{"id":"a97ad9b9-682c-452b-aece-a0e623e95285","arxiv_id":"2412.11575","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"Introduces CAPE, a SCAD-penalized mean-variance estimator that accounts for proportional and quadratic transaction costs, with oracle properties under sparsity and high-dimensional consistency.","lead":"High-dimensional portfolio optimization typically ignores trading frictions. This paper integrates transaction costs directly into the mean-variance objective with a sparsity penalty, and proves that the proposed estimator converges to the optimal cost-aware portfolio in large-asset universes.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The oracle target in Eqs. (2.4)-(2.5) is sparse by definition; the paper never bounds the Sharpe-ratio gap between this constrained optimum and the unconstrained cost-aware optimum, so the practical relevance of the consistency theorem is unestablished.","rationale":"The reader's weakest_assumption is the one I would defend as the most load-bearing. The paper's central theorems are conditional on the true optimal cost-aware portfolio being exactly sparse, with support size s_t = O(√(n/log p)) (Section 2.2, after Eq. (2.5)). If that condition fails, the oracle properties in Theorems 1–4 do not apply, and the LLA estimator converges to a sparse proxy rather than to the investor's true cost-aware optimum. This is not an internal inconsistency: the proofs could be correct for the constrained target. The concern is about external validity, because the sparsity constraint is baked into the definition of the target in Eqs. (2.4)-(2.5), while the abstract and empirical sections implicitly claim relevance to the unconstrained cost-aware problem. The proposed simulation with a dense oracle versus a sparse oracle would directly settle whether the gap is material. A secondary, smaller issue is that the closed-form expression for δ*_A in Section 2.3.2 appears to omit the term 2·1^T\\tildeΣ^{-1}Σ w+ in the Lagrange multiplier and can violate δ^T1 = 0 for β ≠ 0; this should be corrected, but it is not the primary threat to the central claim because the oracle is defined by the optimization problem itself, not by the closed-form expression.","tokens_in":76,"tokens_out":17218,"duration_ms":159285,"concrete_test":"In the Section 3 simulation design (true Σ, µ, factor model, p = 2000, n = 200), compute the unconstrained cost-aware oracle w*_unc by solving (2.11) without the ∥w∥_0 ≤ s0 constraint, and the sparse oracle w*_1 from (2.4) with s0 set to the support size actually selected by CAPE-S. Evaluate both at the true out-of-sample mean and covariance with transaction costs (as in Corollary 1, using Σ_2 and µ_2) and compare their true Sharpe ratios. If the dense oracle's Sharpe ratio exceeds the sparse oracle's by more than the estimation error O(√(log s0/n)) claimed in Theorem 1, then the sparse target is materially suboptimal and the practical relevance of the consistency result is not established.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing assumption is that the true cost-aware optimum is exactly s0-sparse, as imposed in Eqs. (2.4)-(2.5) with s_t = O(√(n/log p)). This is an assumption, not a consequence of the cost structure. For quadratic costs C(w) = (β⊙w)^T w in (2.11), the unconstrained minimizer of w^TΣw − γw^Tµ + (β⊙w)^T w subject only to w^T1 = 1 is generically dense; the ℓ0 constraint in (2.4) changes the target. For proportional costs, the ℓ1 term induces sparsity only for sufficiently large α, which the simulations do not guarantee (α = 0.001 in Section 3). Theorems 1–4 and Corollaries 1–2 establish oracle equivalence and Sharpe-ratio consistency with respect to the constrained sparse oracle, not the unconstrained cost-aware optimum. The approximation error between these two portfolios is never quantified. If the unconstrained optimum has many small positions, CAPE-S can be consistent for the wrong object, and the abstract's phrase 'the optimal cost-aware portfolio' is therefore potentially misleading.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a high-dimensional, multi-period mean-variance portfolio estimator that incorporates transaction costs directly into the optimization. It defines a sparsity-constrained \"cost-aware\" optimal portfolio for the construction stage (Eq. 2.4) and for each rebalancing stage (Eq. 2.5), then approximates the ℓ0 constraint with a SCAD penalty and solves the problem via local linear approximation (LLA). The central theoretical claims are oracle properties: after two LLA iterations, the CAPE-S estimator converges to the oracle estimator at an ℓ∞ rate, and in-sample and out-of-sample Sharpe ratio estimators converge to those of the optimal cost-aware portfolio. The paper also reports Monte Carlo simulations and empirical applications to S&P 500 and Russell 2000 constituents with asset-specific transaction cost estimates.","tokens_in":24055,"tokens_out":5739,"duration_ms":56010,"significance":"If the theorems hold, the paper would be a useful contribution to a relatively thin literature on transaction-cost-aware portfolio optimization in high dimensions, and the LLA-based two-step oracle property is an attractive algorithmic result. The strengths include the explicit treatment of both quadratic and proportional transaction costs, the use of real, asset-specific cost estimates in the empirical section, and a broad simulation study. However, the central claims rest on a sparsity assumption that is imposed rather than derived, the out-of-sample Sharpe consistency statements appear to require information that is not available at the decision date, and the proofs are entirely deferred to a supplementary file that was not available for evaluation. These issues currently prevent the main results from being verified and limit the practical interpretation of the theorems.","major_comments":[{"comment":"The target portfolio is defined as the minimizer of the cost-aware objective subject to a cardinality constraint ∥w∥0 ≤ s0 and ∥δ∥0 ≤ s0. This sparsity is imposed, not derived from the transaction cost structure. For quadratic costs C1(w) = (β⊙w)^T w, the unconstrained minimizer of w^TΣw − γw^Tµ + (β⊙w)^T w subject only to w^T1 = 1 is generically dense, so the ℓ0 constraint changes the economic target. The paper never bounds the approximation error between the constrained optimum in Eqs. (2.4)-(2.5) and the unconstrained cost-aware optimum. Without such a bound, Theorems 1-4 establish consistency to a constrained oracle, not to the 'optimal cost-aware portfolio' of the abstract, and the practical relevance of the consistency theorem is unquantified.","section":"Section 2.2, Eqs. (2.4)-(2.5)"},{"comment":"The out-of-sample Sharpe ratio consistency statements use bµ2 and bΣ2 at the first construction stage and bµt+1 and bΣt+1 at stage t. At the first decision day n+1, however, only the returns R1,...,Rn are available, and no estimator bµ2 or bΣ2 is defined. If these are meant to be estimators formed from period-2 data, the corollaries use future information and are not valid out-of-sample claims. The authors should either define the out-of-sample Sharpe ratio using realized returns realized after the decision date and prove consistency of the resulting estimator, or state clearly that bµ2 and bΣ2 are independent test-period quantities and justify why they are available at the decision date.","section":"Corollary 1 and Corollary 2"},{"comment":"All proofs, including the derivations of the explicit oracle solutions, the oracle properties, and the Sharpe ratio consistency results, are deferred to the Supplementary Material. The main text contains only theorem statements and informal remarks. Because the proofs were not available for review, I could not verify the central claims of the paper. At minimum, the main text should include proof sketches that convey the key concentration arguments and the role of each assumption, so that a referee can check the logic without reconstructing the full supplement.","section":"Theorems 1-4 and Lemmas 1-4"},{"comment":"The condition ∥w+t−1∥1 < C0 is stated as a deterministic assumption, but w+t−1 is a random quantity: it is obtained by applying the return maps f1,...,fn to the previous portfolio weight and therefore depends on the realized returns and on the previous estimator. The probability statements in Theorems 2 and 4 do not account for the event that this condition holds. The authors should either make the condition a high-probability event and include it in the probability bound, or state the theorem conditionally on the event and give a separate bound for its probability.","section":"Theorem 2(i) and Theorem 4(i)"}],"minor_comments":[{"comment":"In the statement of Theorem 1, 'SACD penalty' is a typo and should read 'SCAD penalty'.","section":"Section 2.3, Theorem 1"},{"comment":"Lemma 2 mixes the notation Op with an explicit high-probability statement: it writes 'with probability at least 1−c1p−c2' but then reports Op rates. The high-probability event should be stated explicitly in the same form as Lemma 1.","section":"Lemma 2"},{"comment":"The tuning parameter λopt is selected by the highest in-sample Sharpe ratio, but the theorems require λ ≥ M sqrt(s1 log p/n). The manuscript does not connect the data-dependent selection rule to this theoretical condition, nor does it discuss how the selected λ behaves as n and p grow.","section":"Section 3, tuning parameter"},{"comment":"The conclusion says 'we prove the sign consistency', but sign consistency is not explicitly stated in any theorem; the theorems establish ℓ∞ convergence and oracle equivalence. The conclusion should either add a formal sign-consistency result or rephrase the claim.","section":"Section 5, Conclusion"},{"comment":"The statement uses k ∈ [s1,p] without defining k. The reader is left to infer that k is the active-set size or the dimension of the subvector; this should be defined explicitly in the statement.","section":"Theorem 1"},{"comment":"The text 'Shape Ratio estimation errors' contains a typo and should read 'Sharpe ratio'.","section":"Section 2.3.2, after Corollary 2"}],"recommendation":"major_revision","confidential_remarks":"The manuscript would benefit from making the supplementary proofs available to reviewers and from qualifying the abstract's claim about 'the optimal cost-aware portfolio' in view of the sparsity assumption. The out-of-sample Sharpe ratio issue in Corollaries 1 and 2 is the most serious technical concern, as it appears to involve future information; if it cannot be resolved, the corresponding claims should be removed or reformulated."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The genuinely new part is the package: a multi-period, high-dimensional mean-variance framework that puts transaction costs inside the optimization, handles both proportional and quadratic costs, and proves oracle properties for a SCAD/LLA estimator. That combination is not in Hautsch-Voigt or Ledoit-Wolf, and the rebalancing setup with sparsity is a real extension. The paper is also well executed on the empirical side: asset-specific cost estimates from bid-ask spread data, Ledoit-Wolf covariance shrinkage that satisfies their assumptions, and a sensible comparison set. If the supplementary proofs hold up, the main theorems are plausible and the rates are standard.\n\nThe soft spot is exactly what the stress-test note says. The target portfolio in (2.4)-(2.5) is defined with an explicit ℓ0 constraint, and the sparsity of that target is assumed, not derived from the cost structure. With quadratic costs, the unconstrained optimum is generically dense. Theorems 1-4 give consistency for the sparse oracle estimator, but the abstract says convergence to 'the optimal cost-aware portfolio' without qualification. The approximation gap between the sparse constrained optimum and the unconstrained cost-aware optimum is never bounded. This matters: if the true unconstrained optimum has many small positions, CAPE-S can be consistently estimating the wrong object. That said, for investors who genuinely want a limited number of holdings, the sparse target is the right object, and the paper should just say so explicitly instead of letting the abstract overclaim.\n\nOther concerns are more minor. Proofs are all in the supplement, so I cannot verify them from the main text. The beta-min condition and boundedness of previous weights are strong but are the usual price for oracle results. The real-data evaluation has no error bars or robustness checks across cost specifications. And they cite Ledoit-Wolf (2025) but do not compare against its empirical transaction-cost-aware portfolios, which would be a natural benchmark.\n\nThis deserves serious peer review. The method is useful, the theoretical claims are clearly stated (once you read past the abstract), and the empirical work is careful. I would send it to a referee, with the request that they check the supplement and push the authors to clarify the target of their consistency results.","headline":"A solid high-dimensional portfolio paper with honest scope issues: the oracle theory targets a sparsity-constrained optimum, not the unconstrained cost-aware optimum the abstract advertises.","tokens_in":24501,"tokens_out":1460,"would_cite":true,"duration_ms":14752,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62P05","91G10","62J07"],"pacs":[],"model":"deepseek-v4-flash","headline":"A two-step LLA portfolio estimator with SCAD penalty matches the oracle cost-aware portfolio in high dimensions.","keywords":["high-dimensional portfolio optimization","mean-variance model","transaction costs","SCAD penalty","local linear approximation","oracle property","portfolio rebalancing","Sharpe ratio"],"falsifier":"Construct a simulation where the cost-aware mean-variance optimum (with the same transaction costs but no sparsity constraint) has all $p$ weights nonzero, run CAPE-S, and check whether the $\\ell^\\infty$ error still shrinks at the claimed $\\sqrt{\\log k/n}$ rate and whether the Sharpe gap to the dense optimum vanishes; if either fails, the sparsity assumption is violated and the theorem's conditions do not hold. A complementary check on the paper's own Russell 2000 sample is to compute the unconstrained cost-aware solution and count how many weights exceed a small threshold, comparing that count to the assumed $s_0$.","tokens_in":23510,"feed_emoji":"📈","tokens_out":10051,"duration_ms":84506,"temperature":0.7,"pith_summary":"The paper tries to establish that an investor can do optimal mean-variance portfolio construction and rebalancing even when the number of assets, $p$, is much larger than the number of return observations, $n$, as long as transaction costs are put inside the objective and the true optimal portfolio is sparse. It defines a cost-aware portfolio by adding proportional or quadratic trading costs to the usual mean-variance criterion together with an $\\ell_0$ budget on the number of assets held, then relaxes that budget with a SCAD penalty and solves the nonconvex program with two iterations of local linear approximation. The central claim is that after two iterations the estimated portfolio equals the oracle estimator that already knows the correct asset support, with probability at least $1 - c_1 k^{-c_2}$, and differs from the true optimal portfolio by $\\ell^\\infty$ error $O(\\sqrt{\\log k/n})$. If true, this means fee-aware sparse rebalancing has a rigorous statistical justification in high dimensions, and the in-sample and out-of-sample Sharpe ratios of the estimated portfolio converge to those of the true cost-aware optimum. The paper supports the claim with simulations and with S&P 500 and Russell 2000 applications where the proposed estimator posts the highest net-of-cost Sharpe ratios among the compared strategies.","feed_headline":"Cost-aware portfolios match oracle accuracy in high dimensions","feed_subtitle":"A two-iteration estimator finds the sparse, fee-aware portfolio even when assets outnumber return data.","key_machinery":"The load-bearing object is the CAPE-S estimator: the minimizer of $w^\\top \\hat{\\Sigma}_t w - \\gamma w^\\top \\hat{\\mu}_t + C_t(w) + \\sum_j P_\\lambda(|w_j|)$ under the budget constraint $w^\\top 1 = 1$, and its rebalancing analogue for the weight-difference vector $\\delta_t$, where $C_t$ is either a quadratic cost $(\\beta \\odot w)^\\top w$ or a proportional cost $\\|\\alpha \\odot w\\|_1$ and $P_\\lambda$ is the SCAD penalty, a concave sparsity-inducing penalty that avoids the bias of $\\ell_1$ regularization. The key mechanism is that two iterations of the local linear approximation (LLA) algorithm, initialized at the Lasso-type CAPE-L estimate, produce a solution identical to the oracle estimator restricted to the true support: the $\\beta$-min condition keeps active coefficients above the SCAD threshold while the Lasso initializer already identifies the support with high probability. Quadratic transaction costs enter as a diagonal ridge $\\hat{\\Sigma}_t + \\mathrm{diag}(\\beta)$, which regularizes the covariance matrix; proportional costs add an $\\ell_1$ turnover penalty that directly induces sparsity in rebalancing.","core_discovery":"On the paper's own terms, the discovery is an oracle property for cost-aware portfolios: under Assumptions (A1)-(A3) and a $\\beta$-min condition on the active coefficients, the two-iteration LLA solution of the SCAD-penalized program converges to the oracle estimator with probability at least $1 - c_1 k^{-c_2}$ and achieves $\\| \\hat{w}_1^{\\beta} - w_1^* \\|_\\infty = O(\\sqrt{\\log k/n})$ at the construction stage; Theorems 2-4 give the analogous rates for rebalancing increments $\\delta_t^*$ under quadratic and proportional costs, with the rate improving to $\\sqrt{\\log k/n}$ when previous holdings are sparse. The paper further claims that the estimated portfolio's in-sample and out-of-sample Sharpe ratios converge in probability to those of the optimal cost-aware portfolio under $s_t\\sqrt{\\log s_t/n} = o(1)$, and demonstrates on S&P 500 and Russell 2000 data that the estimator achieves higher overall out-of-sample Sharpe ratios than the equally-weighted, mean-variance, penalized mean-variance, and cost-aware mean-variance benchmarks.","pith_inferences":["The two-iteration result suggests a practical warm-start rule for multi-period rebalancing: initialize each period's LLA from the previous period's selected support, which the paper's sparsity condition on $w^+_{t-1}$ partially anticipates; a head-to-head test of warm-started versus cold-started CAPE-S would quantify the gain.","Because quadratic transaction costs enter as a diagonal ridge on the covariance matrix, raising $\\beta$ should mimic stronger covariance shrinkage; one testable prediction is that turnover falls faster than risk as $\\beta$ increases, independent of $\\hat{\\Sigma}$.","The consistency of the Sharpe ratios points toward a data-driven rebalancing calendar: choose the interval between rebalances by balancing estimated cost parameters against the $\\sqrt{\\log k/n}$ estimation error instead of a fixed annual grid.","Asset-specific cost coefficients in the empirical design imply a liquidity tilt: CAPE-S should hold a selected support whose average bid-ask spread is below the universe average, which is directly checkable from the reported cost estimates."],"forward_implications":["After two LLA iterations, the SCAD-penalized CAPE-S estimator coincides with the oracle that knows the true support, so global optimization of the nonconvex problem is unnecessary for the stated rates.","Under $s_t\\sqrt{\\log s_t/n} = o(1)$, the estimated portfolio's in-sample and out-of-sample Sharpe ratios converge in probability to those of the optimal cost-aware portfolio.","Quadratic transaction costs act as a diagonal perturbation of the covariance matrix, while proportional costs add an $\\ell_1$ turnover penalty; both can be interpreted as regularizers within the same framework.","If previous-period holdings are sparse, rebalancing increments achieve the faster rate $\\sqrt{\\log k/n}$; with non-sparse holdings the rate degrades to $\\sqrt{\\log p/n}$ but consistency is retained.","On S&P 500 and Russell 2000 data from 2017 to 2020, the estimator reports the highest overall out-of-sample Sharpe ratio under both quadratic and proportional cost settings, and in the Russell 2000 it is the only strategy with a positive return in the 2018 stage."],"supporting_citations":[{"why":"Defines the mean-variance optimization objective that the paper extends to a cost-aware, high-dimensional setting.","marker":"Markowitz (1952)"},{"why":"Supplies the gross-exposure $\\ell_1$-penalized mean-variance benchmark and motivates penalized portfolio construction.","marker":"Fan et al. (2012)"},{"why":"Introduces the cost-aware mean-variance model with transaction costs and provides evidence that ex-ante costs improve stability.","marker":"Hautsch and Voigt (2019)"},{"why":"Provides the local linear approximation (LLA) algorithm and the one-step oracle property used to solve the SCAD-penalized programs.","marker":"Zou and Li (2008)"},{"why":"Extends the strong oracle property to high dimensions with Lasso initialization, the template for the two-iteration LLA theorems.","marker":"Fan et al. (2014)"},{"why":"Defines the SCAD penalty and its derivative conditions that the assumptions and proofs rely on.","marker":"Fan and Li (2001)"},{"why":"Documents the irrepresentable condition and bias of $\\ell_1$ regularization, motivating the nonconvex SCAD choice.","marker":"Zhao and Yu (2006)"},{"why":"The linear shrinkage covariance estimator used in simulations and empirics; the paper shows it satisfies assumptions (A2)-(A3).","marker":"Ledoit and Wolf (2004)"},{"why":"The nonlinear shrinkage covariance estimator used as an alternative and shown to satisfy the restricted strong convexity condition.","marker":"Ledoit and Wolf (2020)"},{"why":"Establishes restricted strong convexity of sample covariance matrices under sub-Gaussianity, supporting assumption (A3).","marker":"Loh and Wainwright (2015)"}],"fun_headline_variants":["Cost-aware portfolios hit oracle rates in high dimensions","Fee-aware portfolio optimization achieves oracle convergence","Transaction-cost portfolios beat benchmarks in high dimensions","Cost-aware rebalancing: oracle accuracy with sparse assets"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the true cost-aware portfolio and each rebalancing movement are sparse, with support size at most $s_0$ and $s_t = O(\\sqrt{n/\\log p})$; if the unconstrained fee-aware mean-variance optimum uses many assets, the oracle and consistency claims do not apply.","fun_headline_variants_meta":{"raw":{"variants":["Cost-aware portfolios hit oracle rates in high dimensions","Fee-aware portfolio optimization achieves oracle convergence","Transaction-cost portfolios beat benchmarks in high dimensions","Cost-aware rebalancing: oracle accuracy with sparse assets"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00024,"raw_usage":{"total_tokens":1493,"prompt_tokens":898,"completion_tokens":595,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":514,"completion_tokens_details":{"reasoning_tokens":537}},"tokens_in":514,"tokens_out":595,"duration_ms":5161,"temperature":1.0,"reasoning_tokens":537,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T14:47:13.212233+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Construct a simulation where the cost-aware mean-variance optimum (with the same transaction costs but no sparsity constraint) has all $p$ weights nonzero, run CAPE-S, and check whether the $\\ell^\\infty$ error still shrinks at the claimed $\\sqrt{\\log k/n}$ rate and whether the Sharpe gap to the dense optimum vanishes; if either fails, the sparsity assumption is violated and the theorem's conditions do not hold. A complementary check on the paper's own Russell 2000 sample is to compute the unconstrained cost-aware solution and count how many weights exceed a small threshold, comparing that count to the assumed $s_0$.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines the mean-variance optimization objective that the paper extends to a cost-aware, high-dimensional setting."},{"cited_title":"Zhang, and K","cited_arxiv_id":null,"evidence_quote":"Supplies the gross-exposure $\\ell_1$-penalized mean-variance benchmark and motivates penalized portfolio construction."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Introduces the cost-aware mean-variance model with transaction costs and provides evidence that ex-ante costs improve stability."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the local linear approximation (LLA) algorithm and the one-step oracle property used to solve the SCAD-penalized programs."},{"cited_title":"Xue, and H","cited_arxiv_id":null,"evidence_quote":"Extends the strong oracle property to high dimensions with Lasso initialization, the template for the two-iteration LLA theorems."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines the SCAD penalty and its derivative conditions that the assumptions and proofs rely on."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Documents the irrepresentable condition and bias of $\\ell_1$ regularization, motivating the nonconvex SCAD choice."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"The linear shrinkage covariance estimator used in simulations and empirics; the paper shows it satisfies assumptions (A2)-(A3)."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"The nonlinear shrinkage covariance estimator used as an alternative and shown to satisfy the restricted strong convexity condition."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Establishes restricted strong convexity of sample covariance matrices under sub-Gaussianity, supporting assumption (A3)."}],"review_version":1}