Circular Leverage in Bank-NBFI Synthetic Risk Transfer Networks

Description

Abstract (English)

Audience-Targeted Summaries

For the expert (financial economist or regulator). This paper formalizes circular leverage in bank-NBFI Synthetic Risk Transfer networks through parameter λ (self-funding fraction). Using network contagion simulations across 1,000 Monte Carlo runs, it identifies a two-stage transition: first departure from baseline at λ_onset ≈ 0.85–0.95, then a sharp jump at λ ≈ 0.95 where Dragon King events emerge, cascades that cannot be diversified away because correlation is endogenous to the funding loop. The transition location is invariant across network density, concentration, and shock parameters; only cascade magnitude scales with density (0.18 to 0.61 across the tested range). The paper proposes six publicly observable proxy metrics ranked by sensitivity-weighted ordinal position relative to λ, with four showing stress signals as of Q1 2026. Critical policy implication: disclosure of λ would enable macroprudential supervision at a stable threshold.

For the practitioner (risk manager or bank executive). Your bank extends credit lines to funds that buy your SRT protection, creating a hidden feedback loop. When stress hits, calling those credit lines can force fund failures, wiping out your capital relief and triggering further contractions. The danger zone starts around λ = 0.85–0.95, where λ is self-funded protection divided by total protection. Below that zone the system absorbs shocks; above it, cascades become self-reinforcing. The paper gives you six market signals to watch (BDC prices, PIK ratios, CLO spreads, among others) that fire before traditional funding stress indicators like SOFR-OIS. As of Q1 2026, four of six are flashing red. You cannot measure your own λ without disclosure, but you can watch the cockpit.

For the general public. Banks have found a way to make loans appear less risky on paper without actually reducing the risk. They sell "insurance" on their loans to investment funds, but often lend money to those same funds to buy the insurance. It is like insuring your house against fire and then lending the down payment to the insurance company. If there is a fire, you are partly paying yourself. Researchers found that this arrangement stays stable up to a critical point around 95% self-funding, then suddenly becomes fragile. Four warning signs are already showing stress, but regulators cannot see the key number that would tell them how close we are to danger.

For the skeptic. The paper admits its central parameter λ is unobservable, making direct empirical validation impossible. The six proxy metrics rely on judgment-assigned sensitivity weights, not statistical estimation. The LPPLS framework is explicitly disclaimed as conceptual vocabulary, not empirical prediction. Network topology is stylized (random graphs, not real bank-fund relationships). No central bank intervention is modeled. The 0.95 threshold's stability across parameters is a property of the simulated random graph; real-world heterogeneity (geographic clustering, cross-border fragmentation, correlated reference portfolios) is not modeled and could shift it. The Q1 2026 "red signals" are post-hoc observations, not out-of-sample predictions. These are legitimate limitations flagged directly in §8.

For the decision-maker (policymaker or regulator). You need one number: λ, the fraction of SRT protection financed by the originating bank's own credit. It is already sitting in bank risk systems but is not reported. Disclosure would let you place institutions on a phase diagram with a stable critical zone at 0.85–0.95. The paper proposes a conservative supervisory limit of λ ≤ 0.30, far below the danger zone, with margin for error. Without disclosure, you must rely on six public market metrics (Table 2), four currently showing stress. The traditional interbank funding indicator (SOFR-OIS) ranks last; the upstream signals fire earlier. The policy ask is minimal: one Pillar 3 disclosure item, not new capital requirements or bans.

Methods (English)

AI Utilization Statement

This work was prepared with assistance from Claude Sonnet (Anthropic), Gemini 3.1 (Google), and Kimi 2.5 Thinking (Moonshot), orchestrated through Novakit Utilities v3 (a commercial cognition-as-utility toolkit). AI tools contributed to drafting, critique, code writing and debugging, figure generation, and sensitivity analyses under the author's direction. Final draft inspections used Opus 4.7. The research question, parameterization, empirical interpretations, policy recommendations, and all numerical results (executed from the simulation code) are the author's own. All errors remain the author's responsibility.