QBVK › Finans & BankacılıkFinance & Banking

KKB Kuantum Hazırlık & AvantajKKB Quantum Readiness & Advantage

Kredi bürosu senaryosunda kuantum avantaj stres testi (simülatör + gerçek donanım iskeleti).Credit-bureau quantum advantage stress test (simulator + real-HW scaffold).

headless: kısmenheadless: partial BVK_KKB_Quantum_Advantage_RealHW_Colab.ipynb 22/31 hücre · 642.6 sn22/31 cells · 642.6 s

Yaklaşım ve teknolojiApproach & technology

Defter akışıNotebook outline

• 100,000 synthetic bureau applicants, a quantum-structured risk signal and portfolio tail risk — simulated, rehearsed wit
• Setup — libraries, the BVK engine and the logo
• What "quantum advantage" can mean for a credit bureau
• Hardware configuration — modes, budget presets and credentials
• The data — a synthetic credit-bureau file built to be hard
• Track 1 — Realistic bureau data: scorecard and gradient boosting vs a quantum kernel
• Track 2 — A quantum-structured risk signal, with a hardware-native quantum model
• Track 3 — Portfolio tail risk: amplitude estimation vs Monte Carlo
• Track 4 — Real hardware: IBM Fez and Origin Wukong
• Track 5 — The advantage screen KKB can run on its own data
• Hardware readiness — what a run on today's IBM hardware would face
• Narrator Engine — executive Word report and machine-readable results
• Summary

Sonuç grafikleriResult charts

Çıktılardan metrik satırları (birebir)Metric lines from outputs (verbatim)

0          WOE logistic scorecard (all 55 columns)              50280.0   0.8589    0.7178  0.5633   1.0500
0    Fidelity quantum kernel SVM (10 qubits, ZZ map)  0.8270    0.8464       0.3348              0.3347  5.2006
7  WOE scorecard (all columns, full 50,280 develo...     NaN    0.8444          NaN                 NaN     NaN
1       Projected quantum kernel (10 qubits, ZZ map)  0.8156    0.8384          NaN                 NaN     NaN
8  WOE scorecard (all columns, full 50,280 develo...     NaN    0.8444          NaN                 NaN     NaN
Projected quantum kernel: bandwidth 0.1, γ 0.5, C 1 → test AUC 0.838
• The projected quantum kernel used on hardware in Track 4 scores 0.838 on the same test applicants (best classical on the same features: 0.849).
• Training-size sweep: the quantum-minus-best-classical AUC gap ranges from -0.025 (n=100) to +0.008 (n=1600).
Projected quantum kernel chosen by CV (learner is not told the generating rule): bandwidth 0.5, γ 0.5, C 1 (CV AUC 0.932)
Quantum minus best classical AUC by training size: {100: 0.251, 250: 0.28, 500: 0.289, 1000: 0.289, 2000: 0.266}
Loss distribution (units): [6.810e-01 9.610e-02 1.623e-01 3.460e-02 1.910e-02 5.100e-03 1.400e-03
3.000e-04 1.000e-04 0.000e+00] · limit = 4 units → exact P(loss ≥ limit) = 0.02596
Deepest circuit: k=166 → 921,216 two-qubit gates, ESP ≈ 0 (below 1e-12); needs a two-qubit error rate ≤ 7.5e-07 (today ~3.9e-3).
A · realistic control (Track 1): 100 train / 100 test · best classical by CV: RBF-SVM (AUC 0.806) · noiseless projected kernel AUC 0.782

Kaynak: Source: QBVK headless çalıştırma · 04.11.2026QBVK headless run · 2026-11-04

Colab'da açOpen in Colab

Defteri indirin, ardından Colab'da Dosya → Not defteri yükle. ⭐ = önerilen sürüm. Gizli anahtarlar yayından çıkarıldı.Download the notebook, then in Colab use File → Upload notebook. ⭐ = recommended version. Secrets were removed before publishing. colab.research.google.com ↗