💰 quantitative finance

Landscape-Dependent Performance of Photonic Quantum Solvers in QUBO Feature Selection for Financial Risk Detection

This paper benchmarks classical, photonic, and simulated photonic solvers for feature selection in financial risk detection, revealing that while photonic systems like QCI Dirac-3 can match or exceed classical performance on specific tasks, the choice of feature selection method and the concentration of predictive signal in the dataset are more critical determinants of success than the underlying computing paradigm.

Original authors: Nirvik Sahoo, Paul Robert Griffin

Published 2026-10-05

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Nirvik Sahoo, Paul Robert Griffin

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 authors. For technical accuracy, refer to the original paper. Read full disclaimer

In the high-stakes world of finance, banks and credit card companies constantly battle a silent war against fraud and loan defaults. To win, they rely on massive amounts of data, tracking everything from how much a customer spends to how often they pay their bills. However, having too much data can be a problem. When a computer model is fed thousands of potential clues, it often gets confused, picking up on noise rather than the real signal, much like a person trying to hear a single conversation in a crowded stadium. The solution is to find the specific, most important clues and ignore the rest. This process is called feature selection. For decades, experts have used standard mathematical tools to sift through this data, but a new wave of technology promises to do the job faster and differently. This technology includes photonic computers, which use particles of light instead of electricity to solve complex puzzles, and quantum-inspired machines that explore many possibilities at once. The question researchers have been asking is whether these new, exotic machines can actually find better clues than the old, reliable methods, especially when the data looks very different from one case to another.

A team of researchers at Singapore Management University set out to answer this by pitting three different computing approaches against each other in a real-world test. They chose two very different financial datasets to see if the results would hold up across the board. The first dataset involved nearly 285,000 credit card transactions from Europe, where fraud was rare, occurring in less than two out of every thousand cases. The second dataset was much larger, containing information on over 150 different financial indicators for consumers who might default on their loans. The researchers tested thirteen different ways to select the best clues, running them through a traditional, highly precise classical computer, a photonic entropy computer that uses light to find solutions, and a photonic boson sampler that uses light to randomly pick combinations based on complex mathematical rules. Their goal was to see which combination of method and machine produced the most accurate predictions while using the fewest number of clues.

The results revealed a story of two very different worlds. On the credit card fraud dataset, the researchers found that aggressive compression worked beautifully. They discovered that they could cut the number of clues down from thirty to just thirteen without losing any predictive power. In fact, one of the photonic methods, which used a technique called boson sampling, performed exceptionally well when forced to choose only five clues. It managed to identify fraud with a high degree of accuracy, even outperforming the traditional computer when the number of allowed clues was very small. This happened because the fraud signals in this dataset were concentrated in a few specific, uncorrelated areas. The photonic sampler was able to find these rare, high-value combinations quickly, acting like a skilled detective who knows exactly which few fingerprints matter most in a specific case.

However, when the researchers applied the same logic to the consumer default dataset, the story changed completely. This dataset was larger and more complex, with the warning signs for default spread out across dozens of different indicators. Here, the strategy of picking just a few clues failed. No matter which computer or method they used, the models only reached their peak performance when they were allowed to use almost all 159 available clues. The photonic machines did not show any special advantage in this environment. The researchers observed that the "magic" of the new technology was entirely dependent on the structure of the data. When the important information was scattered and redundant, as it was in the consumer default case, the new machines could not compress the problem effectively. They found that the choice of solver mattered far less than the nature of the data itself.

Perhaps the most surprising discovery was that finding the mathematically perfect solution did not always lead to the best real-world prediction. The traditional computer was designed to find the single, certified best answer to the mathematical puzzle it was given. Yet, on several occasions, this "perfect" answer resulted in poor predictions for fraud or default. In one instance, the traditional computer found the optimal solution for a specific method but produced a prediction score that was less than half as good as a solution found by the photonic entropy computer. This happened because the mathematical formula used to guide the computer was only an approximation of what the bank actually needed. The photonic machine, by exploring a wider range of possibilities rather than locking onto a single mathematical optimum, sometimes stumbled upon a set of clues that worked better in practice, even if it wasn't the "perfect" answer to the equation.

The study concludes that there is no universal winner. The effectiveness of these advanced computing tools depends entirely on the landscape of the data they are analyzing. In compact, clean datasets where the important signals are concentrated, specialized photonic solvers can offer significant advantages, finding high-quality solutions with very few clues. But in larger, messier datasets where the signals are spread out and redundant, the benefits of compression disappear, and the choice of machine matters less than the choice of method. The researchers suggest that before investing in these new technologies, analysts should first understand the structure of their data. If the signals are concentrated, the new machines may be a powerful tool. If the signals are scattered, the traditional approach of using all available data remains the most reliable path. The paper does not claim that quantum or photonic computing has solved the problem of financial risk, but it provides a clear map of where these tools shine and where they do not, offering a practical guide for the future of financial modeling.

Technical Summary: Landscape-Dependent Performance of Photonic Quantum Solvers in QUBO Feature Selection for Financial Risk Detection

Problem Statement
Feature selection for imbalanced financial classification tasks, such as credit card fraud detection and consumer default prediction, requires balancing predictive relevance, inter-feature redundancy, and computational feasibility. The challenge is exacerbated by the "curse of dimensionality" and the specific structural properties of financial data. This paper addresses the performance of three distinct computing paradigms—classical exact optimization, photonic entropic computing, and photonic boson sampling—when applied to Quadratic Unconstrained Binary Optimization (QUBO) formulations of feature selection. The study specifically investigates whether performance trends observed in compact, low-dimensional feature spaces generalize to larger, more redundant datasets, and whether the choice of solver significantly impacts downstream predictive metrics (F1 score and Gini coefficient).

Methodology
The authors benchmarked thirteen feature-selection methods across two datasets with contrasting structures:

ULB Credit Card Fraud: A 30-feature dataset (Time, Amount, and 28 PCA-anonymized components) with 0.17% fraud prevalence. The signal is concentrated in a few uncorrelated components.
AmEx Consumer Default: A 159-feature dataset spanning five semantic categories (Payment, Balance, Spend, Delinquency, Risk). The signal is distributed across many moderately correlated features.

Three solver paradigms were employed, with a "solver-aware" design principle where methods were routed to the solver best matched to their mathematical structure:

Gurobi (Classical): Solved standard and graph-theoretic QUBOs (Tier 1 and Tier 2) to certified global optimality via branch-and-bound.
QCI Dirac-3 (Photonic Entropic): Solved standard, graph-theoretic, and hardware-native methods (Tier 3). Tier 3 methods exploited native capabilities such as multi-level qudit variables and cubic objective terms (HUBO, Integer-Weighted, Community-Poly).
Piquasso (Photonic Boson Sampling): A sampling-based approach using permanent-weighted triplet sampling without an explicit optimization objective.

The evaluation swept feature budgets ( $k$ ) from aggressive compression (e.g., $k=5$ on ULB, $k=25$ on AmEx) to near-full feature sets. Performance was measured using F1 score (primary metric for imbalanced classes) and Gini coefficient (ranking quality). Additionally, Jaccard similarity was used to analyze the diversity of selected feature subsets across solvers and methods.

Key Contributions

Solver-Aware Method Routing: The study introduces a framework where feature selection formulations are matched to solvers capable of handling their specific mathematical structures exactly (e.g., routing higher-order objectives to Dirac-3 rather than reducing them to quadratic forms).
Comprehensive Empirical Sweep: The authors provide a systematic comparison across 13 methods and 3 solver paradigms on two structurally distinct datasets, tracing the compression-performance frontier.
Diversity Analysis: The paper quantifies the overlap between selected feature subsets using Jaccard similarity, comparing observed overlap against the expectation for random subsets to distinguish genuine method agreement from chance.
Landscape-Dependent Insights: The study demonstrates that the efficacy of aggressive feature compression and the impact of solver choice are heavily dependent on the underlying data structure (signal concentration vs. signal dispersion).

Results

ULB Dataset (Compact, Concentrated Signal):
- Compression Advantage: Aggressive compression yielded significant performance. Dirac-3 with the MI-Spearman method at $k=13$ (43% of features) achieved a mean F1 of $0.873 \pm 0.023$ , matching the all-features model.
- Piquasso Outlier: Piquasso achieved the best F1 at the smallest budget ( $k=5$ , F1=0.874), outperforming all other methods and slightly exceeding the all-features model. This advantage was attributed to permanent-weighting, which selects features that are jointly relevant rather than just marginally relevant.
- Solver Sensitivity: Method choice significantly impacted performance at tight budgets. The spread in F1 scores across methods was approximately ten times larger on ULB than on AmEx.
- Optimality vs. Performance: Certified optimality by Gurobi did not guarantee the best downstream F1. For instance, Gurobi's optimal solution for the Dist-Corr method at $k=5$ yielded an F1 of 0.448, while Dirac-3's stochastic search found subsets with higher F1 scores.
AmEx Dataset (Large, Redundant Signal):
- No Compression Advantage: Performance improved steadily as more features were retained. No method achieved a compression advantage; all paradigms approached the all-features F1 ( $\approx 0.80$ ) only at large budgets ( $k=125$ –$150$).
- Solver Robustness: The choice of solver had minimal impact on performance. Gurobi's certified optima and Dirac-3's stochastic samples produced similar F1 scores, and the spread across methods was narrow.
- Optimality Disconnect: A significant disconnect was observed between optimization objective value and predictive performance. For the Dist-Corr method at $k=25$ , Gurobi's globally optimal solution achieved F1=0.422, whereas Dirac-3 achieved a mean F1 of 0.746 on the same method.
Subset Diversity:
- Selected feature subsets across different methods and solvers overlapped at levels close to random expectation on both datasets.
- Despite low overlap, the performance impact of these differences was dataset-dependent: on ULB, different subsets led to vastly different F1 scores, whereas on AmEx, different subsets were largely interchangeable in terms of performance.

Significance and Claims
The paper claims that effective feature selection requires balancing predictive relevance and redundancy against the specific computational capabilities of the solver. The primary finding is that specialized solvers and method choices matter most in compact, largely uncorrelated feature spaces with concentrated signals (like ULB), where aggressive compression is viable and solver stochasticity can sometimes outperform deterministic optima. Conversely, in larger, more redundant feature spaces (like AmEx), solver choice matters less, and aggressive compression often discards necessary signal, leading to performance degradation.

The authors modestly conclude that while photonic solvers (Dirac-3 and Piquasso) offer competitive or superior performance in specific regimes (particularly small budgets on concentrated data), they do not universally outperform classical exact solvers. The study emphasizes that "certified QUBO optimality" is not a guarantee of downstream predictive performance, as the optimization objective is often an approximation of the true classifier goal. The work suggests that the sensitivity of a problem to solver and method choice could potentially be estimated using inexpensive subsets before running full benchmarks.

Technical Summary: Landscape-Dependent Performance of Photonic Quantum Solvers in QUBO Feature Selection for Financial Risk Detection

More like this