Allstate and IBM are exploring whether quantum computing can help insurers build portfolios that are more balanced between value and risk by addressing a complex version of the “knapsack problem” known in computer science. The results of the joint work were published on the arXiv platform in May 2026, while IBM presented details of it on June 18, 2026.
The knapsack problem involves selecting a group of items that achieves the highest possible value without exceeding a specific weight or capacity limit. In its insurance application, insurance policies represent the items to be selected for a portfolio, potential losses represent the weight, and the portfolio’s value represents the expected benefit or return. The problem becomes difficult when the number of items is large and when the weights are not fixed numbers but uncertain ranges.
Why Are Home Insurance Portfolios Different?
Not all risks move in the same way. A car accident may be largely independent of an accident involving another driver, but weather-related hazards can affect a large number of policies at once. Wildfires, hurricanes, and snowstorms can damage many homes in one area, making losses correlated rather than separate events.
Eric Holz, Allstate’s chief analytics and data officer, said that home insurance requires thinking from a portfolio perspective, not only from the perspective of individual risks. It is therefore not enough to assess each property separately and then add the results; the insurer must estimate what may happen when multiple losses occur at the same time and ensure that severe scenarios remain within the level of risk the company accepts.
Allstate currently relies on classical simulation to estimate potential outcomes. The company may run 100,000 future scenarios, but rare and costly events remain a challenge. If the focus is on the worst 1% of these scenarios, only about 1,000 events remain, which may not provide sufficient certainty when analyzing multiple types of hazards across broad geographic regions.
An Approach Combining Quantum and Classical Computing
The team tested a quantum approach tailored to home insurance portfolios. The approach uses a quantum circuit running on an IBM Quantum Heron processor to generate a set of candidate solutions, directing the computation toward combinations that balance value with compliance with the budget or acceptable loss limit.
A classical phase then refines the results. It repairs solutions that exceed the budget, learns which homes most often appear in good solutions, and uses this knowledge to guide the next round of computation. The process therefore repeats in a cycle intended to improve the results with each pass.
The researchers also faced a known problem when training quantum circuits on large problems, as the learning signal can weaken as the problem grows. To overcome this, the team trained the circuit on a small version of the problem and then transferred what it learned to a larger version.
Vaibhav Kumar, an IBM quantum-computing researcher and one of the paper’s authors, explained that current classical methods either rely on simulating multiple scenarios, which is accurate but computationally expensive, or focus on the worst possible range, which is safer but may be highly conservative.
Comparison with Traditional Optimization Methods
To measure the performance of the new approach, the team compared its results with an exact solution capable, if given enough time, of reaching the proven best answer. It also compared the approach with four common approximate methods that balance speed and accuracy:
- Simulated annealing.
- Tabu search.
- Parallel tempering.
- Genetic algorithm.
Each method was given 30 minutes per problem. In problems containing up to 75 items, all the methods reached the proven best result. As the problem sizes increased, the quantum-classical approach remained competitive with the strongest classical heuristic methods, coming close to their performance and marginally outperforming them under strict constraints.
However, the approach cannot yet handle the sizes at which exact solutions become so slow that they lose their usefulness. The results therefore represent an early indication of the possibility of delivering practical value in the future, not evidence that quantum computing has already solved the problem of building insurance portfolios at broad commercial scale.
What Is Allstate Learning from the Experiment?
Allstate sees the goal as building a bridge from two directions: understanding what quantum devices can currently do, identifying business problems that may benefit from them as the devices improve, and then determining where technical capabilities meet practical needs.
Holz said that uncertainty about the timeline for achieving commercial value does not mean waiting, because companies that build their skills and identify suitable problems early may be better prepared when the technology advances. Both Holz and Jean Otte, a data scientist and technical director in Allstate’s insurance products organization, emphasized the importance of starting with a real problem that matters to the business rather than studying the technology separately from its uses.
The partnership began through IBM’s Quantum Accelerator program for organizations new to the field and then developed into joint research lasting two years. According to Otte, the work was characterized by transparency through the publication of the results, code, and data, enabling others to examine and build on them.
Open questions remain about the role that quantum workflows will play in insurers’ decisions and portfolio construction. But the experiment gave Allstate practical experience in formulating the problem, selecting suitable tasks, and integrating quantum processing with classical methods. As hardware improves and noise decreases, IBM expects the portion of the work that must be performed on the classical side to decline.