← Latest papers
⚛️ quantum physics

Insurance portfolio simulation with a quantum circuit

This paper proposes a Quantum Circuit Born Machine (QCBM) approach for insurance portfolio simulation that utilizes tail-focused discretization, a Monte Carlo-style objective, and a two-stage hierarchical training scheme to effectively model complex joint loss distributions and accurately capture tail risks on quantum hardware.

Original author: William Cashman, Marco Caselli, Saad Hamid, Daniel Mills, Pavel Stishenko

Published 2026-11-20
📖 6 min read🧠 Deep dive

Original author: William Cashman, Marco Caselli, Saad Hamid, Daniel Mills, Pavel Stishenko

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the author. For technical accuracy, refer to the original paper. Read full disclaimer

Insurance companies operate on a simple premise: they collect small, steady payments from many people to cover the rare, catastrophic events that might strike a few. To stay solvent, these firms must understand how different risks behave together. If a hurricane hits, it might damage homes and cars simultaneously, causing a sudden surge in claims. At the same time, the same storm might depress the value of the stocks and bonds the insurer holds as investments. This double blow can threaten the company's stability. To prepare, actuaries—specialists who calculate risk—try to build mathematical models that describe how these different losses and asset values move in relation to one another. The challenge is that real-world data is messy. Extreme events are rare, meaning there is very little data to study, yet these are the moments that matter most. Traditional methods often struggle to capture these rare, high-stakes scenarios, especially when the data is limited.

A team of researchers has now explored a different path, using a quantum computer to model these complex relationships. They focused on a specific type of problem: fitting a joint distribution to a mix of insurance losses and stock market returns. In this context, a joint distribution is simply a map that shows how likely it is for different variables to take on specific values at the same time. The researchers wanted to see if a quantum circuit—a program running on a quantum processor—could learn this map more effectively than standard methods, particularly in the "tails" of the data, which represent the most extreme and dangerous outcomes. Their work suggests that by combining a clever way of organizing the data with a specialized training method, quantum machines can indeed learn these intricate patterns, even when the data is sparse.

The researchers began with a dataset containing weekly records of stock market losses and three different types of insurance claims. They knew that standard ways of feeding this data into a computer would fail to capture the worst-case scenarios. If you divide a range of numbers into equal-sized buckets, the buckets covering the extreme ends are just as wide as the ones in the middle. This means the computer has very little room to distinguish between a moderately bad loss and a truly catastrophic one. To solve this, the team designed a system that packed more detail into the extreme ends. They assigned the first half of all observations to the first bucket, the next quarter to the second, the next eighth to the third, and so on. This exponential approach meant that the final bucket, which held the most extreme losses, was incredibly narrow, allowing the model to see the fine details of the worst events.

Once the data was organized, the team faced another hurdle: how to teach the quantum computer to recognize the patterns. Usually, a machine learning model is trained by comparing its guesses to the actual data and adjusting its settings to reduce the difference. However, with so few data points and so many possible outcomes, the standard way of measuring this difference becomes unreliable and noisy. The researchers developed a new method to smooth out this noise. Instead of trying to compare every single bucket of data at once, they randomly grouped the buckets into larger, coarser groups and compared those. By repeating this random grouping many times and averaging the results, they created a stable signal that told the computer which direction to move its settings to improve. This allowed the training to proceed smoothly even with limited data.

The training process itself was built in two stages to ensure the computer learned efficiently. First, the team trained a smaller version of the model on a classical computer simulator. This smaller model used fewer "qubits," the basic units of information in a quantum computer. Once this smaller model learned the general shape of the risk distribution, the researchers transferred its knowledge to a larger, more detailed model. They then fine-tuned only the new parts of the larger model on actual quantum hardware. This approach, known as hierarchical learning, acted like a warm start, giving the complex machine a head start rather than forcing it to learn from scratch. The hardware they used was a trapped-ion quantum computer, which uses individual atoms held in place by electromagnetic fields to perform calculations.

The results showed that this approach worked. The quantum model successfully learned the relationships between the different types of risks. It correctly identified that the different insurance lines tended to move together during bad times, while the stock market returns remained largely independent of the insurance losses. When the team looked at the total potential loss of the portfolio, the model trained with the exponential encoding was able to reach the extreme values seen in the real data. In contrast, a model using the standard, uniform method could not produce values high enough to match the worst observed losses; it simply hit a ceiling far below reality. The exponential model, however, could represent losses that were hundreds of times larger than the median, capturing the true danger of the tail.

There were, of course, limitations. The model trained on the quantum hardware did not perform quite as well as the one simulated on a classical computer. The researchers found that this gap was not primarily due to errors in the quantum machine itself, but rather because they could only run the machine a limited number of times to gather data. The quantum computer needs to be run many times to build a clear picture of the probabilities, and with the current technology, they could not run it enough times to match the precision of the simulation. Nevertheless, the experiment proved that the method is viable. The hierarchical training scheme significantly outperformed starting from scratch, reducing the error in the final model by more than half compared to random attempts.

The study concludes that quantum circuits can be adapted to handle the heavy-tailed, sparse data typical of insurance portfolios, but only if the entire process is designed with these specific challenges in mind. Standard methods developed for smoother, more common data distributions do not work directly. The success of this project relied on three key innovations: a way to encode the data that focuses on the extremes, a training objective that remains stable with few samples, and a two-step learning process that builds complexity gradually. While the researchers did not claim to have solved the problem of risk modeling entirely, they demonstrated a working pipeline that moves from data to a quantum-generated risk model. They noted that for this to become a practical tool for the industry, future work will need to address the number of times the hardware must be run and test the method on even larger portfolios. For now, the work stands as a proof that quantum machines can begin to map the shape of rare and dangerous financial events.

Drowning in papers in your field?

Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.

Try Digest →