Fraud Detection & Anomaly Analysis

100,000 retail transactions  ·  Jan 2023 – Dec 2024  ·  UK Retail Simulation

10,077 transactions flagged

Executive summary

Flagged transactions
10,077
10.1% of total volume
Revenue at risk
£6.81M
of £25.04M total (27.1%)
Critical risk
5,146
risk score ≥ 75 · auto-block
High risk
4,911
risk score 55–74 · manual review

Fraud pattern breakdown

Fraud type distribution
High-Risk Segment Velocity Fraud Off-Hours Large Amount Other
Monthly fraud rate (%)

Hourly risk heatmap

Fraud rate by hour of day — off-hours (00:00–04:59) flag at 100%
Low risk High risk

Risk segmentation & amount exposure

Fraud rate by customer segment
Segment Transactions Flagged Risk rate
High-Risk9,8463,68037.4%
Premium19,9951,4387.2%
Regular54,8683,8867.1%
New15,2911,0737.0%
Fraud rate by transaction amount band

Channel & merchant analysis

Fraud rate by channel
Top suspicious merchants (flagged count)

Category & geographic exposure

Fraud rate by merchant category
Flagged transactions & fraud rate by city

Loss prevention thresholds

Rule-based detection logic & recommended actions
Action Risk tier Trigger conditions
Auto-block Critical (score ≥ 75) risk_score ≥ 90  OR  amount > £5,000  OR  hour in [0–4]
Manual review High (score 55–74) risk_score 55–89  OR  segment = 'High-Risk'  OR  amount > £1,000
Monitor Medium (score 35–54) risk_score 35–54  OR  velocity flag  OR  channel = 'Phone'
Pass Low (score < 35) risk_score < 35  AND  no rule triggers

Key findings & insights

Off-hours transactions (00:00–04:59) show a 100% fraud rate — the strongest single rule-based signal. Immediate auto-blocking of this window is recommended.
High-Risk segment customers are 5.3× more likely to be flagged than Regular customers (37.4% vs 7.1%). Segment-based thresholds should differ substantially.
Transactions above £5,000 carry an 89.6% fraud rate. An amount-based hard ceiling should trigger automatic blocking pending manual review.
Risk scores separate fraudulent from clean transactions by a 5× margin (avg 75.2 vs 15.0), confirming the scoring model as a reliable real-time filter.