Imagine you hire a professional chef (the Agent) to cook a gourmet meal for a dinner party you are hosting (the Principal). You want the meal to be delicious, but you can't stand in the kitchen watching every chop and stir. You have to trust them.

Usually, you agree on a payment plan: a base salary plus a bonus if the food is amazing. But this paper introduces a twist: What if the power goes out, or the gas line bursts, or the building collapses before the meal is even served?

In the financial world, this "power outage" is called a random default. It means the investment market might suddenly stop existing or become inaccessible at an unpredictable moment. This paper asks: How do you design the perfect contract and investment strategy when the clock might stop ticking at any random second?

Here is a breakdown of their findings using simple analogies:

1. The Two Scenarios: "Maybe" vs. "For Sure"

The authors realized there are two very different ways this "power outage" can happen, and they require different math to solve:

The "Maybe" Scenario (Unbounded): Imagine a party where the power might go out, but it could also run all night. The market might crash tomorrow, or it might last for years.
- The Strategy: The chef (manager) knows the power could go out, so they might take a few risks early on to get a great dish ready just in case.
- The Math: This is like solving a puzzle where the finish line is fuzzy. The authors used advanced tools called BSDEs (Backward Stochastic Differential Equations) to figure out the best moves.
The "For Sure" Scenario (Bounded): Imagine a party where you know for a fact the building will be condemned at exactly 10:00 PM. The power will go out before the night ends, but you don't know when (5:00 PM? 8:00 PM?).
- The Strategy: This is trickier. The chef knows the game must end soon. They can't wait until the last minute to cook. They have to work harder and faster as the deadline approaches.
- The Math: This is the harder puzzle. The math equations get "degenerate" (they break down or become unstable) as the deadline nears. The authors had to invent new ways to handle these "exploding" numbers.

2. The Contract: The "Incentive Menu"

In a normal job, you pay someone a salary. In this paper, the "salary" is a complex menu of incentives designed to keep the chef happy and motivated even when the power might go out.

The contract includes:

Base Pay: A guaranteed amount (like a reservation fee).
Performance Bonus: A cut of the profits if the portfolio grows.
The "Black Swan" Insurance: A special payment if the market crashes (the default happens). This is crucial. If the market disappears, the chef needs to know they will still get paid for the work they did before the crash.
Tracking Penalty: If the chef is supposed to follow a safe recipe (like a life insurance fund) but goes off-script to gamble, they get penalized.

3. The "Actor-Critic" Computer Game

Solving these equations is so hard that standard calculators can't do it. The authors had to build a Deep Learning AI (a type of computer brain) to solve it.

They used a method called Actor-Critic, which is like a video game training loop:

The Actor: The AI tries to find the best investment strategy (the "moves").
The Critic: The AI checks if those moves actually lead to the best result (the "score").
They take turns. The Actor makes a move, the Critic grades it, the Actor adjusts, and they repeat until they find the perfect strategy. This allowed them to solve problems that were previously impossible to calculate.

4. What They Discovered (The Results)

By running these computer simulations, they found some interesting behaviors:

Risk Aversion Matters: If the chef is more scared of losing money than the host is, the chef will play it safe, and the meal (portfolio) won't be as good. If the chef is brave and the host is cautious, the chef might take too many risks. The best results happen when their "fear levels" are aligned.
The Shape of Time: If the "power outage" is likely to happen early (like a right-skewed distribution), the chef goes all-out immediately to get the value before the crash. If the outage is likely to happen late (left-skewed), the chef might wait, but the host has to offer bigger and bigger bonuses to keep the chef motivated as the deadline looms.
Linear vs. Complex Contracts: Real-world contracts are often simple (e.g., "I'll give you 20% of the profits"). The authors found that while simple contracts work, they aren't perfect. A complex, custom-tailored contract (like the one they designed) can significantly boost performance, but it's much harder to calculate.

Summary

This paper is about designing the perfect "deal" between an investor and a manager when the future is uncertain and the market might vanish at any moment. They showed that:

You need different math depending on whether the market might vanish or will vanish.
You need to pay the manager extra insurance if the market crashes.
You need super-computers (AI) to figure out the exact numbers because the math is too messy for humans to solve by hand.

It's essentially a guide on how to keep a chef motivated to cook a great meal, even if the kitchen might burn down tomorrow.

Technical Summary: Delegated Portfolio Management with Random Default

1. Problem Formulation

This paper addresses the Principal-Agent (PA) problem in delegated portfolio management under the uncertainty of a random default time ( $\tau$ ). The framework involves an investor (Principal) delegating asset management to a professional fund manager (Agent). The core challenge is the exogenous random time $\tau$ , after which neither party can access the market, rendering the investment horizon uncertain.

The problem is modeled as a bi-level stochastic optimization:

The Agent's Problem: Given a compensation contract $\xi$ , the Agent chooses an investment strategy $\pi$ to maximize their expected utility, subject to a tracking error penalty against a benchmark $\alpha$ . The Agent possesses Constant Absolute Risk Aversion (CARA) utility.
The Principal's Problem: The Principal designs the contract $\xi$ to maximize their own expected utility (derived from the terminal wealth minus the compensation), subject to the Agent's participation constraint (reservation utility) and incentive compatibility (the Agent must optimally choose the strategy intended by the Principal).

The model distinguishes between two distinct mathematical regimes based on the support of the default time $\tau$ :

Unbounded Case (Hypothesis A): The maximum possible default time exceeds the investment horizon $T$ (or is infinite). Default may or may not occur within $[0, T]$ .
Bounded Case (Hypothesis B): The default time is guaranteed to occur within $[0, T]$ . This introduces a singularity in the intensity process as $t \to T$ .

2. Methodology

Stochastic Framework and Filtration Enlargement

The authors employ the theory of filtration enlargement. The market filtration $\mathbb{F}$ (generated by Brownian motion $W$ ) is enlarged to $\mathbb{G}$ to include information about the default time $\tau$ . Under the Density Hypothesis, the default time admits a conditional density, allowing the transfer of the martingale property from $\mathbb{F}$ to $\mathbb{G}$ (the immersion property). The default process $H_t = \mathbb{1}_{\tau \leq t}$ is compensated by a predictable intensity process $\lambda_t$ .

Contract Structure and Reduction

Following the approach of [17] and subsequent literature, the authors define a class of admissible contracts $\Xi$ that are contractible with respect to the observable processes: the asset prices $S$ , the portfolio wealth $X$ , and the default indicator $H$ .

The contract $\xi$ is expressed as a stochastic integral involving processes $Z$ (sensitivity to assets), $Z^X$ (sensitivity to wealth), $U$ (sensitivity to default), and $\Gamma$ terms (sensitivity to quadratic variations).
Using Backward Stochastic Differential Equations (BSDEs), the authors reduce the Agent's optimization problem to finding the optimal control $\hat{\pi}$ that maximizes a specific driver function $F$ .
The Principal's problem is then transformed into a stochastic control problem where the control variables are the processes defining the contract ( $Z, Z^X, \Gamma, U$ ).

Hamilton-Jacobi-Bellman (HJB) Formulation

The Principal's optimization problem is characterized by Integro-Partial Differential Equations (IPDEs) of the Hamilton-Jacobi-Bellman type:

Bounded Case (bHJB): The equation includes a source term weighted by the default density $f(t)$ and a terminal condition of zero. The intensity $\lambda_t$ becomes singular as $t \to T$ , leading to a degenerate driver.
Unbounded Case (uHJB): The equation includes a terminal condition weighted by the survival probability $(1 - F_\tau(T))$ .

The authors prove Verification Theorems for both cases, establishing that if a sufficiently smooth solution to the HJB equation exists, it yields the optimal contract and value function. They also establish that the value function is a viscosity solution to the respective PDEs, relaxing the regularity requirements.

Numerical Solution via Deep Learning

Due to the high dimensionality and the non-linear coupling between the value function and the optimal controls (which depend on the derivatives of the value function), the authors develop a Deep Learning algorithm based on Physics-Informed Neural Networks (PINNs).

Actor-Critic Architecture: The algorithm alternates between two steps:
1. Critic (PDE Solver): Given fixed controls, a neural network approximates the value function $v$ by minimizing the residual of the HJB equation (using automatic differentiation).
2. Actor (Optimizer): Given the updated value function, the algorithm solves the pointwise maximization problem to update the control processes ( $Z^X, \Gamma^X, U$ ).
Stabilization: To handle the fixed-point nature of the problem and potential oscillations, the authors employ exponential moving averages for the control updates and self-aware loss weighting.

3. Key Contributions

Novel PA Framework with Random Horizon: The paper extends the continuous-time Principal-Agent literature by explicitly modeling the investment horizon as a random variable driven by a default time, distinguishing between bounded and unbounded scenarios.
Mathematical Unification: It provides a unified theoretical framework that handles both the unbounded case (standard BSDEs) and the bounded case (degenerate BSDEs with singular drivers), bridging the gap between "black swan" market crashes and life-insurance type horizons.
Explicit Contract Characterization: The authors derive the explicit form of the optimal investment strategy and the structure of the optimal contract, showing that the contract must include incentives based on asset prices, portfolio wealth, quadratic variations, and the default event itself.
Numerical Methodology: The paper introduces a specialized Actor-Critic deep learning scheme to solve the coupled integro-partial HJB equations where the coefficients depend on the solution itself, a challenge that precludes standard numerical solvers.

4. Results and Numerical Findings

The numerical experiments, conducted on a single risky asset with exponential utilities, reveal several qualitative insights:

Impact of Default Distribution:
- Unbounded (Exponential) vs. Bounded (Uniform): Under an unbounded default (high initial hazard rate), the optimal strategy is initially more aggressive to capture value before a potential early default, then becomes conservative. In the bounded case, the strategy is smoother but declines sharply as the terminal date approaches.
- Incentive Structure: In the bounded case, both portfolio-value incentives ( $Z^X$ ) and default-risk incentives ( $U$ ) are systematically higher. The Principal must offer stronger incentives to align the Agent's actions when default is guaranteed to occur within the horizon.
Risk Aversion Alignment:
- When the Agent is less risk-averse than the Principal, the portfolio value increases due to more aggressive investment.
- When the Agent is more risk-averse, the portfolio value decreases. The compensation $Z^X$ and the default insurance $U$ adjust to reflect the Agent's risk profile.
Skewness of Default Time:
- Right-skewed (Early Default): The Agent adopts an extremely aggressive strategy early on, knowing the horizon is likely short. The Principal offers high immediate growth incentives and default insurance.
- Left-skewed (Late Default): The Agent delays aggressive investment. The Principal must continuously increase incentives to motivate the Agent as the probability of default rises in the second half of the horizon. The default insurance component becomes significantly larger to compensate for the lost opportunity of a late default.
Linear vs. General Contracts:
- Restricting contracts to a linear form (a fixed percentage of terminal wealth) results in a significant optimality gap compared to the general optimal contract.
- Linear contracts lead to more conservative strategies. The optimal general contract dynamically balances growth incentives and default insurance, adapting to market conditions and the specific timing of the default risk.

5. Significance and Claims

The paper claims to fill a critical gap in the literature regarding investment horizon uncertainty in delegated portfolio management. By incorporating random default times, the model offers a more realistic representation of financial scenarios such as cryptocurrency markets (prone to sudden termination) and life-insurance portfolios (terminated by death).

The authors modestly note that while the theoretical framework is rigorous, the uniqueness of viscosity solutions for the degenerate bounded case (where the intensity explodes) remains an open theoretical challenge requiring further analysis of second-order BSDEs. However, the proposed numerical approach successfully demonstrates the feasibility of solving these complex, coupled systems and provides actionable insights into how default risk shapes optimal compensation and trading strategies. The work bridges the gap between theoretical stochastic control and practical financial engineering by offering a computational tool to analyze these high-dimensional, non-linear problems.

Delegated portfolio management with random default