Backward Deep BSDE Methods and Applications to Nonlinear Problems
Abstract
:1. Introduction
2. FBSDE for Nonlinear Problems
2.1. Backward TimeStepping
2.1.1. Exact Backward TimeStepping
2.1.2. TimeStepping from Taylor Expansion
3. Deep BSDE Approach
3.1. Forward Approach
3.2. Backward Approach
Algorithm 1 The pathwise Backward deep BSDE Method 

4. Results
4.1. Call Combination
4.2. Straddle
5. Conclusions
6. Disclaimer
Author Contributions
Funding
Data Availability Statement
Acknowledgments
Conflicts of Interest
Notes
1  Set ${Z}_{t}={\Pi}_{t}^{T}\sigma (t,{X}_{t})$ and $\tilde{f}(t,X,Y,Z)=f(t,X,Y,\sigma {(t,{X}_{t})}^{T}{Z}_{t}^{T})$. Then, (Pardoux 1998, Theorem 2.2) with this ${Z}_{t}$ and $\tilde{f}$ shows that one can construct a solution ${Y}_{t}$ and ${Z}_{t}$ of the BSDE from a classical ${C}^{2,1}$ solution of a corresponding PDE and a solution ${X}_{t}$. Rewriting the PDE and BSDE in terms of ${\Pi}_{t}^{T}$ instead of ${Z}_{t}$ and using function f rather than $\tilde{f}$ gives the form reported below. For the opposite direction, (Pardoux 1998, Theorem 2.4) shows how a solution ${Y}_{t}$ of the BSDE (corresponding to a X that starts at x at time t) gives a continuous function $u(t,x)$, which is a viscosity solution of the corresponding PDE. 
2  In general, the terminal condition could be given as a random variate ${G}_{T}$ that is measurable with respect to the information available of time T (i.e., the sigma algebra generated by ${X}_{t}$ with $t\le T$). The FBSDE approach then will be more general than the PDE approach. If there is an exact (or approximate) Markovianization with a Markov state ${M}_{t}$, the strategy ${\Pi}_{t}$ and the solution ${Y}_{t}$ would in general be functions $\pi (t,{M}_{t})$ and $u(t,{M}_{t})$ of that Markov state. We only treat the usual final value case here. 
3  
4  This is actually the expectation of the conditional variance $\mathsf{var}\left({\mathcal{Y}}_{0}^{\pi}\right{X}_{0})$ over the distribution of ${X}_{0}$. 
5  Here, we consider the 1dimensional case, while Han and Jentzen (2017) consider the 100dimensional case. 
6  Notice that (Forsyth and Labahn 2007, Figure 1) do not give the number of space or timesteps used for their plot. 
References
 Forsyth, Peter A., and George Labahn. 2007. Numerical methods for controlled HamiltonJacobiBellman PDEs in finance. Journal of Computational Finance 11: 1–44. Available online: https://cs.uwaterloo.ca/~paforsyt/hjb.pdf (accessed on 15 March 2023). [CrossRef] [Green Version]
 Ganesan, Narayan, Yajie Yu, and Bernhard Hientzsch. 2022. Pricing barrier options with deep backward stochastic differential equation methods. Journal of Computational Finance 25. Available online: https://ssrn.com/abstract=3607626 (accessed on 15 March 2023). [CrossRef]
 Han, Jiequn, and Arnulf Jentzen. 2017. Deep learningbased numerical methods for highdimensional parabolic partial differential equations and backward stochastic differential equations. Communications in Mathematics and Statistics 5: 349–80. [Google Scholar]
 Han, Jiequn, Arnulf Jentzen, and Weinan E. 2018. Solving highdimensional partial differential equations using deep learning. Proceedings of the National Academy of Sciences 115: 8505–10. [Google Scholar] [CrossRef] [PubMed] [Green Version]
 Hientzsch, Bernhard. 2021. Deep learning to solve forwardbackward stochastic differential equations. Risk Magazine. February. Available online: https://ssrn.com/abstract=3494359 (accessed on 15 March 2023).
 Huré, Côme, Huyên Pham, and Xavier Warin. 2020. Deep backward schemes for highdimensional nonlinear PDEs. Mathematics of Computation 89: 1547–79. [Google Scholar] [CrossRef] [Green Version]
 Liang, Jian, Zhe Xu, and Peter Li. 2021. Deep learningbased least squares forwardbackward stochastic differential equation solver for highdimensional derivative pricing. Quantitative Finance 21: 1309–23. Available online: https://ssrn.com/abstract=3381794 (accessed on 15 March 2023). [CrossRef]
 Mercurio, Fabio. 2015. Bergman, Piterbarg, and beyond: Pricing derivatives under collateralization and differential rates. In Actuarial Sciences and Quantitative Finance. Berlin: Springer, pp. 65–95. Available online: https://ssrn.com/abstract=2326581 (accessed on 15 March 2023).
 Pardoux, Étienne. 1998. Backward stochastic differential equations and viscosity solutions of systems of semilinear parabolic and elliptic PDEs of second order. In Stochastic Analysis and Related Topics VI. Berlin: Springer, pp. 79–127. [Google Scholar]
 Raissi, Maziar. 2018. Forwardbackward stochastic neural networks: Deep learning of highdimensional partial differential equations. arXiv arXiv:1804.07010. [Google Scholar]
 Wang, Haojie, Han Chen, Agus Sudjianto, Richard Liu, and Qi Shen. 2018. Deep learningbased BSDE solver for LIBOR market model with application to bermudan swaption pricing and hedging. arXiv arXiv:1807.06622. Available online: https://ssrn.com/abstract=3214596 (accessed on 15 March 2023). [CrossRef]
 Warin, Xavier. 2018. Nesting Monte Carlo for highdimensional nonlinear PDEs. Monte Carlo Methods and Applications 24: 225–47. [Google Scholar] [CrossRef] [Green Version]
Method  Upper Price  Lower Price 

Results from Forsyth and Labahn101 nodes  
Fully Implicit HJB PDE (implicit control)  24.02  23.06 
Crank–Nicolson HJB PDE (implicit control)  24.05  23.09 
Fully Implicit HJB PDE (pwc policy)  24.01  23.07 
Crank–Nicolson HJB PDE (pwc policy)  24.07  23.09 
Forward deep BSDE—20,000 batches, size 256  
Learned ${Y}_{0}$ (shared)  24.06 (23.99–24.14)  23.10 (23.02–23.17) 
Learned ${Y}_{0}$ (separate)  24.07 (24.01–24.12)  23.10 (23.06–23.15) 
Backward deep BSDE—20,000 batches, size 256  
Batch variance/1 (shared)  24.14 (23.98–24.30)  23.19 (23.00–23.37) 
Batch variance/100 (shared)  24.08 (24.06–24.09)  23.13 (23.09–23.16) 
Batch variance/1 (separate)  24.06 (23.94–24.19 *)  23.10 (22.95–23.25) 
Batch variance/100 (separate)  24.07 (24.06–24.09 *)  23.12 (23.10–23.13) 
Learned ${Y}_{0}$ (shared)  24.06 (23.99–24.14)  23.10 (23.02–23.17) 
Learned ${Y}_{0}$ (separate)  24.06 (24.01–24.11 *)  23.10 (23.06–23.15) 
Method  Upper Price  Lower Price 

Results from Forsyth and Labahn—101 nodes  
Fully Implicit (implicit control)  24.02  23.06 
Crank–Nicolson (implicit control)  24.05  23.09 
Fully Implicit (pwc policy)  24.01  23.07 
Crank–Nicolson (pwc policy)  24.07  23.09 
Backward deep BSDE—20,000 batches, size 512, five seeds  
Batch variance/1 (shared)  24.02–24.08 (23.87–24.29)  23.06–23.12 (22.91–23.33) 
Batch variance/100 (shared)  24.06–24.07 (24.03–24.09)  23.11–23.12 (23.09–23.15) 
Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to people or property resulting from any ideas, methods, instructions or products referred to in the content. 
© 2023 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC BY) license (https://creativecommons.org/licenses/by/4.0/).
Share and Cite
Yu, Y.; Ganesan, N.; Hientzsch, B. Backward Deep BSDE Methods and Applications to Nonlinear Problems. Risks 2023, 11, 61. https://doi.org/10.3390/risks11030061
Yu Y, Ganesan N, Hientzsch B. Backward Deep BSDE Methods and Applications to Nonlinear Problems. Risks. 2023; 11(3):61. https://doi.org/10.3390/risks11030061
Chicago/Turabian StyleYu, Yajie, Narayan Ganesan, and Bernhard Hientzsch. 2023. "Backward Deep BSDE Methods and Applications to Nonlinear Problems" Risks 11, no. 3: 61. https://doi.org/10.3390/risks11030061