arxiv: 2605.12189 · v1 · submitted 2026-05-12 · 💱 q-fin.PR · q-fin.CP

Recognition: 2 theorem links

· Lean Theorem

A deep learning approach for pricing convertible bonds with path-dependent reset and call provisions

Nicolas Langren\'e, Qinwen Zhu, Wen Chen

Pith reviewed 2026-05-13 03:20 UTC · model grok-4.3

classification 💱 q-fin.PR q-fin.CP

keywords convertible bondsdeep learningpath-dependent PDEreset provisionscall provisionsdynamic programmingneural networksquantitative finance

0 comments

The pith

Deep learning prices path-dependent convertible bonds by solving high-dimensional PDEs with neural networks.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

This paper develops a method to value convertible bonds that include a downward reset of the conversion price and an issuer call triggered by rolling windows on the stock price history. The bond value is expressed as a path-dependent partial differential equation that tracks dependence on the entire past path of the underlying and the current conversion price. A backward dynamic programming scheme approximates the conditional expectations at each step with neural networks, keeping the calculation feasible under three standard models of asset dynamics. When tested on the China CITIC Bank Convertible Bond, the prices and risk sensitivities remain stable across the different models. The results show that the contract terms themselves drive the bond value more than the choice of how the stock price evolves.

Core claim

The valuation problem is cast as a path-dependent PDE for GBM, CEV, and Heston dynamics. A discrete-time dynamic programming scheme is constructed in which neural networks approximate the conditional expectations required to step the value backward from maturity. Application to the China CITIC Bank Convertible Bond produces stable prices and sensitivities across all three models, yielding the conclusions that contractual features dominate underlying dynamics, the call provision lowers price by truncating upside, and the downward reset lowers price by reducing the effective call threshold.

What carries the argument

Path-dependent partial differential equation (PPDE) discretized via dynamic programming in which neural networks approximate the conditional expectations that depend on historical paths and the evolving conversion price.

If this is right

Contractual features such as the call and reset provisions dominate the choice of underlying asset dynamics in determining convertible bond values.
The issuer call provision decreases bond prices by truncating the holder's upside potential.
The downward reset provision decreases bond prices by lowering the effective call threshold and increasing the likelihood of early redemption.
The framework yields stable and accurate prices together with consistent sensitivity patterns across GBM, CEV, and Heston specifications.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

The same neural-network dynamic programming structure could be applied to other path-dependent contracts such as barrier or lookback options to test computational scaling.
The finding that contract terms outweigh model choice implies issuers may benefit from explicit valuation of reset-call interactions before issuing new bonds.
Adding further state variables such as stochastic interest rates while retaining the neural network approximation provides a direct extension that can be checked numerically.

Load-bearing premise

The neural-network approximation of conditional expectations remains accurate and stable when the state space includes the full path history of the underlying and the dynamic conversion price.

What would settle it

Running the same pricing exercise on the China CITIC Bank Convertible Bond with a standard Monte Carlo simulation under identical path-dependent reset and call rules and comparing the resulting prices and sensitivities.

Figures

Figures reproduced from arXiv: 2605.12189 by Nicolas Langren\'e, Qinwen Zhu, Wen Chen.

**Figure 2.** Figure 2: PPDE-based convertible bond pricing surfaces under different underlying dynamics. [PITH_FULL_IMAGE:figures/full_fig_p014_2.png] view at source ↗

**Figure 3.** Figure 3: State-space evolution of call and reset counters with conversion decisions. [PITH_FULL_IMAGE:figures/full_fig_p017_3.png] view at source ↗

read the original abstract

This paper develops a deep learning-based framework for pricing convertible bonds with path-dependent contractual features, namely downward conversion price reset and issuer call clauses under rolling-window trigger rules, which are widespread in the convertible bond market. We formulate the valuation problem as a path-dependent partial differential equation (PPDE), which explicitly captures the dependence of the convertible bond value on the historical path of the underlying asset and the dynamic evolution of the conversion price. We derive consistent PPDE formulations for three canonical underlying dynamics: geometric Brownian motion (GBM), constant elasticity of variance (CEV) and Heston stochastic volatility. We then construct a discrete-time dynamic programming scheme in which conditional expectations are approximated by neural networks, which remains tractable in such high-dimensional path-dependent setting. Empirical tests on China CITIC Bank Convertible Bond show that our framework produces stable and accurate prices and sensitivity patterns across all model specifications. Three key economic insights emerge: 1. Contractual features dominate underlying dynamics in determining convertible bond values. 2. The call provision decreases convertible bonds prices by truncating upside gains. 3. Counterintuitively, despite improving conversion terms, the downward reset provision further decreases the price of convertible bonds by lowering the effective call threshold and making early redemption more likely. The proposed PPDE-deep learning approach provides an efficient, flexible tool for pricing convertible bonds with complex path-dependent structures.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

This paper sets up a workable NN solver for path-dependent convertible bond pricing but provides no benchmarks or error bounds to support its accuracy claims.

read the letter

This paper gives a practical neural-network method for pricing convertibles with path-dependent resets and calls, but the supporting numerical checks are missing. They formulate the problem as a path-dependent PDE that includes the rolling conversion price in the state, then discretize it and replace the conditional expectations with neural nets. This works for GBM, CEV, and Heston dynamics and they test it on one real bond from China CITIC Bank. The new part is combining the augmented state for the conversion price with the NN solver for the high-dimensional case. It handles the path dependence without exploding the dimension, which is a real advantage over standard PDE or Monte Carlo approaches for these features. The results show that the contractual terms drive the price more than the underlying model, and that both the call and the downward reset lower the value. Those insights line up with intuition once you think about how the call truncates upside and the reset makes early call more likely. The gap is that we get no error metrics, no comparison to a benchmark solver even on the GBM case without features, and no analysis of the approximation error from the neural nets. The claim of stable and accurate prices rests on the output looking reasonable across models, but without independent verification that could be misleading. This is aimed at practitioners pricing complex convertibles or researchers extending numerical methods to path-dependent derivatives. It deserves a serious referee because the formulation and solver are novel enough to warrant checking the details, even if the current evidence is light.

Referee Report

3 major / 1 minor

Summary. The paper develops a deep learning framework for pricing convertible bonds with path-dependent downward conversion price resets and issuer call provisions under rolling-window triggers. It formulates the valuation problem as a path-dependent PDE (PPDE) under GBM, CEV, and Heston dynamics, constructs a discrete-time dynamic programming scheme that approximates conditional expectations via neural networks, and reports empirical results on the China CITIC Bank Convertible Bond showing stable prices and sensitivities, along with three economic insights on the dominance of contractual features over underlying dynamics.

Significance. If the neural-network approximations prove reliable, the method would provide a practical tool for valuing complex path-dependent convertible bonds that are common in markets but intractable with standard PDE or Monte Carlo approaches. The reported dominance of contractual provisions and the counterintuitive price effects of resets could influence valuation practice and hedging strategies.

major comments (3)

[Numerical scheme] Numerical scheme section: the discrete-time dynamic programming recursion replaces conditional expectations with neural-network approximations in a state space that includes the full rolling-window history of the underlying plus the evolving conversion price. No error bounds, convergence rates, or approximation guarantees specific to this high-dimensional path-dependent setting are supplied, even though the central claim of 'stable and accurate prices' rests on this step.
[Empirical tests] Empirical tests section: the assertion that the framework produces 'stable and accurate prices and sensitivity patterns' across GBM/CEV/Heston is supported only by qualitative stability on one real bond; no quantitative error metrics, comparison to benchmark solvers (finite-difference, Monte Carlo with control variates, or closed-form GBM cases without reset/call), or analysis of NN approximation error are reported.
[Model formulations and results] Model formulations and results: the three economic insights (contractual features dominate dynamics, call truncates upside, downward reset lowers effective call threshold) are derived from the NN-based prices. Without independent validation of the scheme on at least the GBM specification, these insights risk being artifacts of approximation bias rather than model behavior.

minor comments (1)

[Abstract] Abstract: the phrase 'stable and accurate prices' is used without defining accuracy criteria or referencing any benchmark.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the constructive and detailed comments. We agree that additional numerical validation and benchmark comparisons would strengthen the manuscript's claims regarding the reliability of the approximations and the robustness of the reported economic insights. We outline our responses and planned revisions below.

read point-by-point responses

Referee: Numerical scheme section: the discrete-time dynamic programming recursion replaces conditional expectations with neural-network approximations in a state space that includes the full rolling-window history of the underlying plus the evolving conversion price. No error bounds, convergence rates, or approximation guarantees specific to this high-dimensional path-dependent setting are supplied, even though the central claim of 'stable and accurate prices' rests on this step.

Authors: We acknowledge that the manuscript does not supply theoretical error bounds, convergence rates, or approximation guarantees for the neural-network approximations in this high-dimensional path-dependent setting. Deriving such guarantees is a significant theoretical undertaking that lies beyond the scope of the current applied work. In revision, we will add a dedicated discussion of the method's properties, citing relevant literature on neural-network approximations for dynamic programming and path-dependent PDEs, and include supplementary numerical experiments demonstrating empirical convergence with respect to network architecture and training parameters. revision: partial
Referee: Empirical tests section: the assertion that the framework produces 'stable and accurate prices and sensitivity patterns' across GBM/CEV/Heston is supported only by qualitative stability on one real bond; no quantitative error metrics, comparison to benchmark solvers (finite-difference, Monte Carlo with control variates, or closed-form GBM cases without reset/call), or analysis of NN approximation error are reported.

Authors: The referee is correct that the current empirical section provides only qualitative evidence from a single bond. We will revise the empirical tests to incorporate quantitative metrics, including the standard deviation of computed prices across repeated independent trainings and sensitivity analyses with respect to hyperparameters. For the GBM specification, we will add direct comparisons to Monte Carlo benchmarks (with and without control variates) and, for cases without reset or call features, to available closed-form or finite-difference solutions. revision: yes
Referee: Model formulations and results: the three economic insights (contractual features dominate dynamics, call truncates upside, downward reset lowers effective call threshold) are derived from the NN-based prices. Without independent validation of the scheme on at least the GBM specification, these insights risk being artifacts of approximation bias rather than model behavior.

Authors: We agree that the economic insights require independent validation on the GBM case to rule out approximation artifacts. In the revised manuscript, we will insert a new validation subsection that applies the scheme to the GBM dynamics in simplified settings (standard convertible bonds without path-dependent reset or call provisions) and compares the results against Monte Carlo benchmarks. The three insights will then be presented only after this validation step, with explicit caveats regarding the scope of the supporting evidence. revision: yes

Circularity Check

0 steps flagged

No circularity: numerical PPDE-DP scheme with NN approximation is independent of its outputs

full rationale

The derivation proceeds from PPDE formulation of the path-dependent convertible bond value (under GBM/CEV/Heston) to a discrete dynamic programming recursion whose conditional expectations are replaced by neural networks. No equation equates the final price to a fitted parameter, no prediction is a renaming of an input, and no load-bearing step reduces to a self-citation or ansatz smuggled from prior work by the same authors. The reported prices and sensitivities are computed outputs of the scheme rather than algebraically forced by its inputs. The absence of error bounds is a correctness concern, not a circularity one.

Axiom & Free-Parameter Ledger

1 free parameters · 2 axioms · 0 invented entities

The framework rests on the standard assumptions of the three diffusion models (GBM, CEV, Heston) plus the usual no-arbitrage and Markovian augmentation of the state space to handle path dependence. No new entities are postulated. The neural-network approximation introduces many free parameters whose values are fitted during training.

free parameters (1)

neural-network weights and biases
The conditional-expectation approximators are trained on simulated paths; their parameters are fitted to minimize the dynamic-programming loss.

axioms (2)

domain assumption The value function satisfies the path-dependent PDE derived from the chosen diffusion and the contractual rules.
Standard Itô calculus applied to the augmented state (stock price, conversion price, time).
ad hoc to paper Neural networks can accurately approximate the conditional expectations arising in the dynamic-programming recursion.
No convergence theorem or error bound is supplied in the abstract.

pith-pipeline@v0.9.0 · 5548 in / 1576 out tokens · 80133 ms · 2026-05-13T03:20:46.844119+00:00 · methodology

discussion (0)

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

IndisputableMonolith/Cost/FunctionalEquation.lean washburn_uniqueness_aczel unclear
We formulate the valuation problem as a path-dependent partial differential equation (PPDE)... discrete-time dynamic programming scheme in which conditional expectations are approximated by neural networks
IndisputableMonolith/Foundation/RealityFromDistinction.lean reality_from_one_distinction unclear
Theorem 2... V^h → V as h→0... neural network class is dense in L^2

Reference graph

Works this paper leans on

38 extracted references · 38 canonical work pages

[1]

Ammann, M., Kind, A., and Wilde, C

doi: 10.1016/S0378-4266(01)00256-4. Ammann, M., Kind, A., and Wilde, C. Simulation-based pricing of convertible bonds.Journal of Empirical Finance, 15(2):310–331,

work page doi:10.1016/s0378-4266(01)00256-4
[2]

Arandjelovi´ c, A., Shevchenko, P

doi: 10.1016/j.jempfin.2006.06.008. Arandjelovi´ c, A., Shevchenko, P. V., Matsui, T., Murakami, D., and Myrvoll, T. Solving stochastic climate-economy models: a deep least-squares Monte Carlo approach.Mathematical Finance,

work page doi:10.1016/j.jempfin.2006.06.008 2006
[3]

Bachouch, A., Hur´ e, C., Langren´ e, N., and Pham, H

doi: 10.3905/jod.2003.319208. Bachouch, A., Hur´ e, C., Langren´ e, N., and Pham, H. Deep neural networks algorithms for stochastic control problems on finite horizon: numerical applications.Methodology and Computing in Applied Probability, 24(1):143–178,

work page doi:10.3905/jod.2003.319208 2003
[4]

Barles, G

doi: 10.1007/s11009-019-09767-9. Barles, G. and Souganidis, P. E. Convergence of approximation schemes for fully nonlinear second order equations.Asymptotic Analysis, 4(3):271–283,

work page doi:10.1007/s11009-019-09767-9
[5]

19 Barone-Adesi, G., Bermudez, A., and Hatgioannides, J

doi: 10.3233/ASY-1991-4305. 19 Barone-Adesi, G., Bermudez, A., and Hatgioannides, J. Two-factor convertible bonds valuation using the method of characteristics/finite elements.Journal of Economic Dynamics and Control, 27(10): 1801–1831,

work page doi:10.3233/asy-1991-4305 1991
[6]

Batten, J

doi: 10.1016/S0165-1889(02)00083-0. Batten, J. A., Khaw, K. L.-H., and Young, M. R. Pricing convertible bonds.Journal of Banking and Finance, 92:216–236,

work page doi:10.1016/s0165-1889(02)00083-0
[7]

Brennan, M

doi: 10.1016/j.jbankfin.2018.05.006. Brennan, M. J. and Schwartz, E. S. Convertible bonds: valuation and optimal strategies for call and conversion.Journal of Finance, 32:1699–1715,

work page doi:10.1016/j.jbankfin.2018.05.006 2018
[8]

Brennan, M

doi: 10.1111/j.1540-6261.1977.tb03364.x. Brennan, M. J. and Schwartz, E. S. Analyzing convertible bonds.Journal of Financial and Quantitative Analysis, 15(4):907–929,

work page doi:10.1111/j.1540-6261.1977.tb03364.x 1977
[9]

Carriere, J

doi: 10.2307/2330567. Carriere, J. Valuation of the early-exercise price for options using simulations and nonparametric regression.Insurance: Mathematics and Economics, 19(1):19–30,

work page doi:10.2307/2330567
[10]

2023.100423

doi: 10.1016/j.rinam. 2023.100423. Costabile, M. and Viviano, F. Combining lattice and regression methods for the evaluation of convert- ible bonds with soft call/put provisions.Journal of Computational and Applied Mathematics, 477: 117208,

work page doi:10.1016/j.rinam 2023
[11]

doi: 10.1016/j.cam.2025.117208. Cox, J. C., Ross, S. A., and Rubinstein, M. Option pricing: a simplified approach.Journal of Financial Economics, 7(3):229–263,

work page doi:10.1016/j.cam.2025.117208 2025
[12]

Ekren, I., Touzi, N., and Zhang, J

doi: 10.1016/0304-405X(79)90015-1. Ekren, I., Touzi, N., and Zhang, J. Viscosity solutions of fully nonlinear parabolic path dependent PDEs: Part I.Annals of Probability, 44(2):1212–1253, 2016a. doi: 10.1214/14-AOP999. Ekren, I., Touzi, N., and Zhang, J. Viscosity solutions of fully nonlinear parabolic path dependent PDEs: Part II.Annals of Probability, 4...

work page doi:10.1016/0304-405x(79)90015-1
[13]

Feng, Y., Huang, B.-H., and Huang, Y

doi: 10.21314/JCF.2021.006. Feng, Y., Huang, B.-H., and Huang, Y. Valuing resettable convertible bonds: based on path decom- posing.Finance Research Letters, 19:279–290,

work page doi:10.21314/jcf.2021.006 2021
[14]

Franco, C

doi: 10.1016/j.frl.2016.09.002. Franco, C. d., Nicolle, J., and Pham, H.Discrete-time portfolio optimization under maximum drawdown constraint with partial information and deep learning resolution, pages 101–136. Springer,

work page doi:10.1016/j.frl.2016.09.002 2016
[15]

doi: 10.1007/978-3-030-98519-6

work page doi:10.1007/978-3-030-98519-6
[16]

Hur´ e, C., Pham, H., Bachouch, A., and Langren´ e, N

doi: 10.3905/jod.1994.407908. Hur´ e, C., Pham, H., Bachouch, A., and Langren´ e, N. Deep neural networks algorithms for stochastic control problems on finite horizon: convergence analysis.SIAM Journal on Numerical Analysis, 59 (1):525–557,

work page doi:10.3905/jod.1994.407908 1994
[17]

Ingersoll, J

doi: 10.1137/20M1316640. Ingersoll, J. E. A contingent-claims valuation of convertible securities.Journal of Financial Economics, 4(3):289–321,

work page doi:10.1137/20m1316640
[18]

Kimura, T

doi: 10.1016/0304-405X(77)90004-6. Kimura, T. and Shinohara, T. Monte Carlo analysis of convertible bonds with reset clauses.European Journal of Operational Research, 168(2):301–310,

work page doi:10.1016/0304-405x(77)90004-6
[19]

doi: 10.1016/j.ejor.2004.07.008. Lau, K. W. and Kwok, Y. K. Anatomy of option features in convertible bonds.Journal of Futures Markets, 24(6):513–532,

work page doi:10.1016/j.ejor.2004.07.008 2004
[20]

Lewis, C

doi: 10.1002/fut.10127. Lewis, C. M. Convertible debt: valuation and conversion in complex capital structures.Journal of Banking and Finance, 15(3):665–682,

work page doi:10.1002/fut.10127
[21]

Li, L., Hu, Z., Guo, H., Wang, L., Zhang, Z., and Xu, Y

doi: 10.1016/0378-4266(91)90091-Y. Li, L., Hu, Z., Guo, H., Wang, L., Zhang, Z., and Xu, Y. Pricing convertible bonds based on GAN and transformer.Computational Economics, 2025a. doi: 10.1007/s10614-025-11021-z. 20 Li, Z., Wang, Y., Qiao, F., and Yu, M. Convertible bond return predictability with machine learning. Journal of Financial Markets, page 101010...

work page doi:10.1016/0378-4266(91)90091-y 2025
[22]

doi: 10.1080/14697688.2026.2616345. Lin, S. and Zhu, S.-P. Numerically pricing convertible bonds under stochastic volatility or stochastic interest rate with an ADI-based predictor–corrector scheme.Computers and Mathematics with Applications, 79(5):1393–1419,

work page doi:10.1080/14697688.2026.2616345 2026
[23]

doi: 10.1016/j.camwa.2019.09.003. Liu, Q. and Guo, S. An excellent approximation for themout ofnday provision.North American Journal of Economics and Finance, 54:101222,

work page doi:10.1016/j.camwa.2019.09.003 2019
[24]

Longstaff, F

doi: 10.1016/j.najef.2020.101222. Longstaff, F. A. and Schwartz, E. S. Valuing American options by simulation: a simple least-squares approach.Review of Financial Studies, 14(1):113–147,

work page doi:10.1016/j.najef.2020.101222 2020
[25]

doi: 10.1093/rfs/14.1.113. Ma, C. and Xu, W. Valuation model for Chinese convertible bonds with soft call/put provision under the hybrid willow tree.Quantitative Finance, 20(12):1999–2014,

work page doi:10.1093/rfs/14.1.113 1999
[26]

doi: 10.1080/14697688.2020. 1814022. McGuinness, P. B. and Keasey, K. The listing of Chinese state-owned banks and their path to banking and ownership reform.China Quarterly, 201:125–155,

work page doi:10.1080/14697688.2020 2020
[27]

Nyborg, K

doi: 10.1017/S030574100999110X. Nyborg, K. G. The use and pricing of convertible bonds.Applied Mathematical Finance, 3(3):167–190,

work page doi:10.1017/s030574100999110x
[28]

doi: 10.1080/13504869600000009. Roch, A. Optimal liquidation through a limit order book: a neural network and simulation ap- proach.Methodology and Computing in Applied Probability, 25(3):1–29,

work page doi:10.1080/13504869600000009
[29]

Warin, X

doi: 10.3905/jfi.1998.408243. Warin, X. Reservoir optimization and machine learning methods.EURO Journal on Computational Optimization, 11:100068,

work page doi:10.3905/jfi.1998.408243 1998
[30]

Xu, W., Yu, B., Yao, L., and Zhou, Y

doi: 10.1016/j.ejco.2023.100068. Xu, W., Yu, B., Yao, L., and Zhou, Y. A tree model for pricing convertible bonds with equity, market and default risk.Computational Economics, 34(2):169–194,

work page doi:10.1016/j.ejco.2023.100068 2023
[31]

Monte Carlo analysis of convertible bonds with reset clause

doi: 10.1007/s10614-009-9170-8. Yang, J., Choi, Y., Li, S., and Yu, J. A note on “Monte Carlo analysis of convertible bonds with reset clause”.European Journal of Operational Research, 200(3):924–925,

work page doi:10.1007/s10614-009-9170-8
[32]

doi: 10.1016/j.ejor.2009. 02.012. Yarotsky, D. Error bounds for approximations with deep ReLU networks.Neural Networks, 94: 103–114,

work page doi:10.1016/j.ejor.2009 2009
[33]

Zabolotnyuk, Y., Jones, R

doi: 10.1016/j.neunet.2017.07.002. Zabolotnyuk, Y., Jones, R. J., and Fan, H. An empirical comparison of convertible bond valuation models.Financial Management, 39(2):675–705,

work page doi:10.1016/j.neunet.2017.07.002 2017
[34]

Going Out

doi: 10.1111/j.1755-053X.2010.01088.x. Zhang, S. and Zhu, X. Financing structure and risk response in overseas investment and M&A by Chinese enterprises. InThe Challenge of “Going Out” Chinese Experiences in Outbound Investment, pages 71–87. Springer,

work page doi:10.1111/j.1755-053x.2010.01088.x 2010
[35]

doi: 10.1007/978-981-99-3326-6

work page doi:10.1007/978-981-99-3326-6
[36]

Zhu, S.-P

doi: 10.3390/axioms13040218. Zhu, S.-P. A closed-form analytical solution for the valuation of convertible bonds with constant dividend yield.ANZIAM Journal, 47(4):477–494,

work page doi:10.3390/axioms13040218
[37]

Zhu, S.-P., Lin, S., and Lu, X

doi: 10.1017/S1446181100010087. Zhu, S.-P., Lin, S., and Lu, X. Pricing puttable convertible bonds with integral equation approaches. Computers and Mathematics with Applications, 75(8):2757–2781,

work page doi:10.1017/s1446181100010087
[38]

doi: 10.1016/j.camwa.2018. 01.007. 21 A Proofs in Section 2 A.1 Proof of Theorem 1 Proof.We derive the PPDE satisfied by the convertible bond value functionalV(t, ω t, Ht) on the continuation regionC. By definition ofC, on the continuation region, neither the reset clause nor the call clause is active, so the conversion priceH t remains locally constant a...

work page doi:10.1016/j.camwa.2018 2018