Continuous-time Optimal Stopping through Deep Reinforcement Learning

Battle Count: 82

Directly applicable to American option pricing, which is ubiquitous in quantitative finance (all single-name Puts in U.S. markets). The CARLOS algorithm closes the Bermudan-American value gap that has persisted for 25 years since Longstaff-Schwartz. Achieves 3-10x speedup over brute-force fine-grid LSMC while delivering higher prices approaching the American upper bound. Highly relevant for derivatives desk pricing, risk management, and hedging of early-exercise features.

The algorithm involves multiple interacting components: LSMC initialization, ADNN training with adaptive sampling, grid refinement scheduling, delayed stopping logic, validation-based grid transitions, and learning rate decay. While the architecture is a standard feedforward NN, the RL training loop with concept drift management, anchor set construction, and the exploration-exploitation balance requires careful implementation. The paper provides detailed pseudocode and parameter guidelines, reducing implementation difficulty. Computation runs on CPU (no GPU required for the tested dimensions).

The paper provides detailed algorithm pseudocode (Algorithms 1-3), comprehensive parameter tables (Tables 2, 4, 5, 6, A.1-A.4), explicit NN architecture specifications, activation functions, learning rates, and grid schedules. All benchmark contracts are fully specified with parameters. However, no code repository is mentioned. The method is model-based (Black-Scholes), making simulation reproducible. Five independent runs with standard deviations are reported.