Optimal Time to Change Premiums

Erhan Bayraktar, H. Vincent Poor

Abstract

The claim arrival process to an insurance company is modeled by a compound Poisson process whose intensity and/or jump size distribution changes at an unobservable time with a known distribution. It is in the insurance company's interest to detect the change time as soon as possible in order to re-evaluate a new fair value for premiums to keep its profit level the same. This is equivalent to a problem in which the intensity and the jump size change at the same time but the intensity changes to a random variable with a know distribution. This problem becomes an optimal stopping problem for a Markovian sufficient statistic. Here, a special case of this problem is solved, in which the rate of the arrivals moves up to one of two possible values, and the Markovian sufficient statistic is two-dimensional.

Introduction

In insurance risk theory, premiums are calculated based on the expected frequency and severity of claims. However, the underlying environment is rarely static. Insurance companies often face "regime shifts"—sudden, unobservable changes in claim intensity or size distribution caused by external factors. If a company reacts too slowly to an increase in risk, it faces insolvency or reduced profit; if it reacts too quickly based on a false alarm, it risks losing customers due to unnecessary premium hikes.

This paper develops a mathematical framework to solve the "quickest detection" problem: identifying the exact moment a regime shift occurs based only on the observed history of claims. By framing this as an optimal stopping problem, the authors provide a method to balance the risk of false alarms against the cost of detection delay.

Theoretical Framework

The authors model the claim arrival process as a compound Poisson process. At an unobservable time $\theta$, the process undergoes a "disorder," where the intensity $\lambda$ and the jump size distribution $\nu$ change. To make the decision-making process mathematically tractable, the authors employ several advanced stochastic tools:

  • Optimal Stopping Theory: The goal is to find a stopping time $\tau$ that minimizes a cost function $R_\tau(\pi)$, which represents the weighted sum of the probability of a false alarm ($P(\tau < \theta)$) and the expected delay ($E[(\tau - \theta)^+]$).
  • Markovian Sufficient Statistics: The entire history of claim observations is compressed into a "sufficient statistic." This compressed state contains all necessary information to determine whether the optimal stopping boundary has been reached.
  • Change of Measure: Using Girsanov’s Theorem, the authors relate the complex real-world model to a simpler reference probability measure, allowing for the derivation of the Radon-Nikodym derivative used in the decision statistic.

The Two-Dimensional Case

The paper focuses on a specific, complex scenario where the post-change arrival rate can shift to one of two possible values (a Bernoulli distribution). This results in a two-dimensional Markovian sufficient statistic $(\Phi^{(0)}, \Phi^{(1)})$.

The evolution of this statistic is characterized by two distinct behaviors:

  1. Deterministic Drift: Between claim arrivals, the statistic moves according to ordinary differential equations (ODEs). This highlights the "informativeness of silence"—the absence of claims is itself a data point that suggests the regime may not have changed yet.
  2. Stochastic Jumps: Whenever a claim occurs, the statistic "jumps" to a new position in the state space, reflecting the immediate increase in the likelihood of a high-intensity regime.

Key Insights

The Geometry of Decision Making

The authors prove that the optimal stopping region is defined by decreasing, convex curves within the two-dimensional state space. Once the sufficient statistic crosses this "free boundary," the insurer should immediately adjust premiums.

Counter-Intuitive Alarms

A significant finding is that it is not always optimal to trigger an alarm at the exact moment a claim arrives. Depending on the specific penalties for delay and false alarms, the drift of the sufficient statistic between claims can hit the free boundary, meaning the optimal time to change premiums can occur during a period of silence.

Results and Implementation

The research provides several key mathematical guarantees:

  • Existence: An optimal stopping time is proven to exist and is finite almost surely.
  • Iterative Solutions: For scenarios where an explicit solution is unavailable, the authors characterize the "value function" (the minimum possible cost) as the fixed point of a functional operator. This allows for an iterative algorithm that converges exponentially fast to the optimal solution.
  • High-Risk Parameters: The paper identifies specific parameter ranges—such as very high penalties for delay—where the optimal stopping rule can be identified exactly.

Conclusion and Future Directions

While this framework provides a robust solution for two-dimensional shifts, the authors acknowledge the "curse of dimensionality" as the number of possible post-change rates increases. Furthermore, the model currently assumes a known prior distribution for the change time and known characteristics for the new regime. Future research may explore adaptive detection models where the insurer must learn the characteristics of the new regime while simultaneously attempting to detect its onset.

Summary updated

Abstract

The claim arrival process to an insurance company is modeled by a compound Poisson process whose intensity and/or jump size distribution changes at an unobservable time with a known distribution. It is in the insurance company's interest to detect the change time as soon as possible in order to re-evaluate a new fair value for premiums to keep its profit level the same. This is equivalent to a problem in which the intensity and the jump size change at the same time but the intensity changes to a random variable with a know distribution. This problem becomes an optimal stopping problem for a Markovian sufficient statistic. Here, a special case of this problem is solved, in which the rate of the arrivals moves up to one of two possible values, and the Markovian sufficient statistic is two-dimensional.
BibTeX
@article{bayraktar2008optimal,
  author = {Erhan Bayraktar and H. Vincent Poor},
  title = {{Optimal Time to Change Premiums}},
  journal = {Math. Methods Oper. Res.},
  year = {2008},
  doi = {10.1007/s00186-007-0182-9},
}
Figure