Definition
Weighted change in fraud detection recall across fraud categories over time, measuring whether model coverage is quietly eroding on the categories that matter most.
The Measurement Gap
Measurement Gap and Decision
What FDCD Measures: FDCD tracks the weighted change in fraud detection recall across fraud categories over time. It reveals whether the detection model is keeping pace with evolving fraud vectors or optimizing for historical patterns while novel fraud goes undetected.
Decision Supported:
Is the fraud detection model keeping up with new fraud types, or is it becoming a historical-pattern detector while eSIM exploitation, IoT device hijacking, and API abuse on 5G slices go undetected? Without FDCD, fraud teams report total fraud loss and detection volume without examining whether coverage is expanding or shrinking on the categories that represent emerging threats.
Why This Is Unmeasured:
Fraud management systems report aggregate detection rates, total fraud value blocked, and false positive rates. None of these reveal whether the model's coverage is shifting: specifically, whether recall is improving on well-known fraud types (SIM swap, subscription fraud) while declining on emerging types (eSIM exploitation, 5G slice abuse, API credential farming). The model's overall detection rate can hold steady or even improve while it loses ground on the fraud categories that represent the highest future risk. FDCD measures the direction of coverage change, weighted by current threat intelligence, to expose this drift before it becomes a financial loss event (GSMA Fraud Forum, "T.A.D. Framework: Fraud Detection Maturity Assessment," 2024; ITU-T, "Security Aspects of Telecommunications Fraud Management," 2023).
Formula
Plain Language: Segment fraud into categories based on the ITU/GSMA taxonomy or internal classification. Measure detection recall per category per quarter. Weight each category by current fraud intelligence frequency and financial impact. Track whether weighted coverage is expanding or contracting.
Notation:
FDCD = Σ (w_c × ΔRecall_c) / Σ w_c
Variable Definitions:
- w_c = weight of fraud category c, based on current fraud intelligence frequency and estimated financial impact (dimensionless weight, higher = more important)
- ΔRecall_c = change in detection recall for fraud category c between the current and previous measurement period (percentage points)
- Recall for a category = confirmed detected fraud events / total confirmed fraud events (detected + missed) for that category
Units: Weighted percentage change per measurement period. Negative FDCD = coverage is shrinking on the weighted-important categories (bad). Positive FDCD = coverage is expanding (good). Zero = stable coverage.
Sign Convention:
Positive is better. Negative FDCD means the fraud detection system is losing ground on the categories that matter most right now.
Edge Cases:
- New fraud categories with no historical detection data: assign recall = 0 for the first measurement period. This correctly registers as a gap.
- Categories with zero fraud incidents in the current period: carry forward the previous recall estimate and flag the category as "untested". Absence of incidents does not mean absence of fraud.
- Red-team testing or fraud simulation may be required for categories with insufficient organic volume to measure recall. Budget 1–2 red-team exercises per year for emerging fraud types.
Proposed formula, pending validation pilots.