In China's domestic corporate bond market, the trustee who is supposed to act for bondholders when an issuer defaults is often the same bank that underwrote the bonds. That bank earns fees from the issuer and wants its future business. Legal scholarship argues that this dual role makes trustees slow to declare defaults and enforce claims. That argument has been made through case comparisons, not measurement. This note proposes a direct test with a sharper question. The dual role may not be uniformly harmful. An underwriter-trustee knows the issuer well, which could let it act faster than an outside trustee. That advantage may hold while the issuer is a minor client and reverse once the issuer becomes a major source of the bank's fees.
The test exploits a feature of the 2021–2022 real-estate default wave. When one of a developer's bonds defaults, cross-default clauses put all of that developer's other bonds under pressure at the same moment, and different bonds often have different trustees. Comparing how quickly each trustee responds to the same shock holds the issuer's finances fixed. Any remaining difference in response speed can then be attributed to the trustee's own incentives.
Core Claim
The claim is that the gap in response speed between dual-role trustees and independent trustees changes sign as the underwriter's dependence on the issuer grows. Dependence is measured as the issuer's share of the underwriter's recent underwriting fees or volume. At low dependence, the dual-role trustee should respond faster than an independent one, because it has inside information and little to lose. At high dependence, it should respond slower, because it is protecting a client. Under this view, the harm is concentrated above a threshold. That matters for policy: an outright ban on dual roles and a cap on client concentration imply different reforms.
What is new is neither "add a squared term" nor "conflicts are bad." The new elements are three. First, the identification comes from many trustees responding to one shared cross-default event. Second, the shape test is a difference in curvature between dual-role and independent trustees, not a curve fitted to the dual-role group alone. Third, the analysis commits to showing that the turning point lies inside the observed range of dependence, rather than treating a significant squared term as proof of a reversal.
Prior Work and Positioning
The motivating evidence comes from Difficulties and Optimization Paths of the Application of "Cross-default" Clause in Chinese Real Estate Corporate Bonds. That paper compares bond documents from developers such as CIFI and Country Garden. It shows that domestic cross-default triggers are loose, requiring both RMB 50 million and 10% of the parent company's net assets, and that Country Garden's USD bond trigger was only 12.2% of its domestic one. It argues that underwriter-trustees delay enforcement. It runs no regressions and does not ask whether the harm depends on how much the trustee relies on the issuer. This proposal turns that legal argument into a measurable relationship.
Case studies of audit failure point to the same mechanism without measuring it. Analyses of Kangmei Pharmaceutical attribute auditors' failure to commercial and social embedding with management (An Analysis of Kangmei Pharmaceutical's Financial Fraud Based on GONE Theory, The Failure of Kangmei, the Lesson of A-Shares). Two further papers describe auditor negligence and contract value overriding professional norms (Zijing Storage Financial Fraud Case Analysis, Business Ethics and Accounting Professional Ethics in Financial Fraud). Outsider Oversight states the general "independence paradox": a monitor paid by the party it monitors drifts toward rubber-stamping. All of these treat capture as present or absent. This proposal asks how capture grows with dependence.
The closest quantitative work constrains the design. Research on the Robustness of Financial Statements of Listed Companies based on Securities Law measures reporting conservatism before and after reforms in Chinese A-shares. It finds an earnings bad-news bias of 0.006 in 2008–2012 and 0.002 in 2013–2018. Because it compares two time periods, it cannot say how incentives at the level of a single relationship shape behavior. Detecting Accounting Frauds in Publicly Traded U.S. Firms shows that non-linear classifiers outperform linear ones on fraud labels. Its BalancedCascade model reaches a balanced accuracy of 77.53%, against 69.74% for a kernel SVM benchmark. That is evidence that non-linearity carries signal in this domain. However, it is a prediction result about fraud, not an estimate of a gatekeeper's incentive, so it is not a baseline for the present question.
The quadratic method comes from applied work that estimates turning points. Examples include the digitalization paradox in small and medium firms (Chen et al.), the U-shaped link between digital finance and poverty (Cheng), and the threshold relationship between financial development and foreign direct investment (Dutta). These papers also show the method's known weakness: a curve can be significantly concave while its peak lies outside the data. This proposal treats that weakness as a test it must pass rather than a footnote.
Research Program
The first step is to build a bond-level panel of all domestic real-estate corporate bonds outstanding during 2021–2022, using the issuance and default records in Wind that Can and Wang draw on. For each bond, the dataset records the lead underwriter and the trustee, and flags bonds where they are the same institution. For each underwriter, it computes dependence on each issuer as that issuer's share of the underwriter's real-estate underwriting volume over the prior three years. Fees are used instead where prospectuses disclose them.
A "trigger event" is the first public payment default or extension by any bond of an issuer. For every other outstanding bond of that issuer, the outcome is the number of days until its trustee takes a first protective step. Protective steps are publishing a credit-risk notice, convening a holders' meeting, or declaring acceleration. Bonds whose trustee never acts within the window are treated as censored. Secondary outcomes are the bond's price or spread reaction in the 30 days after the trigger, and its eventual recovery where restructuring terms are public.
The main model is a hazard regression, which models the probability of the trustee acting on each day given that it has not yet acted. Let $h_{bke}(t)$ be the daily hazard that the trustee of bond $b$ acts $t$ days after trigger event $e$ at issuer $k$. Let $D_b$ indicate a dual-role trustee, and let $X_b$ be the trustee's dependence on the issuer, centered at its sample mean:
$$\log h_{bke}(t) = \lambda_{ke}(t) + \gamma D_b + \beta_1 D_b X_b + \beta_2 D_b X_b^2 + \delta_1 X_b + \delta_2 X_b^2 + \Gamma Z_b$$
Here $\lambda_{ke}(t)$ is a baseline hazard specific to each issuer-event, so all bonds of the same issuer facing the same shock are compared only with each other. $Z_b$ holds bond controls: remaining maturity, size, whether the bond is secured, and the strictness of its cross-default clause. The $\delta$ terms allow dependence to matter for independent trustees too, for example because they have other business ties to the issuer. The parameters of interest are therefore $\beta_1$ and $\beta_2$, which capture how the effect of the dual role itself bends with dependence.
The hypothesis predicts that the dual-role effect $\gamma + \beta_1 X + \beta_2 X^2$ is positive (faster action) at low $X$ and negative at high $X$. Its zero crossing and turning point should both fall between the 5th and 95th percentiles of observed dependence.
Baselines
The naive reference point is a model with only the dual-role indicator and the issuer-event effects. This is the quantitative version of Can and Wang's claim that dual roles delay enforcement on average. If it shows a large uniform delay and adding dependence explains nothing, the threshold story fails. The proposal must beat it with a likelihood-ratio test and with an improvement in held-out concordance, computed by cross-validation that holds out entire issuers.
The strongest flexible competitor is the same hazard model with dependence entering through a restricted cubic spline, which can bend in any smooth way. It is also estimated separately with binned dependence dummies at terciles and quintiles. Both use identical samples, fixed effects, and controls. Knot placement is tuned by the same issuer-held-out cross-validation used for any choice in the quadratic model, so neither side gets extra tuning. "Beating" the spline is not the goal. The quadratic summary is only accepted if the spline's estimated dual-role effect has the same sign change in the same region, and if the quadratic's held-out concordance falls within the spline's confidence interval. If the spline shows an asymmetric or kinked shape that the parabola misses, the spline result is reported as the finding.
A gradient-boosted survival model, given the same features plus dual-role status, serves as an upper bound on predictive fit. It can find interactions the parametric model ignores, and its partial-dependence curve for dependence acts as a check on the shape. No reported result was found on Lacuna for any model of trustee enforcement timing, delay, or recovery as a function of trustee incentives. There is therefore no published number to beat, and all baselines will be estimated fresh on the same panel.
The ablation that isolates the key ingredient removes the issuer-event effects and replaces them with issuer-level credit controls. If the curvature appears only without within-event comparison, it probably reflects which issuers choose dual-role trustees rather than how trustees behave.
Why not just…?
Why not just regress delay on a dual-role dummy? Because a sign-changing effect can average to roughly zero, and a null average would be read as "no conflict." That is the same averaging problem that hides the digitalization paradox in Chen et al.. The dummy-only model is kept as the naive baseline. If its estimate is large and the dependence terms add nothing, this proposal's hypothesis is wrong, and the experiment will say so.
Why not just use bins or splines instead of a quadratic? This is a real threat, and the design concedes part of it. The parabola forces the decline after the peak to mirror the rise before it. Capture plausibly sets in abruptly, which would produce a kink rather than a smooth curve. For that reason the spline and the bins are run side by side, and the proposal's claim is stated in terms of a sign change and a turning point inside the data, not a parabola. The quadratic's job is to give regulators a single estimated dependence threshold with a confidence interval, computed by the delta method from the two coefficients. That role is only justified if the spline agrees.
Why not just fit a boosted model, which needs no functional form? The boosted model will probably predict as well or better. Prediction, however, is not the question. The question is whether the dual-role effect reverses, and where. A tree ensemble's partial-dependence curve has no standard error and can produce artifacts. Silent failures in automated evaluation pipelines (Auditing the Audit) argue for cross-checking it against an interpretable model rather than trusting it alone. The boosted model is included as a fit ceiling and a shape check, not as the estimator of record.
Decisive Checks
Five checks decide the result.
The first is the shape test. The slope of the dual-role effect must be significantly positive at the 5th percentile of dependence and significantly negative at the 95th, tested jointly. A significant $\beta_2$ alone is not enough.
The second is a placebo. The same curvature must be absent, or much weaker, in the $\delta$ terms for independent trustees. If independent trustees show an identical bend, the pattern reflects something about issuers or the bank's broader relationship, not the dual role.
The third addresses selection. Underwriter-trustees may be steered toward bonds expected to fail later. The analysis restricts to bonds issued before 2020, when the crisis was not anticipated. It also uses a trustee's dependence on other issuers, which shifts its incentives without reflecting this issuer's credit quality.
The fourth checks whether trustee speed matters to investors. If faster action is worth anything, the price reaction after the trigger and the eventual recovery should follow the same sign pattern in dependence.
The fifth is an honest power analysis. Wind records 48 defaulting issuers in 2022, so the number of distinct events is modest even if the number of bonds is large. Standard errors will be clustered by issuer-event. The analysis will also report the smallest change in the effect's slope it can detect. If that minimum is too large to be informative, the study will be extended to 2023 defaults before any conclusion is drawn.
References
- Difficulties and Optimization Paths of the Application of "Cross-default" Clause in Chinese Real Estate Corporate Bonds
- An Analysis of Kangmei Pharmaceutical's Financial Fraud Based on GONE Theory and Its Implications for Capital Market Governance
- The Failure of Kangmei, the Lesson of A-Shares
- Zijing Storage Financial Fraud Case Analysis - Audit Risk Prevention for Listed Companies
- Business Ethics and Accounting Professional Ethics in Financial Fraud
- Outsider Oversight: Designing a Third Party Audit Ecosystem for AI Governance
- Research on the Robustness of Financial Statements of Listed Companies based on Securities Law
- Detecting Accounting Frauds in Publicly Traded U.S. Firms: A Machine Learning Approach
- The Non-Linear Impact of Digitalization on the Performance of SMEs: A Hypothesis Test Based on the Digitalization Paradox
- Research on the Impact of The Development of Digital Financial Inclusion on Multidimensional Poverty
- Foreign Direct Investment, Financial Development and Political Risks
- Auditing the Audit: Five Failure Modes in Benchmark-Validity Audits