{"id":"1c5f4ae2-a88f-4e88-aed1-d54592e8bcaa","arxiv_id":"2608.11266","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A behavioral model shows that loss-averse, paycheck-to-paycheck depositors can trigger bank runs when they assign high probability to bad income states, and a Call Report exercise finds modest, imprecise empirical support.","lead":"This paper builds a theory of bank runs in which workers live paycheck to paycheck and fear losses so much that they may suddenly demand their cash. It adds a proof-of-concept test on U.S. bank data, where balance-sheet proxies for that fear add only a little explanatory power.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The Bank Run Exposure State Space rests on Eq (3.3)-(3.4), where the loss-aversion index is conflated with a marginal-utility ratio; under canonical Tversky-Kahneman values the two differ, so the taboo-state bounds do not follow.","rationale":"The reader's verdict is CONDITIONAL, and my analysis does not move the verdict: the paper's central behavioral mechanism, including Lemma 2's stress-state condition, is a plausible sufficient condition, but the formal construction of the Bank Run Exposure State Space in Section 3 has a genuine gap. The reader identified Assumption 8 (linear separability) as the weakest assumption; my concern is adjacent but distinct: even granting Assumption 8, Eq (3.3) misidentifies the loss-aversion index with a marginal value ratio at an arbitrary point, and the bounds in Eq (3.4) are asserted, not derived. Under the paper's own Tversky-Kahneman calibration, the ratio equals 1 rather than λ≈2.25, so the taboo-state set and Theorem 3.1 lose their basis. This is load-bearing because the state space is the headline claim of the paper, and it is in Section 3, not the secondary probability or martingale representations. The issue is repairable by re-deriving the state space directly from Lemma 2 (e.g., taking Ω_run = Ω_admiss or introducing a correct bound on λ(c1)), and the empirical section is explicitly proof-of-concept, so a REJECT verdict would be too strong. A CONDITIONAL verdict requiring the Section 3 derivation to be fixed remains appropriate; hence UNCHANGED.","tokens_in":30207,"tokens_out":9320,"duration_ms":83691,"concrete_test":"Independently re-derive Eq (3.4) from Theorem 2.1 and Assumption 8 using the canonical Tversky-Kahneman value function v_g(c)=c^α, v_ℓ(ℓ)=ℓ^β with α=β=0.88 and λ=2.25. Compute λ(c1)=v'_g(c1)/v'_ℓ(-c1) for c1>0 and check whether it equals 1 (not 2.25) and whether it satisfies k/k-tilde ≤ λ(c1) < 1/k-tilde for the calibrated k and k-tilde. If the bounds fail, the taboo-state space construction in Section 3 is invalid as written.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 3 defines the paper's central object, the Bank Run Exposure State Space Ω_run = Ω_admiss ∩ Ω_taboo, and Theorem 3.1 asserts that this space is induced by loss aversion and large subjective bad-state probabilities. The taboo set is built on Eq (3.3)-(3.7). Eq (3.3) sets λ(c1_ω') = v'_g(c1_ω')/v'_ℓ(-c1_ω'), treating the marginal value ratio at an arbitrary consumption change c1 as the loss-aversion index. But in the model's value function (2.1), λ is a separate multiplicative parameter on the loss branch, and the Kőbberling-Wakker index is the limit of this ratio as c↓0, not at an arbitrary c1. Under the paper's own calibration, with Tversky-Kahneman power forms v_g(c)=c^α, v_ℓ(ℓ)=ℓ^β and symmetric α=β=0.88, the ratio v'_g(c1)/v'_ℓ(-c1) equals 1 for all c1, not the calibrated λ≈2.25. The asserted bounds (3.4), k/k-tilde ≤ λ(c1) < 1/k-tilde, and the consequent taboo set (3.8), are not derived from Theorem 2.1: Theorem 2.1 with separability gives only λ(ω') ≥ k(ρ,d)/k-tilde(ρ,d), and provides no upper bound. Thus Ω_taboo, and the pullback-topology construction in Theorem 3.1, lack a valid derivation. Lemma 2's trigger condition may survive, but the 'Bank Run Exposure State Space' as the intersection with Ω_taboo is unsupported; the central claim's formal content is therefore not established as written.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper develops a behavioral model of bank run exposure in which paycheck-to-paycheck depositors are loss averse over changes in income and consumption. The central mechanism is a sufficient condition (Lemma 2) that a run is triggered when the loss-aversion index exceeds a probability-weighted ratio of marginal value in gains versus losses. On this basis the paper constructs a 'Bank Run Exposure State Space' (Section 3), derives probability estimates for suspension of convertibility via a stopped process (Section 4), and claims a martingale representation for the exposure process (Section 5). A proof-of-concept empirical section uses quarterly Call Report data for 232 banks to compare a retail-share proxy with a composite balance-sheet proxy, finding modest R² improvements and proxy-sensitive interaction terms. The paper is candid that the empirical exercise is not a definitive test of the behavioral mechanism. However, several load-bearing formal steps, especially the definition of the taboo state space, the derivation of Lemma 2, and the martingale representation, are either incorrect or unsupported as written.","tokens_in":30678,"tokens_out":6919,"duration_ms":63148,"significance":"The topic is timely, particularly after the 2023 SVB episode, and the idea of endogenizing liquidity demand through loss aversion is a useful complement to the exogenous stress parameters in Basel III-type regulation. If the formal results were established, the paper would contribute a behaviorally grounded characterization of run-prone states and a dynamic exposure measure. The empirical implementation is transparently labeled as a proof of concept and provides a detailed data-cleaning discipline, which is commendable. The paper also makes falsifiable predictions about the role of subjective bad-state probabilities. However, the significance is undercut by the fact that the central state-space theorem and the martingale representation are not validly derived as they stand, so the theoretical contribution cannot currently be taken as established.","major_comments":[{"comment":"The definition λ(c1ω') = v'_g(c1ω')/v'_ℓ(-c1ω') conflates the Kőbberling-Wakker loss-aversion index, which is a limit of this ratio at the reference point c=0, with the marginal-value ratio at an arbitrary consumption change c1. Under the paper's own Tversky-Kahneman calibration with symmetric power exponents, v'_g(c)/v'_ℓ(-c) equals 1 for all c, not the calibrated λ≈2.25. Moreover, Theorem 2.1 provides only the lower bound λ(ω') ≥ k/k-tilde; the upper bound λ(c1) < 1/k-tilde in (3.4) appears without any derivation. Consequently, the taboo set Ω_taboo in (3.8) and (3.11), and the pullback-topology construction in Theorem 3.1, are not supported by the preceding results. The central state-space object of the paper therefore lacks a valid derivation and must be re-derived or replaced.","section":"Section 3, Eqs. (3.3)-(3.8)"},{"comment":"The derivation of the trigger condition starts from inequality (8.32), which includes the indicator I{c1ω'<0}. The subsequent text cancels this indicator because 'π−ω' > 0 in a loss state', but this cancellation is only valid when c1ω'<0, i.e., when the state is actually a loss state. Lemma 2 as stated in (2.27) does not impose this sign restriction on c1ω'. If c1ω'≥0, the left side of (8.32) is zero, so (2.27) is not a sufficient condition for a run trigger. The lemma and its proof need to either restrict ω' to states with c1ω'<0 or demonstrate that the condition in (2.27) is sufficient without the indicator, which the current argument does not do.","section":"Section 8.3, proof of Lemma 2"},{"comment":"The martingale representation (5.3) with the specific integrand b(u)=σ_x,N(u)/∫₀^T σ²_x,N(v)dv is asserted without derivation. The paper's appeal to the martingale-difference property of correctly specified regression residuals does not imply that the least-squares exposure coefficient admits a representation of this form; one would need a genuine martingale representation theorem applied to the projection coefficient as a stochastic process, together with verification of its hypotheses. As it stands, the main dynamic result of Section 5 is unsupported. The authors should either provide a rigorous derivation of the representation or state the precise theorem and conditions under which it holds.","section":"Section 5, Theorem 5.1"},{"comment":"The negative binomial formula in (4.3) uses τ_n(ω), which is defined in (4.1) as a random stopping time, as the fixed parameter of a negative binomial distribution. In a standard negative binomial, the number of 'successes' (or failures) before the stopping event is fixed; substituting a random stopping time changes the distribution into a mixture over the law of τ_n, so the formula (4.3) is not a valid probability mass function as written. The Poisson approximation in (4.4) inherits the same problem because its mean parameter τ_n(ω)(1-π−ω)/π−ω is random. The authors should condition on τ_n and integrate over its distribution, or redefine τ_n as a fixed threshold.","section":"Section 4, Proposition 2"},{"comment":"The regression-weighted composite proxy is constructed in-sample: the weights in (6.8) are estimated from the interaction coefficients in the same regressions that are then used to evaluate the composite's fit. The reported improvement in R² (0.043 to 0.049 in Table 3) is therefore a within-sample comparison and does not establish out-of-sample content. Moreover, the within R² is reported as 0.000 in all specifications, meaning that after bank and quarter fixed effects the regressors have no explanatory power within banks. The text should present the result as purely illustrative, and the claim of a 'modest improvement' should be qualified by the in-sample weighting scheme and the zero within fit.","section":"Section 6, Tables 3-5"}],"minor_comments":[{"comment":"The symbol σ is used both for the elasticity of intertemporal substitution in (2.18)-(2.19) and for liquidity risk σx in (2.7) and elsewhere; this dual usage is confusing and should be disambiguated.","section":"Section 2.2.2, Eqs. (2.18)-(2.19)"},{"comment":"The proof of Theorem 3.1 is a single sentence that refers to the Internet Appendix for the formal topology, measurability, and random-field details, but the Internet Appendix is not included with the manuscript. The formal construction therefore cannot be checked by a reader or referee.","section":"Section 8.4, proof of Theorem 3.1"},{"comment":"The phrase 'constructed defensively' is vague; please specify the exact rule for when RCFD2200 is used versus the fallback construction, and why the fallback is preferred for all observations in the cleaned panel.","section":"Section 6.2, Eq. (6.2)"},{"comment":"The constant a0 in (2.18) is a free proportionality constant, and the calibration to λ ≈ e is not derived from the model; the text should more clearly state that a0 is a normalization rather than a model output.","section":"Section 2.2.2, Eq. (2.18)"},{"comment":"The loss-aversion proxy LAbt is defined as the retail-funding share, which is admittedly a coarse balance-sheet proxy rather than a measure of loss aversion; the text acknowledges this, but readers would benefit from an explicit statement about the direction of the measurement error and how it would bias β2 in (6.13).","section":"Section 6.2, variable construction"}],"recommendation":"major_revision","confidential_remarks":"The manuscript's central theoretical object, the Bank Run Exposure State Space, is not validly derived because of the conflated loss-aversion index and the unsupported upper bound in (3.4). This is a load-bearing problem that goes beyond presentation. In addition, the paper repeatedly defers technical content to an 'Internet Appendix' that is not part of the submission, and it cites the author's own companion work in progress for the empirical implementation; this makes independent verification difficult. The empirical section is honest about its limitations, but the theoretical gaps are substantial. I would be open to reconsidering a revised version that re-derives the state-space construction with a correct loss-aversion index, provides a genuine derivation of the martingale representation, and fixes the negative-binomial misspecification."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's my read. The paper has a genuinely new behavioral channel: loss-averse, paycheck-to-paycheck depositors whose subjective probability of a bad income state exceeds a threshold can trigger a run. Lemma 2 is a valid sufficient condition, and the LCR framing is a sensible hook. But the central formal object — the Bank Run Exposure State Space — does not hold up as written. Equation (3.3) treats the ratio of marginal utilities at an arbitrary consumption change as the loss-aversion index. That conflates the parameter λ in (2.1) with the Kőbberling–Wakker derivative ratio at the kink. Under the paper's own Tversky–Kahneman calibration with symmetric power coefficients, the ratio v'_g(c1)/v'_ℓ(-c1) equals 1 for all c1, not the calibrated λ ≈ 2.25. So the bounds in (3.4) and the taboo set Ω_taboo do not follow from Theorem 2.1; the upper bound has no derivation at all. Since Ω_run is the intersection with Ω_taboo, Theorem 3.1 and the pullback-topology construction are unsupported.\n\nOther soft spots are real but milder. Proposition 2 uses a random stopping time τn(ω) as the fixed parameter of a negative binomial; that needs a conditional or mixed representation. Theorem 5.1's martingale representation is asserted with a specific integrand and no derivation, and the Internet Appendix that supposedly contains these arguments is not included in the arXiv posting. The empirical exercise is honestly labeled as proof-of-concept, and it is: within R² of 0.000 and overall R² of 0.043–0.061 are weak, and the regression-weighted composite is constructed in-sample, so the modest fit improvement is partly mechanical.\n\nWhat the paper does well: the behavioral story is coherent and the literature review is genuine. The ratchet plus loss-aversion combination is a plausible way to make liquidity demand state-dependent, and the regression specification gives supervisors a concrete monitoring statistic even if public Call Report data can't test it. The honesty about data limitations is welcome.\n\nVerdict: conditional, but the condition is substantial. The trigger condition in Lemma 2 can stand, but the state-space needs to be rebuilt on the actual λ parameter rather than the marginal-utility ratio, and the probability and martingale results need real derivations. As it stands, the claimed formal contributions in Sections 3–5 are not established.\n\nRecommendation: this deserves a serious referee, with instructions to focus on Sections 3–5 and to require the missing Internet Appendix, a fixed negative-binomial statement, and a derivation of the martingale integrand. I wouldn't cite the state-space result until it's repaired.","headline":"A plausible behavioral trigger (Lemma 2) is buried under an unsupported state-space construction that conflates the loss-aversion parameter with a marginal-utility ratio, so the paper's central formal contribution needs substantial rework before it is credible.","tokens_in":31185,"tokens_out":3286,"would_cite":false,"duration_ms":29698,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper claims that bank runs become possible exactly when depositor loss aversion exceeds a belief-weighted threshold set by the odds of a bad paycheck state, and identifies that threshold in closed form.","keywords":["bank runs","loss aversion","liquidity risk","consumption ratcheting","suspension of convertibility","paycheck-to-paycheck economy","Bank Run Exposure State Space","stochastic loss aversion"],"falsifier":"Measure a depositor population's subjective bad-state probability, loss-aversion index, and marginal value-function slopes, then check whether suspension of convertibility actually occurs in the states satisfying the Lemma 2 inequality; if reserve shortfalls arise outside those states, or fail to arise inside them, the central claim is falsified.","tokens_in":29986,"feed_emoji":"🏦","tokens_out":10773,"duration_ms":97123,"temperature":0.7,"pith_summary":"This paper tries to establish that bank run exposure is not only a balance-sheet object but an endogenous behavioral outcome: in a paycheck-to-paycheck economy, depositors who are loss averse over declines in income and consumption will suddenly demand more liquidity when they attach sufficiently high probability to a bad income state. The paper pins down a precise trigger — the loss-aversion index exceeds a probability-weighted ratio of marginal gains to marginal losses — and shows that the states satisfying this condition form a Bank Run Exposure State Space. If true, regulatory run-off factors of the LCR type systematically understate reserve needs in stress states, because they treat withdrawals as exogenous averages rather than as a function of depositor fear. The empirical proof of concept on quarterly Call Report data is deliberately modest: a composite balance-sheet proxy improves fit slightly over retail share, and the theory's real test awaits account-level pay-cycle data.","feed_headline":"One inequality predicts when depositor fear becomes a run","feed_subtitle":"The model pins down the stress states and run odds, then tests a composite exposure proxy on U.S. bank data.","key_machinery":"The load-bearing object is the loss-aversion-adjusted marginal propensity to consume, $\\tilde{k}(\\rho,d;\\lambda)=\\tilde{k}(\\rho,d)\\lambda$, coupled with the consumption ratchet: the one-period ratchet increment is nonnegative exactly when $\\tilde{k}\\ge k$ (Theorem 2.1). This linear separability converts loss aversion into a threshold on the marginal propensity to consume and into the admissible-state inequality of Lemma 2. Around that threshold the paper builds the Bank Run Exposure State Space, the stopping time at which cumulative withdrawals exhaust reserves, and the half-Cauchy stochastic process $\\bar{\\lambda}(t)=m_\\lambda\\tan(\\pi\\Phi(Z_t)/2)$ that turns the static liquidity-risk slope into a dynamic exposure coefficient.","core_discovery":"The paper's central claim is that a bank run is triggered when a loss-averse depositor's belief in a bad income state is strong enough that the state-dependent loss aversion $\\lambda(\\omega')$ exceeds $\\pi^+_{\\omega'} v'_g(c_{1\\omega'}) / (\\pi^-_{\\omega'} v'_\\ell(-c_{1\\omega'}))$ (Lemma 2). Because $\\lambda$ is itself endogenous and random, the collection of admissible and taboo states satisfying this inequality defines a Bank Run Exposure State Space, where the subjective bad-state probability $\\pi^-_{\\omega'}$ is large relative to $\\pi^+_{\\omega'}$. The model converts this state-space description into quantitative run probabilities through a stopped withdrawal process: suspension occurs when cumulative withdrawal demand first exhausts cash reserves, giving a negative-binomial count with a Poisson approximation. It then represents the exposure coefficient as a martingale plus a liquidity-risk stochastic integral, with loss aversion evolving as a positive half-Cauchy process around a median near 2.25. The proof-of-concept Call Report exercise finds that a regression-weighted composite of transaction-deposit, core-deposit, and consumer-loan shares modestly improves fit relative to retail share alone, with suggestive but proxy-sensitive amplification for small banks and the post-SVB window.","pith_inferences":["Beyond the paper: if the Lemma 2 inequality is right, supervisors could measure depositor subjective bad-state probabilities directly — for instance through survey expectations — and use the threshold as an early-warning screen for run-prone clienteles before balance-sheet deterioration appears.","Beyond the paper: a testable extension would compare two banks with identical balance sheets but different depositor paycheck volatility; the model predicts the high-volatility clientele shows larger reserve shortfalls exactly in low-income states.","Beyond the paper: the public-data proxy cannot identify the account-level mechanism, so an out-of-sample prediction is that with payroll-cycle data the interaction coefficient $\\beta_2$ should spike precisely in the states characterized by Lemma 2."],"forward_implications":["Liquidity regulation that treats run-off as an exogenous stress parameter will understate reserve needs: in the paper's simulation, omitting loss aversion cuts fitted reserve-need volatility by roughly 87 percent, making cash demand look smooth just before suspension.","Banks facing loss-averse clienteles will optimally hold higher cash reserves, and those reserves crowd out positive-net-present-value lending, a liquidity externality the model formalizes.","Run probabilities become computable objects: given the withdrawal order, reserves, and depositors' subjective bad-state probability, suspension risk follows a negative-binomial count that can be approximated by a Poisson law.","Bank-run exposure is better treated as a dynamic, clientele-specific monitoring statistic than as a fixed balance-sheet ratio, because the exposure coefficient inherits the volatility of liquidity demand and of loss aversion.","Quarterly Call Report data can implement the exposure regression, and a composite proxy spanning transaction, core, and consumer-loan shares yields modestly better fit than retail share alone, with small-bank and post-SVB amplification appearing only in some specifications."],"supporting_citations":[{"why":"Supplies the canonical bank-run model and the suspension-of-convertibility concept the paper extends to loss aversion.","marker":"Diamond and Dybvig (1983)"},{"why":"Supplies the mean-risk liquidity-preference analysis whose local indifference contour the paper re-derives with a prospect-theory value function.","marker":"Tobin (1958)"},{"why":"Supplies the piecewise value function, reference dependence, and loss-aversion index at the core of the depositor preference specification.","marker":"Kahneman and Tversky (1979)"},{"why":"Provides the myopic-loss-aversion mechanism that the paycheck-to-paycheck evaluation period relies on.","marker":"Benartzi and Thaler (1995)"},{"why":"Supplies the random-walk consumption benchmark whose growth factor is used to calibrate the median loss-aversion level.","marker":"Hall (1978)"},{"why":"Supplies the consumption-saving loss-aversion model that the paper extends by adding borrowing and discounting.","marker":"Bowman et al. (1999)"},{"why":"Defines the L'Hospital-type index of loss aversion used to build the taboo-state threshold.","marker":"Kobberling and Wakker (2005)"},{"why":"Provides parameter-free loss-aversion estimates used to anchor the calibration around $\\lambda \\approx e$.","marker":"Abdellaoui et al. (2007)"},{"why":"Documents SVB contagion and the role of uninsured deposits and unrealized losses, motivating the empirical controls.","marker":"Choi et al. (2023)"},{"why":"Supplies cumulative prospect theory weighting and the median loss-aversion value 2.25 used in the stochastic loss-aversion process.","marker":"Tversky and Kahneman (1992)"}],"fun_headline_variants":["Depositor fear threshold flips calm to bank run","When depositor fear beats hope, bank runs begin","One threshold: loss aversion sparks bank runs","Bank runs ignite when depositor loss aversion crosses a line","Loss-averse depositors: fear threshold decides run or calm"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is Assumption 8: a depositor's change in consumption is exactly proportional to her change in income, with loss aversion multiplying a fixed marginal propensity to consume; if that linear separability fails, the ratchet threshold and the run-state inequalities built on it do not follow.","fun_headline_variants_meta":{"raw":{"variants":["Depositor fear threshold flips calm to bank run","When depositor fear beats hope, bank runs begin","One threshold: loss aversion sparks bank runs","Bank runs ignite when depositor loss aversion crosses a line","Loss-averse depositors: fear threshold decides run or calm"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001493,"raw_usage":{"total_tokens":5999,"prompt_tokens":954,"completion_tokens":5045,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":570,"completion_tokens_details":{"reasoning_tokens":4967}},"tokens_in":570,"tokens_out":5045,"duration_ms":34376,"temperature":1.0,"reasoning_tokens":4967,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T04:14:52.759607+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Measure a depositor population's subjective bad-state probability, loss-aversion index, and marginal value-function slopes, then check whether suspension of convertibility actually occurs in the states satisfying the Lemma 2 inequality; if reserve shortfalls arise outside those states, or fail to arise inside them, the central claim is falsified.","supporting_citations":[{"cited_title":"2023 , month =","cited_arxiv_id":null,"evidence_quote":"Documents SVB contagion and the role of uninsured deposits and unrealized losses, motivating the empirical controls."}],"review_version":1}