Reconciliation identities as acceptance tests of an expected credit loss allowance: detection limits on a public loan book, overlays kept apart, the convex region of scenario weighting and three provisioning regimes on one synthetic bank
Year: 2026
Type: preprint
Abstract: An expected credit loss allowance is published in tables that must reconcile, is moved by overlays and differs across regimes; published work gives no detection limit for the reconciliation and no test of the overlay rule. We test a protocol: the IFRS 7 identities as an acceptance test with a measured detection limit, overlays as separate lines, the region of the non-linearity of the weighted loss, and a bridge between three regimes. Twelve identities on a tolerance of one millionth of the gross amount, run on six dates of an SBA 7(a) loan book, name a planted defect from that size in 4,200 of 4,200 cases of seven classes; the stage arithmetic and a pooled row name 1,800. An absolute tolerance of 10^-6 units flags 4 of six clean tables. No identity sees an amount booked under another line or an overlay merged into the model loss. The weighted PD is never below the PD at the mean factor in 16,906 convex draws and is below it in 796 of 2,046 mixed draws. Three scenarios at one standard deviation understate the expected PD by a median 4.56%. One synthetic bank has an IFRS 9 total of 0.621 of B-1 in the base scenario over 10 seeds (95% interval 0.606 to 0.635), above it only in the combined severe one. Of 27 predictions, 3 were refuted: the books flagged after rounding to millions (97.8%, predicted above 99%), the entries no identity sees, and the largest group of indistinguishable entries. Repository and reproducible pipeline (MIT): https://github.com/fsantibanezleal/CAOS_Contraste .
Author: Santibáñez-Leal, Felipe
Language: English
Subfield: Finance
Domain: Social Sciences
Open Access status: green
OpenAlex ID: W7225649219
Cites:
Cited by:
Related to: