Financial institutions face a familiar constraint in an unfamiliar context: computational limits that are becoming a strategic bottleneck.
From insurers optimizing capital allocation to banks running value adjustments and liquidity simulations, many firms rely on the tried-and-true Monte Carlo simulations that have served the sector well for decades. But as portfolios become more dense, complex, and interconnected, institutions are running headlong into the limitations of brute-force simulations. Runtimes and costs are up and, more important, solution quality is suffering.
The biggest financial institutions have invested in a potentially powerful solution: quantum computing, which is designed to handle exactly the kind of complex problems that large-scale simulations present. But quantum computers are still in development. The primary focus has been on hardware, but software technology has quietly advanced to the stage where a new class of computing approaches known as quantum-inspired algorithms (QIAs) can address the needs of financial firms. And they do so running on hardware—CPUs and GPUs—that these institutions already have. This makes QIAs relatively simple and inexpensive (nothing in tech is ever truly easy or cheap) for most banks and insurance companies to deploy.
The Limits of Monte Carlo at Scale
The strength of Monte Carlo simulations lies in their simplicity: sample widely, evaluate outcomes, and identify optimal tradeoffs between risk and return. But this strength becomes a weakness at scale.
Modern financial portfolios have grown in size and complexity much faster than Monte Carlo’s ability to keep up. As banks and insurers move toward more in-depth portfolio steering, by line of business, geography, client segment, and risk type, the combinatorial challenge expands exponentially. Monte Carlo runs now take hours, consume significant computing power, and still fail to find the most attractive risk-return combinations.
In a controlled experiment, we tested current Monte Carlo capabilities with those involving QIAs using a synthetic portfolio with more than 1,000 segments, each with defined expected return, risk, and cross-segment correlations. The QIA-enabled simulations run on classical computers consistently outperformed their brute-force counterparts on the key metrics of speed and portfolio quality.
Tech + Us: Harness the power of technology and AI
How QIAs Work
QIAs are not quantum computing. (See the exhibit.) They are classical algorithms designed to replicate certain features of quantum systems, such as navigating complex “energy landscapes” or representing high-dimensional states more efficiently. They run on standard CPUs and GPUs. No new hardware is required.
We tested three approaches to QIAs, each representing a fundamentally different way of navigating complex optimization problems compared with random sampling:
- Simulated annealing (SA) is a probabilistic search method algorithm that begins by accepting a wide range of portfolio moves, including unfavorable ones, then progressively narrows its tolerance as the search matures. This structure allows SA to sidestep local optimal solutions that trap greedy or random-sampling methods and converge toward the best global solution.
- Quantum Hamiltonian descent (QHD) is a gradient-based method grounded in quantum energy dynamics. Portfolio objectives are translated into an energy function, and the algorithm evolves toward minimum-energy (maximum quality) allocations using momentum-guided descent. This structure tends to converge faster than SA on well-conditioned problems.
- Tensor networks represent a more nascent and more computationally intensive approach than either SA or QHD. The networks compress high-dimensional decisions with many variables into efficient algebraic structures, enabling the algorithm to reason across many possible states simultaneously. In our early GPU-based experiments, tensor networks demonstrated strong exploration efficiency and favorable scaling properties, suggesting significant potential as implementations mature. Realizing this potential, however, requires more specialized mathematical and engineering expertise, so companies should expect to make greater upfront investment in skills and implementation before tensor networks can scale smoothly.
Speed and Quality
To compare these methods with each other and with traditional Monte Carlo simulations, we constructed a benchmark portfolio optimization problem that reflects a calculation challenge typical of those faced by insurers, banks, asset managers, and other financial institutions. Our base portfolio contained approximately 600 active segments with the option to incorporate around 400 more. We imposed practical constraints that capped allocation per segment at 2% of the total portfolio and imposed limits on how much the total allocation could shift within a single rebalancing step.
We then ran each simulation method across multiple configurations and hardware options: Monte Carlo on CPUs with up to millions of sampled portfolios, SA and QHD on CPUs, and tensor networks on GPUs. For each run we tracked the runtime to a stable solution and assessed portfolio quality by the set of portfolios maximizing returns at each level of risk (known as the Pareto-efficient frontier).
Across multiple configurations, the results were consistent—and strategically significant. Both SA and QHD consistently outperformed Monte Carlo along the Pareto-efficient frontier. At equivalent levels of risk, they identified portfolios with higher expected returns. At equivalent return targets, they achieved lower risk—sometimes reducing volatility by several percentage points. These are not marginal improvements; they represent a shift in the attainable frontier.
QIA-enabled simulations also win more quickly. While Monte Carlo simulations required tens of minutes to hours to do their work, SA reached comparable or better solutions in minutes. QHD often converged in under a minute for similar problem sizes. Tensor network approaches, while still maturing, showed strong scaling potential on GPU infrastructure. Tensor networks also demonstrated significantly higher exploration efficiency, identifying more feasible solutions per unit of computing capacity. This improves the likelihood of finding superior outcomes as problem complexity grows.
Practical Benefits
For financial institutions, these improvements translate directly into strategic advantage.
Superior Capital Allocation. QIAs consistently reach regions of the risk-return frontier that Monte Carlo does not, which translates directly to better capital deployment, whether the goal is higher return on equity at constant risk or lower volatility at constant return.
Faster Decision Cycles. Optimization processes that once ran overnight can now be executed in near real-time. This opens the door to more frequent rebalancing, richer stress testing, and faster responses to market dislocations or regulatory changes.
No New Infrastructure Required. Because QIAs run on existing infrastructure, adoption does not depend on uncertain quantum computing timelines. QIAs are available to deploy now.
Future-Proof Scalability. As GPU utilization deepens and tensor networks mature, the performance gap is likely to widen—particularly for large, complex portfolios.
A Pragmatic Path to Adoption
Adopting QIAs does not require a wholesale transformation. A staged approach is both feasible and advisable.
Phase 1. Introduce challenger models. Deploy QIAs alongside existing Monte Carlo frameworks on selected high-complexity portfolios. Build internal evidence on performance.
Phase 2. Focus on bottleneck use cases. Prioritize areas where Monte Carlo is already under strain: large portfolios, tight constraints, frequent re-optimizations, or computationally intensive trading books.
Phase 3. Strengthen governance. Subject outputs to rigorous validation: back-testing, stress scenarios, and model risk review. Establish clear performance benchmarks, mandate transparency, and perform regular audits.
Phase 4. Scale into production. Integrate validated approaches into core processes such as strategic asset allocation, capital management, and risk analytics. Expand GPU infrastructure where tensor networks show the most incremental value.
The Bottom Line
Monte Carlo simulations will remain part of the optimization toolkit for years, particularly in cases where regulatory transparency and methodological familiarity are priorities. But they need not remain the ceiling.
QIAs are a practical, proven enhancement to today’s optimization capabilities. The question for the financial industry is not whether QIAs work, it is how quickly they can be moved from benchmark to production. For institutions willing to act early, the opportunity is clear: better decisions, delivered faster, using infrastructure already in place.
