{"id":"4626a511-1531-4f94-8e2e-ab7909cc5795","arxiv_id":"2507.01545","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"ERSE rotates paired eigenvectors of the sample correlation matrix toward the uniform vector and re-estimates eigenvalues, cutting out-of-sample GMV portfolio variance for positively correlated assets by about 10-12% relative to linear and nonlinear shrinkage benchmarks.","lead":"A new covariance matrix estimator, ERSE, gently rotates the eigenvectors of the sample correlation matrix so that the virtual portfolios they define are less extreme, targeting assets whose returns are all positively correlated. In tests on U.S. factor-sorted stock portfolios, it lowers out-of-sample portfolio risk by about 10-12% compared with standard shrinkage estimators.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Algorithm 2's reconstruction step is ambiguous: retaining rotated eigenvectors silently drops nonzero off-diagonal entries of Q̂ᵀ R_S Q̂, while retaining sample eigenvectors contradicts the rotation narrative; the headline risk reductions hinge on which is implemented.","rationale":"The reader's weakest assumption partially coincides with ours: both flag the off-diagonal truncation. I would sharpen it to an ambiguity: the printed Algorithm 2 is internally inconsistent, because Step 8 refers to rotated eigenvectors while Step 9 writes Q_S, so the estimator is not uniquely defined without reading intent into the notation. This is load-bearing because the empirical claim cannot be reproduced or interpreted until the reconstruction is fixed. A secondary concern is the in-sample selection of δ = 0.25: Section 6.3 evaluates the δ-sweep and recommends the interval [0.15, 0.35] on the same out-of-sample period used for Table 2, so the reported advantage may be partially selected. I do not treat this as the primary issue because the δ-sweep shows a plateau rather than a knife-edge, and because the reconstruction ambiguity affects every δ value. I also note that Proposition 3's bound (15) appears to require a denominator argument the paper does not supply, but that proposition supports motivation rather than the headline empirical claim. The appropriate verdict remains conditional: release code, disambiguate the reconstruction, and re-report results under both implementations.","tokens_in":26076,"tokens_out":14699,"duration_ms":172075,"concrete_test":"Run the Table 2 out-of-sample GMV experiment under two implementations: (A) Σ̂ = D_S Q̂ Diag(diag(Q̂ᵀ R_S Q̂)) Q̂ᵀ D_S (rotated eigenvectors retained, off-diagonal dropped); (B) Σ̂ = D_S Q_S Diag(diag(Q̂ᵀ R_S Q̂)) Q_Sᵀ D_S (sample eigenvectors retained). If the average risk reductions versus LINC and QIS differ by more than two percentage points across the ten datasets, or if ERSE no longer achieves the lowest variance in a majority of datasets under one reading, the headline claim is not robust to the reconstruction ambiguity. Also report the maximum magnitude of the discarded off-diagonal entries cosθ sinθ(λ_j − λ_i) at the final iteration of each dataset.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central empirical claim in Section 6.2 depends on a reconstruction step that is not uniquely defined. Algorithm 2 first rotates sample eigenvectors via PER, then Step 8 sets λ̂_i = q̂_iᵀ R_S q̂_i using the rotated eigenvectors, and Step 9 writes the estimator as D_S Q_S Diag(λ̂) Q_Sᵀ D_S. If Q_S in Step 9 denotes the original sample eigenvectors, the procedure is only an eigenvalue shrinkage rule and the paper's 'eigenvector rotation' narrative is a misdescription. If Q_S is meant to be the rotated matrix Q̂, then the estimator keeps only the diagonal of Q̂ᵀ R_S Q̂ and discards the off-diagonal entries; for a paired rotation those entries equal cosθ sinθ(λ_j − λ_i), which are generally nonzero. The paper provides no bound or analysis of this truncation, and Proposition 5 only characterizes the diagonal entries, not the full estimator. Because the reported 10.52% and 12.46% average risk reductions are the paper's strongest claim, the results must be reproduced under both readings before the claim can be considered established.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes the Eigenvector Rotation Shrinkage Estimator (ERSE), a rotation-equivariant covariance matrix estimator designed for positively correlated assets. The method begins from the sample correlation matrix, imposes a minimum-deviation constraint on each eigenvector from the null space of the uniform vector, and enforces this constraint through paired eigenvector rotations (PER). The paper proves several properties of the deviation measure and of single PER steps, and then evaluates ERSE in a global minimum-variance (GMV) portfolio framework on ten factor-sorted Ken-French datasets. The headline empirical claim is that ERSE achieves the lowest out-of-sample variance across all datasets, with average risk reductions of 10.52% relative to the best linear shrinkage method (LINC) and 12.46% relative to quadratic-inverse shrinkage (QIS), together with lower condition numbers and more stable portfolio weights.","tokens_in":26356,"tokens_out":10901,"duration_ms":125420,"significance":"If the central empirical claim survives scrutiny, ERSE is a useful addition to the covariance-estimation toolbox for high-correlation settings. The paper deserves credit for explicitly targeting the positive-correlation case, for the Perron-Frobenius-based analysis of eigenvector misalignment (Propositions 1-3), and for a broad robustness campaign covering subperiods, estimation windows, and 150 random subsamples. The algebraic proofs of Propositions 1-6 are self-contained. The significance of the result, however, rests on two load-bearing points that are not yet settled: the exact estimator implemented in Algorithm 2, and the use of the same datasets to select the hyperparameter δ that drives the headline performance numbers. The theoretical connection to sequential linear shrinkage also needs strengthening for iterations after the first rotation.","major_comments":[{"comment":"The reconstruction in Algorithm 2 is not uniquely defined. Step 8 computes λ̂_i = q̂_i^T R_S q̂_i using the rotated eigenvectors, while Step 9 sets Σ̂ = D_S Q_S Diag(λ̂) Q_S^T D_S with Q_S denoting the original sample eigenvector matrix. If Q_S is the original matrix, the final estimator contains only the original eigenvectors, and the eigenvector rotation enters only through the scalar λ̂_i; the procedure is then an eigenvalue-shrinkage rule, and the paper's 'eigenvector rotation' description is misleading. If Q_S is meant to be the rotated matrix Q̂, then the estimator keeps only the diagonal of Q̂^T R_S Q̂ and discards the off-diagonal entries; for a single paired rotation those entries equal cosθ sinθ(λ_j − λ_i), which are generally nonzero, and Proposition 5 provides no analysis of this truncation. Because the out-of-sample results in Table 2 depend on which implementation is used, the authors must state explicitly which matrix is used in Step 9, make the algorithm consistent with that statement, and ideally report results under both readings if the ambiguity persists.","section":"Section 4, Algorithm 2, Steps 8-9"},{"comment":"The key hyperparameter δ is selected on the same data used for the headline evaluation. Section 6.1 states that δ = 0.25 is 'empirically determined', and Section 6.3 recommends the interval [0.15, 0.35] based on out-of-sample variance curves for the same ten datasets. Under this protocol, the reported average risk reductions of 10.52% versus LINC and 12.46% versus QIS are not purely out-of-sample, since the parameter was chosen by inspecting the test results. The authors should either pre-specify δ before seeing the out-of-sample data, select it via a nested procedure within the estimation window, or demonstrate that the rankings in Table 2 are insensitive to δ over a wide interval containing 0.25 for every dataset. The evidence in Figure 4 is suggestive but still relies on the same evaluation period.","section":"Sections 6.1 and 6.3"},{"comment":"The claimed equivalence of ERSE to sequential linear shrinkage on pairs of eigenvalues is established only for the first PER step. Proposition 5 assumes that q1 and q2 are eigenvectors of R_S, so that cross terms in q̂^T R_S q̂ vanish. After the first iteration of Algorithm 2 the rotated vectors are no longer eigenvectors of R_S; for such vectors q̂1^T R_S q̂2 = cosθ sinθ(λ_j − λ_i) is generally nonzero, and the updated λ̂_i is not a convex combination of two current eigenvalues as written in Eq. (38). The statement that ERSE is 'equivalent to performing multiple linear shrinkage on two distinct eigenvalues' therefore needs either a generalized proof for the iterated procedure or a careful restatement limiting the equivalence to the initial pair.","section":"Section 5, Proposition 5"}],"minor_comments":[{"comment":"The vector λ̂ is written as (λ̂_1, . . . , n)^T; it should be (λ̂_1, . . . , λ̂_n)^T.","section":"Section 4, Algorithm 2, line 8"},{"comment":"The asterisk convention is never defined. The tables should state explicitly that asterisks indicate the significance level of the difference between ERSE and the listed strategy, and should name the test (e.g., the stationary bootstrap of Ledoit and Wolf 2011) and the direction of the alternative.","section":"Tables 2, 4, 5, and 6"},{"comment":"The Lagrangian derivation has sign inconsistencies between Eqs. (21) and (24): stationarity gives x = −(1_n + λρ_i)/(2ν), while the text writes x = (1_n − λρ_i)/(2ν). The final bound in Eq. (14) is correct, but the intermediate algebra should be corrected for clarity.","section":"Section 3, Proposition 3 proof"},{"comment":"The caption does not explain what the plotted points and axes represent. Please add a short description of the two-dimensional and three-dimensional illustrations so the reader can interpret the rotation geometrically.","section":"Figure 3"},{"comment":"The citation 'Vaart, 1961' appears with the author name 'Vaart, H. R.'; if this refers to H. R. van der Vaart, the spelling should be standardized.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The reconstruction ambiguity in Algorithm 2 is the most serious technical issue and should be resolved before further review. The δ-selection problem could be mitigated by an extra validation experiment showing that the reported rankings are stable over a reasonable range of δ. The paper is within the scope of the journal and the empirical setup is otherwise carefully executed."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things before anything else. The core idea—rotating sample eigenvectors so each has a minimum deviation from the uniform vector, then using the rotated eigenvectors to compute eigenvalues—is genuinely new relative to the Ledoit–Wolf eigenvalue-only shrinkage literature. The Perron–Frobenius intuition is made precise in Propositions 1–2, and Proposition 5 correctly shows that each paired rotation is pairwise linear shrinkage. The empirical section is broad: ten datasets, subperiod splits, estimation windows from 60 to 240 months, and 150 random subsamples all point the same direction. The authors also explicitly say they are not claiming universal dominance, which is honest.\n\nNow the load-bearing problem. Algorithm 2 does not define the estimator uniquely. Step 8 sets λ̂_i = q̂_i^T R_S q̂_i with the rotated eigenvectors. Step 9 then writes Σ̂ = D_S Q_S Diag(λ̂) Q_S^T D_S. If Q_S is the original sample eigenvector matrix, the rotation is cosmetic—the final matrix uses original eigenvectors and only changes eigenvalues. If Q_S is meant to be the rotated matrix Q̂, then the estimator keeps only the diagonal of Q̂^T R_S Q̂ and discards off-diagonal entries that, after a paired rotation, equal cosθ sinθ(λ_j − λ_i). Those entries are generally not zero, and the paper never analyzes the truncation. Proposition 5 only characterizes the diagonal. The headline 10.52% and 12.46% risk reductions cannot be interpreted until the authors state which construction they implemented. Released code would settle this immediately; its absence is a real problem for a paper whose central claim is empirical.\n\nThe second issue is smaller but real. δ = 0.25 is 'empirically determined' on the same ten datasets used for the headline results. The sensitivity analysis shows a broad sweet spot in [0.15, 0.35], which softens the concern, but an out-of-sample or cross-validated choice would be far more convincing. Also minor: Proposition 3's proof is terse, though the min-over-i bound is valid; and the Sharpe-ratio results are mixed, which the paper does not hide.\n\nThe audience is empirical researchers in covariance estimation and portfolio optimization, especially those working with factor-sorted portfolios. On balance, this deserves a serious referee. The idea is novel, the math is mostly sound, and the empirical scope is above average for this literature. But the current draft is not ready to publish. The estimator must be unambiguously defined, the off-diagonal truncation either analyzed or the algorithm corrected, and δ selection justified out-of-sample. If those are fixed, I'd expect a solid EJOR paper. I would not cite it in its current form.","headline":"Genuinely new rotation idea and broad empirics, but Algorithm 2 is ambiguous and δ is chosen in-sample, so the headline risk reductions are not yet reproducible.","tokens_in":26814,"tokens_out":4239,"would_cite":false,"duration_ms":45052,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62H12","62H25","91G10"],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper claims that rotating the eigenvectors of the sample correlation matrix—rather than only shrinking its eigenvalues—produces a lower-risk covariance estimator for positively correlated assets, cutting out-of-sample…","keywords":["covariance matrix estimation","shrinkage estimator","positively correlated assets","global minimum variance portfolio","eigenvector rotation","rotation-equivariant estimation","factor-sorted portfolios","condition number"],"falsifier":"Re-run Algorithm 2 but keep the full rotated matrix $\\hat{Q}^T R_S\\hat{Q}$ instead of keeping only its diagonal $\\hat\\lambda_i$; if the resulting GMV portfolios have out-of-sample variance similar to or lower than ERSE's, then the claimed improvement is not attributable to the eigenvalue shrinkage that the paper identifies as the mechanism.","tokens_in":25877,"feed_emoji":"📉","tokens_out":8230,"duration_ms":85687,"temperature":0.7,"pith_summary":"The paper claims that when assets are positively correlated, the eigenvectors of the sample correlation matrix corresponding to weak factors sit very close to the null space of the uniform vector, and that pushing these eigenvectors away from that null space—while preserving orthogonality—yields a better covariance matrix for minimum-variance portfolios. The proposed estimator, ERSE, pairs the eigenvector with the smallest deviation with the one with the largest deviation and rotates them until every eigenvector satisfies a minimum deviation threshold. Each rotation is exactly a linear shrinkage of two eigenvalues toward each other while preserving their sum, so ERSE behaves as a data-driven, multi-pair shrinkage method. On ten factor-sorted portfolio datasets, ERSE reports the lowest out-of-sample variance in every dataset, with average risk reductions of 10.52% versus the best linear shrinkage benchmark and 12.46% versus quadratic-inverse nonlinear shrinkage. If these results hold, ERSE gives practitioners a computationally cheap covariance estimator for decision settings where comovement makes most pairwise correlations positive.","feed_headline":"Rotation estimator cuts correlated-asset portfolio risk by 10-12%","feed_subtitle":"Pairwise rotating weak-factor eigenvectors beats linear and nonlinear shrinkage on out-of-sample variance.","key_machinery":"The central object is the Paired Eigenvector Rotation (PER). PER takes two orthogonal unit eigenvectors $q_1,q_2$ with $T(q_1)<\\delta<T(q_2)$, rotates them in their plane via $\\hat{q}_1=\\cos\\theta\\,q_1+\\sin\\theta\\,q_2$ and $\\hat{q}_2=-\\sin\\theta\\,q_1+\\cos\\theta\\,q_2$, and chooses the smallest rotation angle making $T(\\hat{q}_1)=\\delta$. This preserves orthogonality and the sum of the two deviation degrees, and it transforms the paired eigenvalues by $\\hat\\lambda_1=\\cos^2\\theta\\,\\lambda_1+\\sin^2\\theta\\,\\lambda_2$ and $\\hat\\lambda_2=\\sin^2\\theta\\,\\lambda_1+\\cos^2\\theta\\,\\lambda_2$. ERSE iterates PER at most $n-1$ times, always pairing the current smallest-deviation eigenvector with the largest-deviation eigenvector, then rebuilds eigenvalues as $\\hat\\lambda_i=\\hat{q}_i^T R_S\\hat{q}_i$ and the covariance estimate as $D_S\\hat{Q}\\,\\text{Diag}(\\hat\\lambda)\\,\\hat{Q}^T D_S$. The machinery converts an eigenvector alignment constraint into an eigenvalue-shrinkage scheme without directly modifying the sample covariance or correlation matrix.","core_discovery":"The discovery is a structural fact about positively correlated assets plus an estimator built on it. For such assets, the sample correlation matrix has a dominant eigenvector whose deviation degree $T(q)=(\\mathbf{1}_n^T q)^2$ is larger than its largest eigenvalue, while the remaining eigenvectors are squeezed toward the null space of the uniform vector; Propositions 1–3 show that the deviation degrees sum to $n$ and that weak-factor eigenvectors have deviation bounded near zero when average correlation is high. The paper argues that moderately amplifying these weak-factor deviations is the right shrinkage direction, and it constructs ERSE to do this. Algorithm 2 pairs the eigenvector with the smallest deviation against the one with the largest deviation and applies a Paired Eigenvector Rotation until every eigenvector satisfies $T(\\hat{q}_i)\\ge\\delta$. Because one rotation sends two eigenvalues to $\\cos^2\\theta\\,\\lambda_a+\\sin^2\\theta\\,\\lambda_b$ and $\\sin^2\\theta\\,\\lambda_a+\\cos^2\\theta\\,\\lambda_b$, ERSE is equivalent to repeated linear shrinkage on pairs of eigenvalues, pulling small eigenvalues up and large ones down while preserving their sum. The empirical claim is that plugging ERSE into the global minimum-variance portfolio achieves the lowest out-of-sample variance on all ten datasets, with statistically significant gains in most comparisons and average reductions of 10.52% versus LINC and 12.46% versus QIS.","pith_inferences":["Our inference: because the positive-correlation assumption is the only structural input, ERSE should transfer to other positively correlated universes such as government-bond portfolios or country-industry indices; the paper does not test these.","Our inference: the threshold $\\delta=0.25$ is selected using the same datasets that are later used for evaluation, so the headline reductions likely include an in-sample tuning component; a held-out calibration would be needed to confirm the out-of-sample edge.","Our inference: Algorithm 2 drops the off-diagonal entries of $\\hat{Q}^T R_S\\hat{Q}$ after rotation, and no analysis quantifies the resulting information loss; keeping those entries or shrinking them separately is a natural modification that could change the risk comparison.","Our inference: since each PER pair is exact two-eigenvalue linear shrinkage, the same rotation geometry could produce adaptive shrinkage by letting $\\delta$ vary across eigenvectors or by targeting a different aggregate rather than a uniform floor."],"forward_implications":["GMV portfolios built from ERSE should have lower out-of-sample variance than portfolios built from the compared linear and nonlinear shrinkage estimators whenever the sample correlation matrix is positive, with average reductions near 10.5% and 12.5% on factor-sorted portfolios.","ERSE covariance estimates have lower condition numbers than the benchmark estimates, which translates into more concentrated, more stable portfolio weights and modest short positions in positively correlated settings.","The out-of-sample risk reduction is stable: it appears in both subperiods, across estimation windows from 60 to 240 months, and in all 150 random 200-asset subsamples drawn from the largest dataset.","ERSE has low computational cost, requiring at most $n-1$ closed-form rotations per estimation window, making it feasible for portfolios with hundreds of assets."],"supporting_citations":[{"why":"Supplies the linear shrinkage baseline and the eigenvalue-shrinkage identity that ERSE generalizes by shrinking pairs of eigenvalues.","marker":"Ledoit and Wolf (2004b)"},{"why":"Defines the rotation-equivariant oracle $\\hat\\lambda_i=q_i^T\\Sigma q_i$ and the framework in which ERSE revises eigenvalues while keeping the sample eigenvectors.","marker":"Ledoit and Wolf (2017)"},{"why":"Documents eigenvalue overdispersion, argues eigenvectors are hard to improve, and provides the dataset-construction scheme used for the larger test sets.","marker":"Shi et al. (2020)"},{"why":"Establishes the $\\ell^2$-norm portfolio constraint that Proposition 6 equates with ERSE's deviation threshold.","marker":"DeMiguel et al. (2009a)"},{"why":"Defines the LIS, QIS, and GIS nonlinear shrinkage benchmarks that ERSE is compared against.","marker":"Ledoit and Wolf (2022b)"},{"why":"Provides the single-dominant-factor evidence for large covariance matrices that motivates concentrating on weak-factor eigenvectors.","marker":"Dai et al. (2024)"},{"why":"Supplies the stationary bootstrap test used to claim statistical significance for the variance reductions.","marker":"Ledoit and Wolf (2011)"}],"fun_headline_variants":["Eigenvector rotation trims correlated-asset risk 10-12% over shrinkage","Pairwise rotation beats shrinkage on correlated assets by 10-12%","ERSE rotation cuts variance 10-12% vs linear and nonlinear shrinkage","Rotate weak-factor eigenvectors to slash correlated-asset portfolio risk 10-12%","Eigenvector rotation outperforms shrinkage on correlated assets, cutting risk 10-12%"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The paper silently assumes that discarding the off-diagonal entries of the rotated correlation matrix in the eigenvalue reconstruction step does not materially distort the estimate; it never analyzes this truncation, so ERSE's gains could partly reflect an unintended distortion rather than the intended shrinkage.","fun_headline_variants_meta":{"raw":{"variants":["Eigenvector rotation trims correlated-asset risk 10-12% over shrinkage","Pairwise rotation beats shrinkage on correlated assets by 10-12%","ERSE rotation cuts variance 10-12% vs linear and nonlinear shrinkage","Rotate weak-factor eigenvectors to slash correlated-asset portfolio risk 10-12%","Eigenvector rotation outperforms shrinkage on correlated assets, cutting risk 10-12%"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001222,"raw_usage":{"total_tokens":5048,"prompt_tokens":990,"completion_tokens":4058,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":606,"completion_tokens_details":{"reasoning_tokens":3951}},"tokens_in":606,"tokens_out":4058,"duration_ms":30814,"temperature":1.0,"reasoning_tokens":3951,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T20:49:45.278827+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run Algorithm 2 but keep the full rotated matrix $\\hat{Q}^T R_S\\hat{Q}$ instead of keeping only its diagonal $\\hat\\lambda_i$; if the resulting GMV portfolios have out-of-sample variance similar to or lower than ERSE's, then the claimed improvement is not attributable to the eigenvalue shrinkage that the paper identifies as the mechanism.","supporting_citations":[{"cited_title":"and Wolf, M","cited_arxiv_id":null,"evidence_quote":"Defines the rotation-equivariant oracle $\\hat\\lambda_i=q_i^T\\Sigma q_i$ and the framework in which ERSE revises eigenvalues while keeping the sample eigenvectors."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Documents eigenvalue overdispersion, argues eigenvectors are hard to improve, and provides the dataset-construction scheme used for the larger test sets."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the single-dominant-factor evidence for large covariance matrices that motivates concentrating on weak-factor eigenvectors."},{"cited_title":"and Wolf, M","cited_arxiv_id":null,"evidence_quote":"Supplies the stationary bootstrap test used to claim statistical significance for the variance reductions."}],"review_version":1}