion_tokens":2441,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":543,"completion_tokens_details":{"reasoning_tokens":2360}},"tokens_in":543,"tokens_out":2441,"duration_ms":20835,"temperature":1.0,"reasoning_tokens":2360,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T13:58:46.911292+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Simulate a heavy-tailed return process with known parameters, form rolling 5% and 50% sample quantiles over time, and compare the empirical quadratic covariation of those quantile paths with the value implied by $\\Lambda$; a mismatch means the quantile-SDE assumption is false, so formula (7) would not be the optimum for the simulated process.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the asymptotic joint normality and the covariance matrix of sample quantiles that becomes the paper's quantile SDE."},{"cited_title":"Yong & X","cited_arxiv_id":null,"evidence_quote":"Provides the stochastic maximum principle theorem and adjoint-equation framework used for the optimality derivation."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Invoked through Ekeland's theorem to argue that the simplified solution is an approximate solution of the original non-smooth problem."},{"cited_title":"Markovich, Non parametric analysis of univariate heavy-tailed distribution, Wiley (2007)","cited_arxiv_id":null,"evidence_quote":"Supplies the kernel density estimation method used to obtain the quantile densities nonparametrically."},{"cited_title":"Biswas & D","cited_arxiv_id":null,"evidence_quote":"The discrete-time version of this median/VaR heavy-tail problem that the paper generalizes to continuous time."},{"cited_title":"Agarwal & R","cited_arxiv_id":null,"evidence_quote":"Motivates converting return dynamics into quantile dynamics, a step the paper adopts for the state process."},{"cited_title":"Aït-Sahalia & A","cited_arxiv_id":null,"evidence_quote":"Offers the nonparametric calibration idea that the paper's estimation strategy builds on."}],"review_version":1}