{"id":"3440d161-83ae-4daa-8fa6-86677652c5eb","arxiv_id":"2501.19260","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":9,"one_line_summary":"Heterogeneous investment sizes (a binary big/small mixture) destabilize the financial network and the commonly used Corsi approximation underestimates this instability.","lead":"This paper extends a standard model of financial contagion to allow banks to make large and small investments, and finds that such heterogeneity pushes the system toward instability. It also shows that a common analytical shortcut, the Corsi approximation, underestimates this risk, while a replica-based method tracks the numerical transition more closely.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Method #3's first-moment matching of W by cX is unvalidated at q=8, so the 'never false stable' claim and Figure 3 boundaries rest on an unquantified proxy.","rationale":"The reader's conditional verdict is appropriate. The paper's core qualitative result—investment-size heterogeneity moves the system toward instability—is demonstrated by direct numerical diagonalization (method #1), which is exact for finite matrices and does not depend on the cX proxy. The same mean investment is enforced by Eq. (4), so the homogeneous-versus-heterogeneous comparison is fair. Thus I do not see a viable attack on the central claim itself. The load-bearing weakness is the paper's strongest methodological selling point: method #3 is claimed to be reliable enough that it 'will never result in a false stable diagnosis,' and Figure 3 presents its phase boundary as the safe estimate. That claim rests entirely on replacing the correlated, column-stochastic W by cX and applying the unpublished replica result [17]. Matching first moments (Eq. 25) is a weak constraint for a spectral quantity: E[lambda_max] is governed by the top eigenvalue of W W^T, which depends on column normalization, correlations, and sparsity in a way that E[W_ij] does not capture. Appendix C's derivation of c even uses the unconditional expectation including zero-investment columns, whereas the W used in the model is column-stochastic; the discrepancy is small at alpha q ~ 6.5-9.2, but it is not quantified. Figure 7's relative-gap curves do not report the value at q=8, the parameter of Figures 2-3, and the text only says to 'select a value of q for which our approximation is under 5% off.' A direct test—comparing the exact and proxy phase boundaries at q=8—would settle whether the proxy can misclassify stable as unstable (a safe error) or unstable as stable (the claimed impossibility). If the latter occurs anywhere, Section 4.3's safety claim is overbroad and the Figure 3 transition lines need caveats, but the qualitative heterogeneity result survives. Hence the verdict should remain conditional: the central claim is credible, but the method #3 guarantee and phase boundaries need verification or qualification.","tokens_in":20342,"tokens_out":7171,"duration_ms":69780,"concrete_test":"At the exact parameters of Figure 3 (alpha=sqrt(200/300) and sqrt(400/300), q=8, zeta=1.85, sigma_s^2=0.009, sigma_d^2=0.03, gamma=50), generate 10^4 realizations of X for a dense grid of (pB, phi). For each realization compute lambda_max(W W^T) and lambda_max(c^2 X X^T) with c from Eq. (26), and average over realizations. Mark the E[lambda_max]=1 contours for both. If at any grid point the cX contour lies below 1 while the exact W W^T contour lies above 1, method #3 has produced a false stable diagnosis and the 'never false stable' claim in Section 4.3 is refuted. Reporting the signed relative gap at q=8 in the style of Figure 7 would also settle whether the paper's stated 'under 5% off' selection criterion is actually met at the parameter values used for the phase diagrams.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 5.2 replaces the column-stochastic, correlated matrix W by cX, with c=(1-e^{-alpha q})/(alpha q) fixed by matching first moments (Eq. 25, Appendix C). This proxy is then fed into the replica solver of [17] to produce the method #3 transition lines in Figure 3 and the claim that method #3 'will never result in a false stable diagnosis.' First-moment matching does not control the top eigenvalue of W W^T, which depends on column normalization, correlations, and sparsity rather than on the mean entry alone. Appendix C actually computes E[W_ij] unconditionally, including institutions with zero investments; the exact column-stochastic W has E[W_ij]=P(active)/N, so the two differ by the fraction of empty columns, a discrepancy that is small at alpha q ~ 6.5-9.2 but is not quantified. Figure 7 reports relative gaps in E[lambda_max] only in aggregate and does not state the gap at q=8, the value used for the phase diagrams. If the cX proxy overestimates or underestimates lambda_max differently across the (pB, phi) plane, the method #3 boundary can cross the true transition, and the 'no false stable' guarantee fails. The qualitative finding that heterogeneity destabilizes is supported by direct numerical diagonalization (method #1), so the central policy conclusion does not collapse; what is at risk is the quantitative phase boundary and the safety claim attached to method #3.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies a random-matrix model of portfolio rebalancing in which M financial institutions invest in N assets, with heterogeneous investment sizes drawn from a two-point distribution. The stability of the linear price dynamics is governed by the average largest eigenvalue E[lambda_max] of the matrix Phi = ((eta-1)/gamma) alpha^2 W W^T, where W is the column-stochastic weight matrix. The authors evaluate E[lambda_max] by direct numerical diagonalization, by the largest eigenvalue of E[Phi] (the \"Corsi\" method), and by a replica calculation applied to the approximate matrix c X X^T, with c fixed by matching first moments. They find that increasing investment-size heterogeneity (parameter phi) pushes the system toward instability, that increasing diversification q is stabilizing at low q and destabilizing at high q, and that the Corsi approximation underestimates systemic risk whereas the replica method is more accurate and, in the simulations, conservative in the sense of never misclassifying an unstable system as stable.","tokens_in":20637,"tokens_out":15270,"duration_ms":131826,"significance":"If the result holds, the paper makes a useful policy-relevant point: homogeneous-investment models can understate systemic risk, and the stability phase diagram depends non-trivially on the interaction between diversification and heterogeneity. The paper's main qualitative finding is supported by direct numerical diagonalization of the full random matrix, which does not rely on the two analytical approximations, and the comparison of three independent methods is a strength. The authors are also transparent about the failure of the Corsi approximation, which is itself a useful cautionary result. However, the quantitative reliability of the replica-based phase boundaries and the associated safety claim require additional validation, so the paper in its current form is not ready for acceptance.","major_comments":[{"comment":"The cX proxy used by method #3 is calibrated only by matching first moments, which does not control the top eigenvalue of W W^T. Figure 7 reports the relative gap in E[lambda_max] only for a single parameter set (B=3, s=0.3, pB=7/27) and does not state the gap at q=8, the connectivity used in the phase diagrams of Figures 2 and 3. Since B and s vary strongly across the (pB, phi) plane (for phi close to 1 and small pB, B becomes very large), the accuracy of the proxy at q=8 must be quantified across that plane before the method #3 transition lines can be used as a quantitative phase boundary.","section":"Section 5.2, Eq. (25) and Figure 7"},{"comment":"The statement \"Using method #3 will never result in a false stable diagnosis\" is a universal safety claim, but the evidence is empirical: Figure 3 shows no false-stable region for the finite set of parameters tested at q=8. Nothing in the first-moment matching of Eq. (25) guarantees that the proxy underestimates or overestimates lambda_max in a fixed direction. Please either prove a bound on the approximation error or replace the sentence with a weaker claim, e.g., \"no false stable diagnoses were observed in our simulations.\"","section":"Section 4.3"},{"comment":"The derivation of c in Eq. (C.9) uses only the unconditional mean E[W_ij]. However, the exact W is column-stochastic with column sums S_j = sum_i X_ij, and the top eigenvalue of W W^T is sensitive to the distribution of these column sums, because W_ij = X_ij / S_j takes large values when S_j is small. The paper does not check whether the distribution of S_j (or the second moment of W W^T) is reproduced by cX. A second-moment-matched proxy, or a direct report of the S_j statistics, would be needed to justify the replica-based phase boundary.","section":"Appendix C"}],"minor_comments":[{"comment":"The arrow meanings appear inconsistent: both rows for \"system unstable\" use a down arrow and both rows for \"system stable\" use an up arrow, regardless of whether the approximate method is correct. Clarify whether the arrow denotes the actual state or the approximate method's verdict, and adjust the entries so that the four cases are distinguishable.","section":"Table 1"},{"comment":"The text says \"each asset has E[sum_j X_ij] = q/alpha institutions investing in it on average,\" but sum_j X_ij is the total monetary investment in the asset, not a count of institutions. The count is sum_j delta_{c_ij,1}; the two quantities coincide only because the mean investment size is normalized to 1. Please rephrase.","section":"Section 3.1 and Eq. (14)"},{"comment":"The same symbol Phi is used for the exact matrix in Eq. (7) and for the approximate matrix in Eq. (26); using a distinct notation (e.g., hat_Phi) would avoid confusion.","section":"Section 5.2, Eq. (26)"},{"comment":"The number of Monte Carlo samples is not stated for Figure 5, although Figures 6 and 7 mention 10^4 realizations; please specify the sample size for all numerical averaging and indicate whether error bars are smaller than the marker size.","section":"Figures 5 and 7"},{"comment":"The typeset formulas for the diagonal and off-diagonal entries of E[Phi] and for the largest eigenvalue are difficult to parse in the current version, with fractions and square roots not cleanly separated; please ensure the final version renders these formulas unambiguously.","section":"Equations (17)-(19)"}],"recommendation":"major_revision","confidential_remarks":"The paper relies heavily on the companion preprint [17] for the replica calculation, which the referee could not check; it would strengthen the submission to include a self-contained summary of the relevant replica equations or to submit [17] alongside. The safety claim about method #3 is stronger than the evidence presented and should be softened or supported. The central qualitative result, however, is convincing because it is visible in the direct numerical diagonalization, and the paper is generally well organized."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper does something real: it pushes the Corsi et al. portfolio-rebalancing model beyond the homogeneous 1/N rule and shows, mostly through direct numerical diagonalization, that heterogeneity in investment sizes pushes the stability transition toward instability. The central qualitative result is solid, and the paper deserves credit for isolating the heterogeneity effect via the constraint pBB + pss = 1, which keeps the average investment size fixed. The finding that the widely used Corsi approximation (method #2) severely underestimates risk in the heterogeneous case is also well supported and practically relevant.\n\nThe soft spots are real but not fatal. The biggest issue is the claim that the replica method (method #3) “will never result in a false stable diagnosis.” That guarantee rests on a first-moment matching of the correlated, column-stochastic matrix W by cX. Matching first moments does not control the top eigenvalue of WW^T, and the paper only validates the proxy through aggregate relative gaps (Figure 7), not at the q=8 parameters used in the phase diagrams. The auxiliary calculation in Appendix C also conditions differently on active investments, which is fine but not quantified. So the qualitative phase structure is credible, but the quantitative boundary from method #3 and the “never false stable” safety claim are not established at the level the paper asserts. A related concern: method #3 relies on an unpublished companion paper [17], which is acceptable but should be made available or at least clearly summarized.\n\nMinor points: the model has many free parameters and is not calibrated to data, which is normal for a stylized theoretical model but limits direct policy takeaways. The paper also provides no code or data, making the numerical figures hard to reproduce. And the literature review is broad but somewhat bloated, with several citations that are not clearly tied to the model.\n\nOverall, the paper is a genuine contribution to the quantitative literature on portfolio overlap and systemic risk. It is written clearly and the main claims (heterogeneity destabilizes, Corsi underestimates risk) are backed by unambiguous numerics. The reviewer's job should be to push for a more honest statement of method #3's limitations, a robustness check of the cX approximation at the parameters actually used, and ideally a release of code. This deserves a serious referee, not a desk reject. I would accept it for review, and I'd probably bring it to our reading group; the results might be worth citing once the overclaim is softened.","headline":"A well-executed extension of Corsi et al. that convincingly shows investment-size heterogeneity destabilizes the system, but the paper overstates the safety guarantee of its replica method and leans on an unvalidated matrix proxy.","tokens_in":21230,"tokens_out":1469,"would_cite":true,"duration_ms":17120,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Uneven investment sizes push a rebalancing financial network across the stability boundary, even where equal-weight models predict safety.","keywords":["systemic risk","financial contagion","portfolio rebalancing","portfolio diversification","random matrix theory","largest eigenvalue","replica method","heterogeneous investments"],"falsifier":"Using the parameter values of Figures 2 and 3 ($q=8$, $\\zeta=1.85$, $\\sigma_s^2=0.009$, $\\sigma_d^2=0.03$, $\\gamma=50$), compute the replica method's predicted transition line and compare it cell-by-cell with $E[\\lambda_{\\max}]$ from numerical diagonalization of $10^4$ realizations of $\\Phi$ over the $(\\phi, p_B)$ grid; any cell in which the replica method predicts stability while diagonalization gives $E[\\lambda_{\\max}] > 1$ would falsify the claim that Method #3 never issues a false stable verdict.","tokens_in":20103,"feed_emoji":"📉","tokens_out":13354,"duration_ms":109048,"temperature":0.7,"pith_summary":"This paper extends an established random-matrix model of financial contagion through overlapping portfolios by letting each institution split its money unevenly between small and large positions, and asks whether this added realism changes the stability verdict. The answer it argues for is yes: the boundary between stable and unstable markets is set by the average largest eigenvalue of the matrix that drives the endogenous part of returns, and investment-size heterogeneity systematically raises that eigenvalue, so a market that looks calm under the usual equal-weight investment rule can be genuinely unstable. The paper also identifies two competing effects of diversification: widening the number of assets each bank holds deepens the market and stabilizes prices at low connectivity, but beyond a threshold it destabilizes the system by creating more overlapping portfolios. Finally, it shows that the standard shortcut of replacing the random matrix by its average badly underestimates the risk, whereas a replica-based calculation tracks the transition much more closely.","feed_headline":"Unequal bank bets destabilize financial networks","feed_subtitle":"Investment-size gaps trigger runaway volatility where the 1/N rule predicts calm.","key_machinery":"The load-bearing object is the average largest eigenvalue $E[\\lambda_{\\max}]$ of the random matrix $\\Phi = \\frac{\\eta-1}{\\gamma \\alpha^2} W W^T$, which encodes the feedback loop in which price changes force portfolio rebalancing, which in turn moves prices; the stability/instability transition is the crossing of the threshold $E[\\lambda_{\\max}] = 1$. Because the entries of the column-stochastic weight matrix $W$ are correlated through the column normalization, the paper computes this quantity two approximate ways: Method #2 takes the largest eigenvalue of the average matrix $E[\\Phi]$, while Method #3 uses a replica representation of the top eigenvalue as the zero-temperature free energy of a partition function (the machinery of [17]) applied to a surrogate matrix in which $W$ is replaced by $c X$ with $c = (1-e^{-\\alpha q})/(\\alpha q)$, turning the problem into the top-eigenvalue statistics of a diluted Wishart matrix $X X^T$ with independent sparse entries. Population dynamics solves the resulting distributional equations, and direct numerical diagonalization of $\\Phi$ serves as the reference standard.","core_discovery":"The central claim, stated on the paper's own terms, is that the typical stability of an investment network is governed by the average largest eigenvalue of $\\Phi = \\frac{\\eta-1}{\\gamma \\alpha^2} W W^T$, the matrix appearing in the return dynamics $e_t = \\Phi(e_{t-1}+\\varepsilon_t)$; whenever $E[\\lambda_{\\max}] > 1$, at least one endogenous return process grows without bound and the market is unstable. With a binary Big/Small investment distribution constrained so that the mean investment per institution is unchanged, the paper finds that heterogeneity raises $E[\\lambda_{\\max}]$ and can convert a stable homogeneous market into an unstable one: small positions propagate shocks between institutions during rebalancing, while large positions magnify the resulting price moves. Diversification acts non-monotonically, deepening markets at low $q$ and increasing portfolio overlap at high $q$, and the paper shows that the largest-eigenvalue-of-the-average-matrix method (Method #2) underestimates this risk, whereas the replica approximation (Method #3), applied to a surrogate matrix with $W$ replaced by $cX$, $c = (1-e^{-\\alpha q})/(\\alpha q)$, tracks the numerical transition without ever issuing a false stable verdict.","pith_inferences":["An untested consequence of the model's mechanism is that the destabilizing effect should strengthen as the investment-size distribution becomes heavy-tailed: with the same mean constraint, rare very large positions would amplify the rebalancing sell-off effect, pushing the instability region toward lower heterogeneity values than the binary distribution does.","The paper verifies the Method #3 proxy only through an aggregate relative-gap curve; a sharper test is a pointwise comparison of the replica phase boundary and numerical diagonalization on the full $(\\phi, p_B)$ grid at $q = 8$, since any cell where the proxy predicts stability while diagonalization gives $E[\\lambda_{\\max}] > 1$ would falsify the never-a-false-all-clear claim for that parameter se","Because the transition is defined by an average eigenvalue, a finite real market near the threshold could cross into instability through fluctuations of $\\lambda_{\\max}$ even when $E[\\lambda_{\\max}]$ is slightly below 1; the paper lists eigenvalue fluctuations as future work, and this observation suggests the safe regulatory region is inside, not merely on, the stable side of the boundary."],"forward_implications":["A market that the homogeneous 1/N rule classifies as stable can be unstable once banks are allowed to hold large and small positions at the same average total investment, so risk assessments built on the 1/N assumption are systematically over-optimistic.","Raising the diversification parameter $q$ first decreases $E[\\lambda_{\\max}]$, because more holders per asset means deeper markets and smaller price impact, then increases it through portfolio overlap, producing a U-shaped stability response with a minimum near $q \\approx 10$ in the tested regimes.","Method #2 underestimates $E[\\lambda_{\\max}]$ substantially at low connectivity and can misclassify unstable markets as stable, while Method #3 may overestimate the risk but never labels an unstable market as stable, so a regulator using Method #3 would never receive a false all-clear.","In the asset-rich regime with $N > M$, representative of globally systemically important banks, each asset is held by very few institutions, so individual trades are a large fraction of the asset's volume and $E[\\lambda_{\\max}]$ reaches higher values than in the asset-poor regime."],"supporting_citations":[{"why":"Supplies the base rebalancing model, the $\\Phi$ matrix, the $E[\\lambda_{\\max}]$ stability criterion, and the Method #2 approximation that the paper extends with heterogeneous investment sizes.","marker":"[1]"},{"why":"Provides the replica computation of the top eigenvalue of diluted Wishart matrices that Method #3 applies after substituting the proxy matrix $c^2 X X^T$.","marker":"[17]"},{"why":"Source of the linear price-impact model used in Appendix A to derive the closed dynamics for the endogenous return component.","marker":"[28]"},{"why":"Establishes the overlapping-portfolio contagion framework and the high-connectivity diversification regime the paper's $q$-dependence builds on.","marker":"[6]"},{"why":"A Langevin price-dynamics model cited alongside [28] as the basis for the price-impact equation.","marker":"[52]"}],"fun_headline_variants":["Unequal bets flip stable markets into turmoil","Diversification helps sparse, hurts dense networks","Standard risk model misses unequal portfolio hazards","Heterogeneous investments destabilize markets"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The replica method's phase boundary assumes that matching the average entry of the weight matrix $W$, whose columns each sum to one, with the scaled matrix $c X$ is enough to reproduce the largest eigenvalue of $W W^T$; the paper tests this only in aggregate, not at the $q = 8$ used for the phase diagrams.","fun_headline_variants_meta":{"raw":{"variants":["Unequal bets flip stable markets into turmoil","Diversification helps sparse, hurts dense networks","Standard risk model misses unequal portfolio hazards","Heterogeneous investments destabilize markets"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00155,"raw_usage":{"total_tokens":6160,"prompt_tokens":874,"completion_tokens":5286,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":490,"completion_tokens_details":{"reasoning_tokens":5232}},"tokens_in":490,"tokens_out":5286,"duration_ms":37385,"temperature":1.0,"reasoning_tokens":5232,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-09T20:46:28.782506+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Using the parameter values of Figures 2 and 3 ($q=8$, $\\zeta=1.85$, $\\sigma_s^2=0.009$, $\\sigma_d^2=0.03$, $\\gamma=50$), compute the replica method's predicted transition line and compare it cell-by-cell with $E[\\lambda_{\\max}]$ from numerical diagonalization of $10^4$ realizations of $\\Phi$ over the $(\\phi, p_B)$ grid; any cell in which the replica method predicts stability while diagonalization gives $E[\\lambda_{\\max}] > 1$ would falsify the claim that Method #3 never issues a false stable verdict.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the base rebalancing model, the $\\Phi$ matrix, the $E[\\lambda_{\\max}]$ stability criterion, and the Method #2 approximation that the paper extends with heterogeneous investment sizes."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the replica computation of the top eigenvalue of diluted Wishart matrices that Method #3 applies after substituting the proxy matrix $c^2 X X^T$."},{"cited_title":"D., Foti, N","cited_arxiv_id":null,"evidence_quote":"Source of the linear price-impact model used in Appendix A to derive the closed dynamics for the endogenous return component."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Establishes the overlapping-portfolio contagion framework and the high-connectivity diversification regime the paper's $q$-dependence builds on."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"A Langevin price-dynamics model cited alongside [28] as the basis for the price-impact equation."}],"review_version":1}