The paper offers a genuine modelling advance, but its evidence stops where a trader would most want it to begin. Every RMSE is measured on the price series used for calibration. Between four and ten parameters are selected from that series, including the conversion trigger barrier. The introduction says pricing errors decline by roughly 30% relative to existing approaches. The conclusion raises the claim to up to 35%. Both rely on in-sample fits.

The model they built

A CoCo pays bond coupons until the issuer's CET1 ratio crosses a contractual threshold, triggering equity conversion or a write-down. March 2023 exposed the limits of treating that contract as the sole trigger. Swiss regulators erased roughly 16 billion Swiss francs ($17.3 billion) of Credit Suisse AT1 while its reported capital ratio remained above minimums. The Bloomberg USD AT1 Contingent Capital Index dropped 16%, its largest single-day fall. CS equity lost about 50% on 20 March. Shareholders received UBS paper through a roughly $3.2 billion deal, while bondholders received nothing. A mechanical-trigger model mispriced that outcome by construction. The global market at stake is around $275 billion as of 2024, using the Nuveen figure cited by the paper.

Chen, Wang and Xia build two linked processes. They model the CET1 ratio as a bounded decreasing (inverse-tangent) transform of a compensated compound Poisson process with Erlang jumps. Capital moves downward through jumps only, without diffusion or mean reversion. This explicitly departs from the Ornstein-Uhlenbeck specification of Chung and Kwok (2016).

The stock follows a Merton/Kou jump-diffusion and shares the CET1 jump component through a leverage parameter eta. A separate intervention shock is then added. Its intensity combines a squared Brownian motion with drift and an OU component driven by negative stock jumps. Closed-form densities using generalised confluent hypergeometric U, together with a Takács running-supremum result, produce the accounting trigger in semi-closed form. Conversion follows the one-parameter power scheme, chi = KR/(S_0^p S_tau^(1-p)). The authors show that it approximates the 30-day-average-with-floor conversion prices found in issued contracts.

Implementation has three phases. Term-sheet and rate inputs enter first. MLE then estimates the three solvency-shock parameters from quarterly CET1 and the stock parameters from daily CRSP returns. Finally, observed CoCo prices and credit spreads calibrate the intervention parameters, barrier J, write-down fraction w, conversion power p, independence probability varpi and shock scale gamma.

There are five case studies: Lloyds (2021-2023 and 2009-2011), Credit Suisse (2020-2023 and 2011), and China Construction Bank (2019-2023). On $100 notional clean prices, reported RMSE is 5.04%, 7.95%, 6.47%, 3.28% and 1.51%.

The default leg carries the model

Write-down and equity-convertible CoCos have identical non-default legs. The third term contains the substantive difference, with flexibility coming from the power parameter p. At p = 0, the default leg becomes a recovered cash payment. At p = 1, it is fully exposed to the stock at conversion.

Calibrated values are 0.9797 and 0.9999 for Lloyds, 1.0000 for CS 2011 and CCB, and 0.6235 for CS 2020-2023. Write-down fractions are essentially zero in four cases. The 2009-2011 Lloyds note has a value of 0.4582. Regulatory write-off also appears in the equity-convertible formula because recovery is zero when intervention arrives before the accounting trigger. All five observed instruments convert into equity, leaving the write-down formula in Theorem 1 untested against market prices.

For CS 2020-2023, intervention intensity is almost entirely jump-driven. The diffusion coefficient is near zero, jump sensitivity is near one and mean reversion is near zero. Once intervention risk appears, it therefore remains elevated. The independence probability lands exactly at 1.0000, the boundary of its permitted range. The paper interprets this as evidence that the market sees a relatively weak connection between the accounting trigger and regulatory action. An intervention shock scale of 0.0212 implies almost no immediate equity response: bonds move first, then the stock.

How much empirical support does intervention receive?

The honest count is one case out of five. The paper's own table records regulatory intervention modelling as no, no, yes, no, no. Support comes from one issuer over one period, with varpi fixed at the boundary value 1.0000.

The same table makes the interpretation of eta difficult to sustain. For CS in 2011, the authors estimate 0.0839, roughly a quarter of Lloyds' 0.2196, and describe this as a weakness in CS's risk structure "that persisted over time." Yet their CS estimate for 2020-2023 is 0.7412, the largest stock-CET1 leverage in the table. Those estimates also come from different procedures. CET1 data were unavailable for 2009-2011, so the relevant parameters were inferred from daily returns alone. Across these five fits, "persisted" carries more weight than the evidence can bear.

Ten parameters, one price series

The paper recommends calibrating trigger barrier J to CoCo prices instead of deriving it from the term sheet and Basel threshold. It describes the term-sheet approach as "overly susceptible" to one CET1 observation. Price calibration is defensible, though it gives the barrier room to absorb model error. Across five cases, calibrated values range from 0.0101 to 1.8732. The paper advises keeping a single calibration to no more than ten parameters, without reporting the count for each case. Between four and ten parameters therefore come from the same price series used to calculate RMSE.

Eight parameters, to the paper's credit, never use a CoCo quote. In the three cases with CET1 observations, likelihood estimation on quarterly capital ratios supplies three solvency-shock parameters. Daily returns supply five stock parameters. The case for data adaptation has some force. The abstract, however, promises short-term predictions, while the empirical section evaluates only fitted periods.

Nothing in Section 6 is evaluated out of sample.

The benchmark figures, 4.98% for CS 2011 and 11.32% best for Lloyds 2009-2011, come directly from Wilkens and Bethke's published table. They were not reproduced on identical data under a matched parameter count. A comparison at that level cannot be verified.

The paper says CS data are "uniformly truncated as of March 17, 2023, which marks the onset of a regulatory-triggered default," and the modelling rationale is coherent. Its stock process is defined only until the intervention time. The model intended to explain the AT1 write-off consequently receives no evaluation across the write-off itself, the 16% index day or the 50% equity collapse. The conclusion acknowledges that forecasting regulatory action during unprecedented systemic events remains inherently difficult. Agreed. Macro-prudential indicators and network effects across institutions are left for future work. The abstract still advertises short-term predictions without a forward test in Section 6.

The hedging results deserve the same caution. Average error ranges from 0.00% to 0.03%, with total error from 0.17% to 0.78%. The hedge uses stock delta and gamma only, rebalanced daily, and no transaction costs are mentioned. Its model prices also depend on parameters fitted to the realised price path. I would not take a total hedging error of 0.53% for a Credit Suisse AT1 from 2020 to March 2023 into a risk meeting. The authors acknowledge that jumps make perfect hedges infeasible and describe this hedge as partial.

Two implementation concerns

The default probability is approximated by a multiquadric RBF surface fitted over 10,000 uniform draws in five dimensions. Evaluation time falls from 0.6 second to under 0.1 millisecond, a 6,000-fold speedup. Ten thousand points across five dimensions amount to about six per axis. We did not find an interpolation error bound away from those sample points.

A separate issue concerns the single-step estimator used for the 2009-2011 cases. The paper states that it is valid provided lambda_1 >= 4 log 100, about 18.4207. Table 2 gives 8.1326 for Lloyds 2009-2011 and 16.7256 for CS 2011, precisely the two cases using that route. Both values sit below the paper's own stated threshold.

We could not test any of this ourselves. CoCo pricing requires security-level clean prices and contract terms for the particular AT1 instrument, together with a point-in-time CET1 and risk-weighted-asset panel for the issuer. We have neither. The three issuers examined are UK, Swiss and Chinese banks outside our tradable universe.

A persuasive test would refit the CS parameters using data through mid-2022, freeze them and price the CoCo forward into March 2023. Because the paper's sample ends on March 17, 2023, it contains no such test. If the model's default leg widens before the write-off, the early-warning claim begins to carry weight.