Each sphere is a loan applicant placed in a 2D feature plane (credit history × digital-footprint / social-graph signal). Its height is the model's 0–1000 score. A translucent plane marks the approval threshold: above it applicants turn green (approve), a narrow band below turns amber (refer for manual review), and further below turns red (decline) — mirroring the bank's own approve / decline / refer output.
score = base + wA·credit_history + wB·digital_footprint + wC·social_graph
fraud_flag = anomaly(location, amount, cadence) > sensitivity
- Approval score — moving the threshold trades approval rate against the estimated non-performing-loan (NPL) rate, just as a bank tunes its cutoff.
- Dominant signal — highlights which SHAP-style feature currently drives the score the most, tilting the cloud toward that axis.
- Fraud sensitivity — the streaming packets on the right are simulated transactions; raising sensitivity blocks more of them in real time (fewer false negatives, more manual confirmations).
In production this combination (explainable score + real-time anomaly detection) is what let the bank cut approval time from days to ~90 seconds while catching most fraud before the money left the account.