💰 quantitative finance

The Compound BSDE Method: A Fully Forward Method for Option Pricing and Optimal Stopping Problems in Finance

This paper introduces the Compound BSDE method, a fully forward deep-learning-based algorithm that reformulates option pricing and optimal stopping problems as systems of backward stochastic differential equations to achieve accurate and efficient solutions for high-dimensional financial problems.

Original authors: Zhipeng Huang, Cornelis W. Oosterlee

Published 2026-02-02

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Zhipeng Huang, Cornelis W. Oosterlee

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to price a very complicated financial contract, like an option that gives you the right to buy another option, which in turn gives you the right to buy a third one. Or imagine an American-style option where you can choose to cash out at any of 50 different dates. Traditionally, solving these problems is like trying to navigate a maze by starting at the exit and working your way backward to the entrance. It's difficult, especially when the maze has hundreds of dimensions (many different variables changing at once).

This paper introduces a new way to solve these problems called the Compound BSDE Method. Instead of walking backward through the maze, this method walks forward from the start, using a smart computer program (a neural network) to learn the path as it goes.

Here is how it works, broken down into simple concepts:

1. The Problem: The "Russian Doll" of Options

Think of a Compound Option like a set of Russian nesting dolls.

The outer doll is the main option you are looking at.
Inside it is the inner doll, which is another option.
To know the value of the outer doll today, you need to know the value of the inner doll at a future date. But to know the value of that inner doll, you might need to know the value of a third doll inside it.

Traditionally, financial math tries to solve this by starting at the very end (when the last doll is opened) and calculating the value step-by-step backward to today. This gets messy and slow when there are many dolls or many variables.

2. The Solution: The "Forward Train"

The authors propose a method that treats the whole chain of options as a single, connected system that moves forward in time.

Imagine a train with several cars (representing the different time periods of the option).

The Engine (The Forward Process): The train moves forward along the tracks, simulating how the stock price or asset value changes day by day. This is the "forward" part.
The Coupling (The Compound Condition): At specific stations (dates), the cars are linked together. The value of the car arriving at Station A must match the value of the car leaving Station B, adjusted for the rules of the contract (like paying a strike price or deciding whether to exercise the option).
The Smart Driver (The Neural Network): Instead of calculating the value of every car manually, a computer program (a neural network) acts as the driver. It guesses the values for the cars at the start. As the train moves forward, the driver checks if the links between the cars are tight. If the values don't match the rules at the stations, the driver adjusts its guess and tries again.

3. How It Learns: The "Group Hug"

In older methods, you might try to fix one car at a time, working backward. This new method uses a Compound BSDE (Backward Stochastic Differential Equation).

Think of it like a group of friends trying to agree on a story.

Friend A (Time 1) needs to tell a story that connects perfectly to Friend B (Time 2).
Friend B needs to connect to Friend C (Time 3).
Friend C tells the final story at the end.

Instead of asking Friend C to tell the story first and then working backward to Friend A, the Compound BSDE method asks everyone to tell their part of the story simultaneously while moving forward. They all adjust their stories until the connections between them are perfect. The computer minimizes the "mistakes" (errors) in these connections until the whole story makes sense from start to finish.

4. Why It's Special

It's Fully Forward: Unlike other methods that require looking backward (which is hard to do with complex, high-dimensional data), this method simulates everything forward, which is much faster and easier for computers to handle.
It Handles Complexity: It works great for "high-dimensional" problems. Imagine a basket of 50 different fruits (stocks) instead of just one apple. Traditional methods struggle with 50 variables, but this "forward train" method handles them easily.
It's Accurate: The authors proved mathematically that as you make the time steps smaller (more stops on the train), the answer gets closer to the true value. They tested this on various "nesting doll" options and "early exercise" options (where you can cash out anytime), and the results were very accurate.

Summary

The paper presents a new, "fully forward" way to price complex financial options. Instead of solving a puzzle by starting at the end and working backward, it simulates the future forward in time, using a smart computer program to ensure that all the different stages of the option (the "compound" parts) fit together perfectly. It's like training a team to walk forward in perfect sync, rather than trying to march backward in a line. This makes it possible to price very complex, multi-step financial contracts that were previously too difficult to calculate accurately.

Technical Summary: The Compound BSDE Method

Problem Statement

Financial mathematics faces significant challenges in pricing derivatives with complex payoff structures, early-exercise features (such as American and Bermudan options), and high-dimensional underlying dynamics. While traditional methods like PDEs, Fourier techniques, and Monte Carlo regression exist, they often struggle with the "curse of dimensionality" or the computational complexity of backward induction.

Backward Stochastic Differential Equations (BSDEs) offer a flexible framework for these problems via the Feynman–Kac representation. However, existing deep-learning-based approaches for BSDEs, such as the Deep BSDE method, are primarily designed for single BSDEs. While recent variants attempt to address optimal stopping (e.g., Bermudan options) by minimizing the variance of the initial value, these methods often rely on backward simulation or lack established convergence guarantees for early-exercise problems. The core problem addressed is the need for a fully forward, deep-learning-based numerical method capable of solving coupled systems of BSDEs that naturally model multi-stage payoffs and optimal stopping without requiring backward induction.

Methodology: The Compound BSDE Method

The paper proposes the Compound BSDE method, a reformulation of option pricing and optimal stopping problems into a system of coupled backward stochastic differential equations.

1. The Compound BSDE Formulation

The method introduces an $M$ -fold compound BSDE defined on a filtered probability space with time intervals $0 = T_0 < T_1 < \dots < T_M = T$ . The system consists of:

A single forward SDE for the underlying asset $X_t$ .
$M$ backward equations defined on successive intervals $[T_{j-1}, T_j]$ .
Coupling Conditions: The equations are linked at intermediate times $T_j$ via compounding functions $g_j$ . Specifically, the terminal condition of the $j$ -th backward equation depends on the solution of the $(j+1)$ -th equation at time $T_j$ :
$Y_{j, T_j} = g_j(X_{T_j}, Y_{j+1, T_j})$
The final equation ( $j=M$ ) terminates with a standard condition $g_M(X_T)$ .

This structure allows the representation of complex financial products where the value at an intermediate time depends on the value of a subsequent option or an exercise decision.

2. The Numerical Algorithm

The proposed algorithm is a fully forward method. Unlike traditional backward schemes that simulate paths forward and optimize backward, this method:

Simulates the forward SDE and all $M$ backward equations simultaneously in the forward time direction.
Parameterizes the initial values $\{Y_{j, T_{j-1}}\}$ and the processes $\{Z_{j, t_i}\}$ using neural networks.
Minimizes a joint objective functional that enforces both the terminal condition and the intermediate compounding conditions:
$\mathcal{L} = \sum_{j=1}^{M-1} \mathbb{E}\|g_j(X^\pi_{T_j}, Y^\pi_{j+1, T_j}) - Y^\pi_{j, T_j}\|^2 + \mathbb{E}\|g_M(X^\pi_{T_M}) - Y^\pi_{M, T_M}\|^2$
The neural networks are trained to minimize this loss, effectively learning the solution to the entire coupled system simultaneously.

3. Theoretical Foundations

The paper establishes the well-posedness of the compound BSDE under standard Lipschitz and regularity assumptions. Key theoretical results include:

Existence and Uniqueness: The system admits a unique adapted solution, which corresponds to the viscosity solution of a system of semilinear PDEs.
Convergence Analysis: Building on the a posteriori error estimates for the standard Deep BSDE method, the authors derive an a posteriori error estimate for the Compound BSDE method. The total error is bounded by the time step size $h$ and the training loss (the value of the objective functional). This proves that if the neural networks successfully minimize the loss, the numerical solution converges to the true solution.

Key Contributions

New Formulation: The introduction of the Compound BSDE, a new mathematical formulation that unifies the pricing of multi-stage derivatives (like compound options) and optimal stopping problems (like Bermudan options) under a single framework.
Fully Forward Algorithm: The development of a numerical algorithm that is naturally fully forward. It avoids the backward induction step required by many existing methods for early-exercise options, allowing for simultaneous learning of all stages.
Convergence Guarantees: The establishment of rigorous convergence properties and an a posteriori error estimate for the proposed method, addressing a gap in the literature where the convergence of deep learning methods for Bermudan options had not been fully established.
Unified Framework: Demonstrating that diverse financial problems, from simple compound options to high-dimensional Bermudan basket options, can be solved using the same algorithmic template by simply adjusting the compounding functions $g_j$ .

Numerical Results

The authors validate the method through extensive numerical experiments:

Plain Compound Options ( $M=2$ ): The method accurately prices call-on-call, call-on-put, put-on-call, and put-on-put options, with relative mean squared errors (reMSE) as low as $10^{-7}$ . The results confirm first-order convergence with respect to the time step size.
M-fold Compound Options: The method scales effectively to $M \in \{2, 3, 4, 5\}$ , maintaining accuracy and stability as the number of compounding stages increases.
High-Dimensional Bermudan Options: The method is applied to Bermudan geometric basket put options in dimensions up to $d_1 = 50$ . The results show that the method achieves accurate prices and hedges (deltas) in high-dimensional settings, with errors remaining stable as dimensionality increases. The convergence behavior aligns with the theoretical error bounds derived in the paper.

Significance and Claims

The paper claims that the Compound BSDE method provides a unified numerical framework for both multi-stage derivatives and discrete-time optimal stopping problems. Its primary significance lies in:

Overcoming the Backward Limitation: By being a fully forward method, it circumvents the need for backward simulation, which is often computationally expensive or difficult to implement for high-dimensional optimal stopping.
Theoretical Rigor: It provides a mathematically grounded approach with proven convergence, unlike some heuristic deep learning approaches for early-exercise options.
Flexibility: The method's ability to handle high-dimensional problems and complex payoff structures (via the choice of $g_j$ ) makes it a versatile tool for modern financial engineering.

The authors conclude that while the method is currently limited by Lipschitz assumptions on coefficients, it offers a flexible and theoretically grounded route to forward deep-learning solvers for complex pricing problems. Future work may involve relaxing these assumptions and incorporating variance reduction techniques.