TL;DR
This paper introduces "conditional optimal stopping," a new problem where an agent's decision to stop depends on a future event occurring. They develop an equilibrium framework to handle the inherent time-inconsistency of this problem, showing that it leads to unique solutions in finite horizons but complex, non-unique equilibria in infinite horizons.
What Problem Does This Paper Address?
The classical optimal stopping problem involves maximizing an expected payoff by choosing the best time to stop a process. However, real-world decisions often involve conditions. For example, a company might only care about payoffs if it remains solvent, or an individual might only consider outcomes while they are alive. This paper addresses the challenge of modeling such "conditional optimal stopping" problems. The core difficulty is that the probability of the conditioning event changes dynamically as the agent makes decisions, leading to time-inconsistency – a strategy optimal today might not be optimal when reconsidered tomorrow. This is akin to Strotz's concept of sophisticated agents who anticipate their future selves' potentially different preferences.
Methods
The authors introduce a novel framework for optimal stopping that accounts for time-inconsistency arising from conditioning events. They adapt Strotz's concept of sophisticated agents, focusing on "subgame perfect Nash equilibria" where agents choose strategies that their future selves won't want to deviate from.
For finite time horizons, they develop a backward recursion that constructs an equilibrium. This recursion involves not only a value process (similar to the classical Snell envelope) but also a "survival process" that tracks the evolving conditioning probability.
For infinite time horizons, they establish existence results and then explore the complexities of non-uniqueness and non-Markovian equilibria, even in Markovian settings. They also introduce the concept of a "Snell pair" (a value process and a survival process) as a generalization of the Snell envelope, which is shown to be in one-to-one correspondence with equilibria.
Datasets
Not discussed in the paper.
Key Insights and Ideas
- Time-Inconsistency is Inherent: Conditioning on future events fundamentally breaks the time-consistency of optimal stopping problems, even with standard discounting.
- Finite Horizon Simplicity: In finite horizons, a unique equilibrium can be constructed via backward recursion, similar to classical optimal stopping, but requiring an additional "survival process."
- Infinite Horizon Complexity: The infinite horizon case is much richer, allowing for multiple equilibria, including non-Markovian ones even in Markovian settings. This challenges previous conjectures.
- "Always Stopping" Isn't Always Optimal: Unlike some time-inconsistent problems, "always stopping" is not necessarily an equilibrium in this framework.
- Snell Pairs: The introduction of Snell pairs (value and survival processes) provides a unifying theoretical tool that generalizes the classical Snell envelope and connects directly to equilibria.
Key Results
- Finite Horizon: Existence and uniqueness of an equilibrium (with a specific early-stopping preference) are proven for finite horizons. This equilibrium is Markovian if the underlying process is Markovian.
- Infinite Horizon: Existence of equilibria is established under certain conditions. However, examples demonstrate that equilibria can be non-unique and non-Markovian, even in Markovian settings.
- Equilibrium Construction: A constructive method for finding equilibria is provided, based on backward recursion for finite horizons and a limiting procedure for infinite horizons.
- Snell Pair Equivalence: A direct correspondence is established between Snell pairs and equilibria with early stopping preference.
Limitations and Future Work
- Non-Markovian Settings: While the paper focuses on Markovian settings for tractability, the theory is developed to handle non-Markovian cases, suggesting future work in this direction.
- Infinite Horizon Nuances: The paper highlights the surprising non-uniqueness and time-dependent nature of equilibria in infinite horizons, suggesting further exploration of these phenomena.
- General Preferences: The focus is on equilibria with early stopping preference. Extending the theory to other preferences (e.g., late stopping) is mentioned as a possibility.
- Numerical Methods: While theoretical results are presented, numerical methods for solving these conditional optimal stopping problems in more complex settings could be a future research avenue.
Summary updated