Source-linked AI summary
Quantum Machine Learning for Credit Scoring
Nikolaos Schetakis, Davit Aghamalyan, Michael Boguslavsky, Agnieszka Rees, Marc Raktomalala, Paul Griffin
TL;DR
This paper examines whether hybrid quantum machine learning can be applied to SME credit scoring, where small firms may lack rating-agency coverage and practical QML use cases remain limited. It compares a full hybrid classifier with classical models on Singapore SME data, finding comparable AUC performance with fewer training epochs while noting simulator scaling constraints.
Problem
The paper asks whether QML can be applied to SME credit scoring, where small firms are often not covered by rating agencies and practical real-world QML use cases remain limited.
Source: §1 · pp. 1–2Method
The authors benchmark a full hybrid classifier combining classical neural-network layers with variational-circuit and data-reuploading components on Singapore SME data.
Source: §2.2 · pp. 4–5Results
The hybrid model achieved an ROC/AUC of 0.75, compared with 0.73 for the classical counterpart at 350 epochs; the classical model reached 0.75 when trained up to 5,000 epochs.
Source: §3 · p. 6Takeaways & Limitations
The Singapore-data results support further exploration of hybrid QML for SME credit scoring, with comparable prediction performance reported using fewer training epochs.
Source: §3 · p. 6, §4 · p. 8Simulator execution has exponential time complexity as the number of qubits and blocks increases, while real-device feasibility depends on circuit width and depth staying within device specifications.
Source: §4 · p. 9
Abstract
from arXiv · showhide
In this paper we explore the use of quantum machine learning (QML) applied to credit scoring for small and medium-sized enterprises (SME). A quantum/classical hybrid approach has been used with several models, activation functions, epochs and other parameters. Results are shown from the best model, using two quantum classifiers and a classical neural network, applied to data for companies in Singapore. We observe significantly more efficient training for the quantum models over the classical models with the quantum model trained for 350 epochs compared to 3500 epochs for comparable prediction performance. Surprisingly, a degradation in the accuracy was observed as the number of qubits was increased beyond 12 qubits and also with the addition of extra classifier blocks in the quantum model. Practical issues for executing on simulators and real quantum computers are also explored. Overall, we see great promise in this first in-depth exploration of the use of hybrid QML in credit scoring.
1. Introduction
Small and medium-sized enterprises (SMEs) may have reduced access to credit because they are often not covered by rating agencies. This paper applies hybrid quantum machine learning to SME credit scoring and benchmarks it against classical models.
1. Introduction
- The paper situates credit scoring within a search for practically useful real-world problems for quantum computing and quantum machine learning.It identifies credit scoring and portfolio optimization among computational-finance problems relevant to quantum technologies.Source: §1 · pp. 1–2
- SMEs are often not covered by rating agencies, and limited credit scoring can reduce their access to credit.The paper notes that even small improvements in scoring efficiency could lower rejections of healthy SMEs and reduce risk for lenders.Source: §1 · pp. 1–2
- The study applies the authors’ full hybrid classical-quantum neural network to credit scoring and benchmarks it against classical machine-learning models.The authors describe the algorithm as heuristic and provide extensive benchmarking against classical models.Source: §1 · p. 2
2. Materials and Methods
The section reviews credit-scoring evaluation and classical approaches, then describes the full hybrid classifier and its Singapore SME dataset and experimental setup.
2. Materials and Methods
- The experiments compare credit-scoring classifiers using the trade-off between rejecting healthy companies and identifying companies that default.AUC is the paper’s performance metric for comparing models.Source: §2 · p. 3
2.1. Classical models
- Classical credit scoring is framed as imbalanced binary classification, and model performance can vary with the data, model choices, and evaluation metric.The paper uses a Singapore model with AUC 0.73 as its classical benchmark and does not include XGBoost in the comparison.Source: §2.1 · pp. 3–4
2.2. Quantum models
- The full hybrid model combines classical neural-network layers with variational-circuit and data-reuploading components in its quantum part.The quantum circuit encodes data, applies trainable gates and entangling layers, and passes measurements to a classical decision layer.Source: §2.2 · pp. 4–5
2.3. Data description
- The Singapore dataset covers nearly 2,300 SMEs and uses 24 primary features, with 246 defaulting firms and a random subsample of 2,000 healthy firms.Limited borrower financial information restricts the analysis to binary default/no-default models and panel data rather than time-series analysis.Source: §2.3 · p. 5
2.4. Experiments
- The experiments evaluate an FH:NN/VC-DRC model with a classical neural network before the quantum circuit and a final sigmoid decision layer.The setup uses 21 input features; the feeding classical layer has the same number of neurons as qubits.Source: §2.4 · pp. 5–6
- Quantum and classical models were compared across 6–18 qubits and 1–10 DRC blocks, averaging AUC over five simulations per configuration.The models used 350 epochs for a fair comparison, though classical performance could improve with training up to 5,000 epochs; simulations were limited to 18 qubits by execution time.Source: §2.4 · p. 6
3. Results
The FH model matched the classical model’s best reported ROC/AUC with fewer training epochs, while its performance varied with qubit and block counts. Simulator timings also showed different scaling with qubits and blocks.
3. Results
- 0.75 ROC/AUC was reported for FH, versus 0.73 for CC at 350 epochs and 0.75 for CC trained up to 5,000 epochs.Figure 3 compares the models’ ROC/AUC scores.Source: §3 · p. 6, Figure 3 · p. 7
- With one block, FH ROC/AUC rose through 12 qubits and then fell; adding blocks showed the same behavior, with the highest results at no data-reupload.Source: Figure 3 · p. 7
- In all Figure 4 simulations shown, FH achieved its highest testing ROC/AUC with fewer training epochs than CC.Figure 4 shows CC at 3,500 epochs in the left column and 350 epochs in the middle column, with FH in the right column.Source: Figure 3 · p. 7, Figure 4 · p. 7
- On the simulator, embedding time doubled per added qubit, model execution scaled quadratically with qubits and linearly with blocks, and tests used up to 20 qubits.Figure 5 plots execution time per epoch and training step as qubit count or block count increases.Source: §3 · p. 8, Figure 5 · p. 8
4. Discussion
The discussion reports comparable accuracy with fewer FH training epochs, but notes performance declines with more qubits or blocks. It also highlights simulator, cloud-execution, dataset-size, and tuning constraints.
4. Discussion
- FH accuracy equaled CC accuracy while requiring at least two times fewer training epochs.The authors describe training time as a significant business driver.Source: §4 · p. 8
- The chosen FH architecture uses a classical feeding layer to reduce features entering the quantum layer, which is limited by qubit count for this dataset.Source: §4 · p. 8
- Accuracy declined above 12 qubits and with added blocks; the authors suggest barren plateaus or a mismatch with trigonometric data structure, but rule out overfitting.Source: §4 · p. 8
- Simulator time complexity grows exponentially with qubits and blocks, while real-hardware feasibility depends on circuit width and depth meeting device specifications.Source: §4 · p. 9
- For 350 epochs on 1,798 training examples, the model requires 629,300 circuit executions; at one second each, cloud overhead and queuing could make execution take seven days.The authors say Qiskit runtime could reduce that estimate by 120 times, to 1.5 hours.Source: §4 · p. 9
- The dataset is limited by available class-1 examples, and the authors say FH and CC hyperparameters could still be tuned further.Source: §4 · p. 9
5. Conclusions
The authors conclude that hybrid quantum/classical models are promising for comparable results with fewer training epochs. They identify further hardware and software improvements, while noting that XGBoost remains the best classical model for Singapore data.
5. Conclusions
- Hybrid quantum/classical models achieved comparable results to a purely classical counterpart with much fewer training epochs, which the authors regard as promising.Source: §5 · p. 9
- The authors expect future hardware and software improvements to make real advantages, at least in model training, possible.Source: §5 · p. 9
- XGBoost remains the best classical model for Singapore data; the authors propose extending the work to other datasets with more features and qubits.They also state that machine-learning researchers can relatively easily experiment with QML for their own problems.Source: §5 · p. 9