The 2.095 billion yuan valuation Cao, Guo and Su report for Yingfeng 2024-1 tells a trader nothing: their own numbers cannot reproduce it. Two arithmetic defects sink the calculation. The spread's sign is reversed, and the amounts discounted as cash flows are outstanding balances. The paper's entire evidence on certification pricing comes down to one 8 bp regression coefficient.
Building the valuation
The premise is that green asset-backed securities should trade tighter than conventional securities. Third-party certification should reduce the compensation investors require. Standard ABS models that ignore this effect should produce theoretical values that diverge from issue prices. Such a gap would be a place to look for mispricing.
The authors begin with Chinese five-year Treasury yields from 2019 to 2024, fitting a discretized CIR model. Their estimates are α=0.5006, μ=0.0206 and σ=0.0309. From a starting yield of 1.84%, they simulate 1,000 forward-rate paths and turn them into spot rates.
They estimate the spread separately, through an OLS regression covering 475 green ABS issued 2018 to 2023. The dependent variable is the issue rate less the one-year Treasury yield. Regressors include log issue size, tenor and rating, alongside a state-owned dummy, a regulator dummy and a green-certification dummy. Adjusted R² comes to 0.5432; the certification coefficient is -0.080 (p=0.001). The authors add the predicted spread to the simulated spot rates, use those rates to discount prospectus cash flows, and average the results across paths. They call this last calculation their option-adjusted spread (OAS) formula.
For the test, they choose Dongfeng Motor Finance's deal backed by 36,000 personal new energy vehicle loans. The deal raised 2.09 billion yuan. Its priority securities priced at 1.75%, with the book covered 2.33 times. The model produces 2.095 billion yuan, compared with the 2.09 billion yuan raised, which we take as the issuance price used in the paper's comparison.
The authors describe their valuation as "marginally higher than the actual issuance price." They attribute the difference primarily to the limitations of mainstream pricing models and their heavy dependence on historical data. They also cite the difficulty of measuring green attributes without unified disclosure standards in China.
That difference amounts to 5 million yuan, about 0.24% of the raise. To support such a small claim, the inputs need to be right within a few basis points. With a sign error in the spread and balances substituted for cash flows, the gap cannot be attributed to model limitations or anything else.
How does Yingfeng 2024-1 get a +0.85% spread?
Start with the paper's substitution line. It uses ln(2,090,000,000), approximately 21.46. Multiplying by -0.078 gives -1.674, almost cancelling the constant of 1.68. Add 0.035×5, followed by -0.152×4, -0.058, -0.285 and -0.080, and the result is -0.85%. For Yingfeng 2024, the paper reports 0.85%. The magnitude agrees. The sign is reversed.
Even the regulator coefficient changes between displays. Table 2 reports -0.285; the prediction equation uses -0.258; the substitution line returns to -0.285. Substituting -0.258 gives approximately -0.82%, which remains negative. The coding raises a separate question: the deal is described as an interbank bond market issue, yet Regulator=1 assigns it CSRC supervision.
The negative sign makes more sense than the printed positive spread. At 1.75%, the priority coupon is below the 1.84% Treasury yield used to start the simulation. Against that 1.84% yield, the market therefore paid a negative spread. The benchmarks remain inconsistent, however. The regression measures spreads over the one-year Treasury, then the valuation adds them to a curve constructed from five-year yields.
Outstanding balances in the "cash flow" column
The first row opens with 2,237,241,950.15 and deducts 58,932,759.83 of principal. It labels the resulting 2,178,309,190.32 as "Sum of cash flows." That amount is the remaining balance, down to the cent. Principal plus interest, the cash actually received by the trust during the period, is about 59.77 million. Another row needs checking before discounting begins: interest for 2025-6 is 71,455,585.13, compared with 7,396,203.41 in the preceding month.
The valuation example discounts the 2.178 billion balance as though the trust receives it as a payment. Payment periods also become simulation paths in the example. PV(path2) discounts the period-2 balance over two periods. For PV(path1000), the calculation raises 1.10505 to the 1,000th power against a flow of 3,626,920.64. Its rate of 0.10505 is absent from the spot-rate table. The paper never supplies the calculation that averages these terms to 2.095 billion.
Using actual cash flows exposes the problem. Total pool principal is 2,237,241,950.15 yuan, with interest added, and payments run for 55 months from 2024-12 to 2029-6. A discount rate of roughly 36% cannot value that stream at 2.095 billion.
We did not find a prepayment or default model anywhere in the paper. Its OAS section nevertheless says the method captures how interest-rate and prepayment risk affect value. Without assumptions for prepayment and default, OAS amounts to plain discounted cash flow. The tranche comparison is mismatched as well. The quoted market rate of 1.75% belongs to the priority tranche, whereas the model values the entire pool against the full 2.09 billion raised.
Can the simulated rates have these summaries?
Table 5 puts the 2025 mean spot rate at 0.34896 and its minimum at 0.77992. For 2028, it gives a mean of 0.69044 against a maximum of 0.46079. The defect also appears in Table 4: the 2029 mean forward rate is 0.71445, above the reported maximum of 0.70062.
No set of simulated paths can yield those summaries.
The rate levels present a separate failure. A CIR process starting at 1.84% and reverting to 2.06%, with σ of 3.09%, does not produce mean forward rates ranging from 9.9% to 71%. The paper writes its period-1 discount rate as "0.34896 + 0.85% = 0.35746," giving an AAA auto pool a yield of 35.7%.
One 8 bp coefficient carries the certification case
The certification coefficient is 8 basis points. The authors themselves say that "third-party certification offers only limited advantages in financing costs." Their evidence is a single cross-sectional regression, where the dummy may capture issuer or deal characteristics missed by the other five regressors. We did not find an out-of-sample regression test or a valuation of the same deal against a non-green baseline. Validation consists of comparing one deal with its own raise.
The authors acknowledge that the pricing methods depend heavily on historical data and omit investor behaviour and carbon policy. They present those limitations as the explanation for the discrepancy. But the calculation has yet to establish a discrepancy. Until the inputs are corrected, a 5 million yuan gap produced by a sign-flipped spread and discounted balances is noise. The same arithmetic also supports the conclusion's claim that this case demonstrates the model's effectiveness.
We could not rerun the study. We have neither Chinese ABS prices nor a Chinese Treasury curve. Most importantly, we lack an issuance panel comparable to the authors' 475 products, leaving us unable to re-estimate the spread regression.
A corrected run could earn the green-mispricing story a second look. It would need to retain a premium, with removal of the 0.080 certification term moving value by more than 5 million yuan.