{"paper_id":"2509.19151","title":"Sharp Large Deviations and Gibbs Conditioning for Threshold Models in Portfolio Credit Risk","authors":["Fengnan Deng","Anand N. Vidyashankar","Jeffrey F. Collamore"],"paper_date":"2025-09-23","paper_type":"theoretical","paper_url":"https://arxiv.org/pdf/2509.19151v2","page_url":"https://thequant.space/flowcharts/sharp-large-deviations-and-gibbs-conditioning-for-threshold/","scores":{"empirical_rigor":3.2,"math_complexity":9.5,"quadrant":"Lab Rats"},"proposes":{"status":"available","statement":"The paper proposes sharp large deviation estimates for exceedance probabilities in dependent triangular array threshold models with many latent factors, and establishes a Gibbs conditioning principle in total variation for large-loss events.","quote":"We obtain sharp large deviation estimates for exceedance probabilities in dependent triangular array threshold models with a diverging number of latent factors. Using these estimates, we establish a Gibbs conditioning principle in total variation: conditioned on a large exceedance event, the default indicators become asymptotically i.i.d., and the loss-given-default distribution is exponentially tilted (with the","summary":"We obtain sharp large deviation estimates for exceedance probabilities in dependent triangular array threshold models with a diverging number of latent factors. The prefactors quantify how latent-factor dependence and tail geometry enter at leading order, yielding three regimes: Gaussian or…","strategy":null},"assumptions":{"status":"available","items":[{"assumption":"Normalization of factor loadings: sum of squared loadings plus squared idiosyncratic weight equals one for all n.","type":"model","stated":"explicit","quote":"We assume throughout the article that the nonnegative constants {a(n) j : j = 1, . . . , kn} and b(n) have been normalized such that kn X j=1 (a(n) j )2 + (b(n))2 = 1, for all n; (A1) and we further assume that b := lim n→∞b(n) ∈(0, 1).","support":"full"},{"assumption":"Definition of aggregate common factor Zn as the weighted sum of latent factors.","type":"model","stated":"explicit","quote":"(A2) For ease of notation, denote the aggregate common factors by Zn := a(n) 1 Z(n) 1 + · · · + a(n) kn Z(n) kn , n ∈Z+.","support":"full"},{"assumption":"Limit of idiosyncratic weight b(n) exists and lies in (0,1).","type":"model","stated":"explicit","quote":"We assume throughout the article that the nonnegative constants {a(n) j : j = 1, . . . , kn} and b(n) have been normalized such that kn X j=1 (a(n) j )2 + (b(n))2 = 1, for all n; (A1) and we further assume that b := lim n→∞b(n) ∈(0, 1).","support":"full"},{"assumption":"Loss sizes U(n)i are i.i.d. and independent of default indicators and latent factors.","type":"model","stated":"explicit","quote":"Next, let U (n) i denote the loss incurred from the ith loan (the “loss given default”, or LGD), and assume that {U (n) i : i = 1, . . . , n} is an i.i.d. sequence (possibly dependent on n). We assume that the loss size is independent of the event of default; thus, {U (n) i : i = 1, . . . , n} may be taken to be independent of {Z(n) j : j = 1, . . . , kn} and of {ϵ(n) i : i = 1, . . . , n} (see Appendix F, Remark","support":"full"},{"assumption":"Log‑smooth tail conditions for the idiosyncratic factor ε (Gumbel‑type) and the common factor Z.","type":"statistical","stated":"explicit","quote":"(Log-smooth tail condition for ϵ (Gumbel-type)). (Log-smooth left-tail condition for Zn).","support":"full"},{"assumption":"Balance and window interiority conditions (W1‑W3) ensuring appropriate scaling of the tail region.","type":"statistical","stated":"explicit","quote":"(Balance condition). There exist a sequence Mn ∈(an, bn), where an, bn →∞, such that R1n := n Cx b(n) fϵ( v+Mn b(n) ) rZn(Mn) →1 and R2n := nCxfϵ( v+Mn b(n) )Q′ ϵ( v+Mn b(n) ) (b(n)rZn(Mn))2 = Θ(1). (W1).","support":"full"}]},"datasets":{"status":"available","items":[{"name":"Synthetic Gaussian ε and Gaussian Z","provider":null,"kind":"synthetic","access":"unknown","period":null,"universe":null,"quote":"Case 1: Gaussian ϵ and Gaussian Z. Table E.1 reports VaR and ES estimates for the setting ϵ ∼N(0, 1) and Z ∼N(0, 1). n = 10 n = 50 n = 100 n = 500 n = 1000 VaR α = 0.95 0.565 0.522 0.514 0.505 0.503 α = 0.99 0.630 0.550 0.534 0.513 0.509 α = 0.999 0.711 0.589 0.561 0.526 0.518 ES α = 0.95 0.692 0.597 0.572 0.537 0.528 α = 0.99 0.692 0.583 0.558 0.526 0.518 α = 0.999 0.745 0.607 0.574 0.532 0.522 Table E.1: VaRα and","support":"full"},{"name":"Synthetic Gaussian ε and Pareto Z","provider":null,"kind":"synthetic","access":"unknown","period":null,"universe":null,"quote":"Gaussian Z. Case 2: Gaussian ϵ and Pareto Z. Table E.2 presents the results when ϵ ∼N(0, 1) and Z ∼ Pareto(xm = 1, α = 2). Compared with the Gaussian case, the heavy-tailed distribution of Z yields systematically larger values of both VaRα and ESα, reflecting the heavier tail risk. n = 10 n = 50 n = 100 n = 500 n = 1000 VaR α = 0.95 0.664 0.560 0.539 0.514 0.510 α = 0.99 0.718 0.587 0.559 0.524 0.516 α = 0.999 0.777","support":"full"}],"known":[],"sample":{"start":null,"end":null,"market":null,"asset_class":null,"frequency":null},"data_resolution":"none"},"transaction_costs":{"status":"not_applicable","modeled":"n/a","headline_net_of_costs":null,"detail":null,"quote":null,"source":"front_matter","rule":"R3","candidates":{"front_matter":"n/a","evidence_card":"n/a","evidence_card_quoted":false},"meaning":"the paper reports no trading or execution result"},"results":{"status":"available","evaluation_type":"none","findings":[{"finding":"Sharp large deviation prefactors exhibit three regimes: polylogarithmic refinements for Gaussian or exponential‑power tails, polynomial scaling for regularly varying tails, and an n^{-3/2} prefactor for bounded‑support tails.","kind":"theoretical","quote":"The prefactors quantify how latent-factor dependence and tail geometry enter at leading order, yielding three regimes: Gaussian or exponential-power tails produce polylogarithmic refinements of the Bahadur-Rao $n^{-1/2}$ law; regularly varying tails yield index-driven polynomial scaling; and bounded-support (endpoint) cases lead to an $n^{-3/2}$ prefactor.","support":"full"},{"finding":"Conditioned on a large loss event, default indicators become asymptotically i.i.d. and the loss‑given‑default distribution is exponentially tilted.","kind":"theoretical","quote":"Using these estimates, we establish a Gibbs conditioning principle in total variation: conditioned on a large exceedance event, the default indicators become asymptotically i.i.d., and the loss-given-default distribution is exponentially tilted (with the boundary case handled by an endpoint analysis).","support":"full"},{"finding":"Second‑order expansions for Value‑at‑Risk and Expected Shortfall are derived, showing when these risk measures lie in the genuine large‑deviation regime.","kind":"theoretical","quote":"As illustrations, we obtain second-order approximations for Value-at-Risk and Expected Shortfall, clarifying when portfolios operate in the genuine large-deviation regime. Setting xα,n = µU + yn and applying a Taylor expansion up to third order around µU, we obtain Λ∗ U(xα,n) = Λ∗ U(µU + yn) = Λ∗ U(µU) + (Λ∗ U)′(µU)yn + 1 2(Λ∗ U)′′(µU)y2 n + O(y3 n), (E.1) 47 Noticing that Λ∗ U(µU) = 0, (Λ∗ U)′(µU) = 0, and (Λ∗","support":"full"},{"finding":"Numerical experiments confirm that heavy‑tailed common factors (Pareto Z) produce systematically larger VaR and ES values compared to Gaussian factors.","kind":"performance","quote":"Compared with the Gaussian case, the heavy-tailed distribution of Z yields systematically larger values of both VaRα and ESα, reflecting the heavier tail risk. n = 10 n = 50 n = 100 n = 500 n = 1000 VaR α = 0.95 0.664 0.560 0.539 0.514 0.510 α = 0.99 0.718 0.587 0.559 0.524 0.516 α = 0.999 0.777 0.619 0.582 0.535 0.524 ES α = 0.95 0.712 0.587 0.561 0.526 0.518 α = 0.99 0.752 0.606 0.573 0.531 0.521 α = 0.999 0.801","support":"full"},{"finding":"The Gibbs conditioning principle holds in total variation, identifying the tilted limit law and demonstrating disappearance of dependence under conditioning.","kind":"theoretical","quote":"We also established a Gibbs conditioning principle in total variation for the conditional law under rare-loss events, identifying the tilted limit law and clarifying when dependence disappears at the conditioning scale.","support":"full"}],"reported_performance":null,"out_of_sample":{"method":"none","headline_out_of_sample":null}},"independent_evidence":{"status":"none_found","supports":[],"contradicts":[],"qualifies":[],"compared_with":22,"citations":{"cited_by_in_corpus":0,"by_independent_authors":0,"how_cited":{"used_as_baseline":0,"uses_method_or_data":0,"extends":0,"discusses_findings":0,"mentions":0,"unclassified":0},"independent_papers":[],"scope":"Citations between papers in this corpus only; citations from outside it are not counted."}},"implementation":{"status":"available","public":false,"code_url":null,"note":"No repository presented by the authors as their own was found in the paper. Absence here is not proof that no code exists."},"replication":{"status":"none_found","card":null,"note":"No replication of this paper has been run by thequant.space yet."},"extraction":{"text_source":"fulltext","model":"openai/gpt-oss-120b","version":2,"extracted_at":"2026-11-14T19:47:51Z"},"provenance":{"claims":"Everything here is what the authors state in the paper. Nothing was re-run or reproduced by the platform.","quotes":"Assumptions, datasets, cost detail and findings were extracted by a language model. Every quote is verbatim paper text. Where an item has `support`, \"full\" means every figure in it appears in the quoted sentences and \"partial\" means a figure is stated elsewhere in the paper; items with a figure found nowhere in the paper were discarded.","datasets_known":"datasets.known lists well-known vendors and public datasets by a stable id, found by a dictionary over the full text where the sentence says the data is used; it is independent of the model and covers only names in the dictionary.","flags":"transaction_costs.modeled, results.evaluation_type and results.out_of_sample.method are platform classifications (structured-decision models); see https://thequant.space/score-guide/#penalty-flags","comparisons":"supports / contradicts / qualifies are a language model's reading of two papers' quoted findings, restricted to closely related papers that share no author. They are pointers to read, not verified replications, and coverage is partial.","implementation":"code_url is a repository the paper itself presents as its own (rule-scored, ambiguous cases judged by a model), checked for liveness when it was recorded."},"docs":"https://thequant.space/agents/"}