Budgeted Class-Balanced Drift Gating for Directional Temporal Graph Fraud Detection (Journal of Mathematical Finance and Risk Management)

Summary

• BCG-5 combines a class-balanced predictive-loss gate, a five-update stream budget and recent-only plastic fitting for graph-based fraud detection under concept drift.p.1
• Experiments use synthetic directed transaction graphs with 49 chronological snapshots and five paired seeds in three regimes; no production data are used.p.1
• In the mixed regime BCG-5 reaches average precision 0.3175 and F1 0.3413, against 0.3004 and 0.3367 for the loss-gated ATGNN-SP update rule.p.1
• BCG-5 reduces mean updates from 8.0 to 5.0.p.1
• Total plastic fitting time is 0.0864 s for BCG-5 versus 0.2756 s for the replaying baseline, a 68.7% reduction.p.5
• A plastic-only detector ranks better, with AP 0.3709 and F1 0.4361, at the cost of more updates.p.4

Key insights

1. BCG-5 improves on the ATGNN-SP update rule by 0.0171 AP and 0.0046 F1 in paired mean differences in the mixed regime.p.4
2. The update budget falls by 37.5% in the mixed setting, from 8.0 to 5.0 updates.p.5
3. An always-updated model uses 15 updates and reaches mixed AP 0.2963, below BCG-5, so unselective fitting need not help.p.5
4. Removing historical replay lowers BCG-5 mixed AP from 0.3175 to 0.2963 in the ablation.p.5
5. The plastic-only result shows a stability-adaptability trade-off: higher accuracy but more updates and no frozen prediction branch.p.6
6. At a positive prior of 0.30 the ordinary alarm fires 0.8517 of the time versus 0.1938 for the balanced alarm.p.7
7. All predictive metrics come from deliberately simplified synthetic flows and need testing on real financial transaction records.p.8