Bayesian optimal investment and reinsurance with dependent financial and insurance risks

Basic Information

• Paper ID: 2103.05777
• Title: Bayesian optimal investment and reinsurance with dependent financial and insurance risks
• Authors: Nicole Bäuerle, Gregor Leimcke (Karlsruhe Institute of Technology)
• Classification: q-fin.PM (Portfolio Management), math.PR (Probability Theory)
• Publication Date: March 2021 (arXiv preprint)
• Paper Link: https://arxiv.org/abs/2103.05777

Abstract

Major events such as natural disasters or the COVID-19 crisis impact both financial markets and the claim arrival intensity and claim sizes of insurance companies. Therefore, when determining optimal investment and reinsurance strategies, it is crucial to consider models that reflect such dependencies. This paper proposes a method for generating dependencies between financial markets and claim sizes during crisis periods, and determines optimal investment and reinsurance strategies that maximize the expected exponential utility of terminal wealth through stochastic control methods. Furthermore, the model allows for learning the claim size distribution. Comparisons and bounds on optimal strategies are provided through simplified models. Surprisingly, numerical results demonstrate that even minimal dependencies created in the model have enormous impacts on control, making insurance companies more cautious.

Research Background and Motivation

Problem Background

1. Practical Challenges: Major events such as the COVID-19 pandemic simultaneously affect financial markets and insurance operations, demonstrating the interdependence between financial and insurance risks
2. Theoretical Gaps: While traditional actuarial literature models dependencies among different business lines, few studies connect financial market evolution with claim occurrence
3. Practical Needs: Insurance companies need to formulate optimal investment and reinsurance strategies under incomplete information

Research Motivation

• Dependency Modeling: Propose new methods to model dependencies between financial markets and insurance operations
• Learning Mechanisms: Enable insurance companies to learn claim size distributions from observed data
• Risk Management: Optimize investment and reinsurance decisions while accounting for dependencies

Core Contributions

1. Innovation in Dependency Modeling: Proposes a novel method for modeling dependencies between financial markets and insurance risks through a threshold mechanism—when claim sizes exceed threshold L, risky asset prices experience stochastic downturns
2. Bayesian Learning Framework: Introduces a Bayesian approach that allows learning of claim size distributions, addressing incomplete information issues
3. Optimal Control Solutions: Characterizes optimal investment and reinsurance strategies through generalized Hamilton-Jacobi-Bellman equations
4. Comparative Results: Provides comparative analysis and bounds between dependent and independent models
5. Numerical Findings: Reveals the important phenomenon that even minimal dependencies have significant impacts on optimal strategies

Methodology Details

Problem Formulation

Objective: Maximize expected exponential utility of terminal wealth $U(x) = -e^{-\alpha x}, \quad \alpha > 0$

Control Variables:

• Investment strategy $\xi = (\xi_t)_{t \geq 0}$: Amount invested in risky assets
• Reinsurance strategy $b = (b_t)_{t \geq 0}$: Retention level, $b_t \in [0,1]$

Model Architecture

1. Claim Process

• Claim Arrivals: Poisson process $N_t$ with intensity $\lambda > 0$
• Claim Sizes: Sequence $(Y_n)_{n \in \mathbb{N}}$ with unknown distribution requiring learning
• Aggregate Claims: $S_t = \sum_{i=1}^{N_t} Y_i$

2. Financial Market

• Risk-free Asset: $dB_t = rB_t dt$
• Risky Asset: $dP_t = P_{t-}\left(\mu dt + \sigma dW_t - \int_E z\mathbf{1}_{(L,\infty)}(y)\Psi(dt,d(y,z))\right)$

where the risky asset price drops by proportion $Z_n$ when claim size $Y_n > L$.

3. Wealth Process

$dX_t^{\xi,b} = \left(rX_t^{\xi,b} + (\mu-r)\xi_t + c(b_t)\right)dt + \xi_t\sigma dW_t - \int_E (b_t y + \xi_t z\mathbf{1}_{(L,\infty)}(y))\Psi(dt,d(y,z))$

Bayesian Learning Mechanism

Filtering Process

Posterior probability of claim size distribution: $p_j(t) = P(\vartheta = j|\mathcal{F}_t^\Psi), \quad j = 1,\ldots,m$

Through Bayesian update: $p_{T_n} = J(p_{T_n^-}, Y_n)$ where $J(p,y) = \left(\frac{f_1(y)p_1}{\sum_{k=1}^m f_k(y)p_k}, \ldots, \frac{f_m(y)p_m}{\sum_{k=1}^m f_k(y)p_k}\right)$

Optimal Control Solution

HJB Equation

Through variable separation $V(t,x,p) = -e^{-\alpha x}e^{r(T-t)}g(t,p)$, the generalized HJB equation is obtained: $0 = \inf_{(\xi,b) \in [-K,K] \times [0,1]} \{L^g(t,p;\xi,b)\} + \inf_{\phi \in \partial^C g_p(t)} \{\phi\}$

where $L^g$ is the corresponding operator and $\partial^C g_p(t)$ is the Clarke generalized subdifferential.

Optimal Strategy

The optimal strategy $(ξ^*, b^*)$ is determined through first-order conditions: $v_1(t,p,\xi,b) = \mu - r$$v_2(t,p,\xi,b) = (1+\theta)\kappa$

Experimental Setup

Parameter Configuration

• Financial Market Parameters: $\sigma = 0.4$, $\mu = 0.3$, $r = 0$
• Insurance Parameters: $\lambda = 10$, claim sizes follow exponential distribution $Y \sim \text{Exp}(0.1)$
• Utility Parameters: $\alpha = 0.05$ (approaching risk-sensitive case)
• Reinsurance Parameters: $(1+\theta)\kappa = 350$

Numerical Experiments

Investigate the impact of different thresholds $L$ on optimal strategies, where $L$ varies from near 0 to 1000.

Experimental Results

Main Findings

1. Investment Strategy Sensitivity to Dependency:
   • Optimal investment in independent case: $\xi^* = 37.5$
   • When $L = 100$ (threshold is 10 times expected claim): optimal investment is only 25.19
   • Even with high thresholds (occurrence probability only 4.5×10⁻⁵), investment strategy remains significantly conservative
2. Threshold Effects:
   • Negative investment (short selling for hedging) occurs when $L < 64.35$
   • As $L \to 0$, optimal investment converges to $-86.57$
   • As $L \to \infty$, reinsurance level stabilizes at approximately 0.93

Comparative Results

Theorem 5.1 (Comparison with Independent Case)

For any $t \in [0,T]$, we have $\xi_t^* \leq \tilde{\xi}_t^*$, meaning investment is more conservative with dependencies.

Theorem 5.4 (Comparison with Complete Information)

Under the same retention level, $\xi_t^* \leq \xi_{F_{\bar{p}_{t^-}}}^*(t)$, meaning investment is more conservative under incomplete information.

Related Work

Dependency Modeling Methods

1. Common Shock Models: Generate joint claims through additional Poisson processes
2. Lévy Copulas: Mathematically elegant but less interpretable
3. Interacting Intensities: Model dependencies through intensity interactions

Learning and Control

• Hidden Markov Models: Latent factors influence model parameters
• Bayesian Methods: Learn from finite distribution sets
• Partial Information Control: Handle unobservable parameters

Conclusions and Discussion

Main Conclusions

1. Enormous Impact of Minimal Dependency: Even minimal dependencies have significant impacts on optimal strategies
2. Enhanced Conservatism: Dependencies lead insurance companies to adopt more conservative investment strategies
3. Hedging Mechanisms: Negative investments can hedge claim risks

Limitations

1. Model Simplification: Considers only one business line; actual scenarios may involve multiple related lines
2. Dependency Form: Considers only downside effects; actual scenarios may involve bidirectional impacts
3. Parameter Constraints: Exponential utility parameters must satisfy specific conditions to ensure integrability

Future Directions

1. Multi-line Extension: Extend the model to multiple interdependent business lines
2. More Complex Dependency Structures: Consider more general dependency relationships
3. Empirical Validation: Validate model effectiveness using real data

In-Depth Evaluation

Strengths

1. Theoretical Innovation: Proposes novel dependency modeling methods with clear economic interpretation
2. Mathematical Rigor: Uses Clarke generalized subdifferentials to handle non-differentiable cases with rigorous theoretical analysis
3. Practical Value: Reveals important impacts of dependencies on investment decisions with guidance for practice
4. Numerical Insights: Discovers surprising phenomena—enormous impacts of minimal dependencies

Weaknesses

1. Model Limitations: Relatively simple dependency modeling, considering only unidirectional effects
2. Empirical Gaps: Lacks validation and calibration with real data
3. Computational Complexity: Numerical solution process is complex, potentially challenging for practical applications

Impact

1. Academic Contribution: Provides new perspectives for joint modeling of insurance and financial risks
2. Practical Guidance: Offers important insights for insurance company risk management strategies
3. Methodological Value: Clarke subdifferential application provides technical reference for similar problems

Applicable Scenarios

• Investment decisions for insurance companies facing systemic risks
• Regulatory policy formulation considering correlations between financial and insurance risks
• Joint pricing of catastrophe insurance and financial derivatives

References

The paper cites 26 related references, primarily including:

• Classical literature on risk theory and stochastic control
• Related research on dependency modeling
• Important works on Bayesian learning and partial information control
• Recent research on COVID-19 impacts on insurance and financial markets