Fraud Scoring Pipeline

Agentic Orchestration - Real Time Fraud Scoring

Two independent signals score every transaction — a trained model reading 28 spend & behavior features, and a peer-cohort comparison against the customer's category history — then combine into one final score.

Training rows
284,807
Fraud rate
0.173%
Signals used
28 / 28
ROC-AUC
0.971
PR-AUC
0.708
Recall
0.918

The 28 modeled signals

The dataset's real columns are PCA-anonymized (V1…V28) — never publicly mapped to real attributes. Names below are illustrative stand-ins used only to make this diagram readable.

Spend level & category mix

avg_transaction_amount_3_monthsamount_spent_restaurants_3_monthsamount_spent_travel_3_monthsamount_spent_electronics_3_monthsamount_spent_groceries_3_monthsavg_bill_3_monthsmax_single_transaction_amount_3_months

Behavioral history & travel

travel_history_3_monthsmerchant_category_diversity_3_monthsrepeat_merchant_ratio_3_monthsaccount_age_daysavg_daily_spend_30_days

Frequency & velocity

transaction_frequency_7_daystransaction_frequency_24_hoursdistinct_merchants_30_daysavg_time_since_last_transactionspend_velocity_change_7_daysdeclined_transaction_count_30_days

Timing patterns

night_time_transaction_ratioweekend_transaction_ratioavg_transaction_gap_hours

Merchant & channel risk

new_merchant_transaction_ratiohigh_risk_merchant_category_ratiocross_border_transaction_ratiocard_present_ratioonline_transaction_ratio

Statistical deviation

transaction_amount_zscore_vs_categoryavg_transaction_amount_deviation_daily

Training — offline, once

Historical transactions teach the model what fraud looks like across all 28 signals at once.

Historical data
284,807 transactions
0.173% fraud
→
Features
28 spend & behavior signals
→
Train
Fraud Propensity Model
→
Saved
Model artifact
(reused for every future score)

Agent A — propensity scoring (real time)

Every new transaction's 28 signals are fed through the trained model in milliseconds.

Spend & category mix
Behavioral history & travel
Frequency & velocity
Timing patterns
Merchant & channel risk
Statistical deviation
Fraud
Propensity
Model
Propensity Score
0 – 1

Agent B — behavioral scoring (real time)

Not a trained model — a live comparison of this transaction against how similar customers spend in the same category.

This transaction
(amount, category, merchant)
Category peer average & std. dev.
Customer's recent history
Peer-Cohort
Comparison
Behavior Anomaly
Score 0 – 1

Combine & decide

Both scores blend into one number, which routes the transaction to an outcome.

Propensity Score
Behavior Score
Weighted
Combine
0.7 · A + 0.3 · B
Auto-approvescore < 0.45
Human review0.45 – 0.85
Blockscore ≥ 0.85
Why recall over precision: the model is tuned to rarely miss real fraud — a low score never finalizes a decision alone. Why no Amount/Time feature: Time is a row serial number here, not a real timestamp; spend level is already captured across the 28 signals.