anomaly-fraud-scanner

Anomaly Fraud Scanner

Table of Contents

• Overview
• Input contract
• Detection rules
• Workflow
• Output contract
• Guardrails

Overview

Fraud detection on a household budget is a different problem from fraud detection at a bank. The bank already runs sophisticated rules; this skill exists to catch what the bank misses — small recurring fraud below their threshold, unfamiliar merchant patterns, and post-hoc audit of charges the user has not yet noticed.

It runs five complementary detectors, each producing severity-tagged alerts.

Input contract

The caller provides:

• transactions_new — transactions added by the most recent drop.
• transactions_history — last 12 months of confirmed transactions.
• accounts — for context (account type, owner, geography).
• today — ISO date.
• thresholds — optional overrides:

  {
    "duplicate_window_hours": 48,
    "anomaly_sigma": 3.0,
    "first_merchant_high_dollar_cents": 50000,
    "geo_radius_miles": 1000
  }
  

Detection rules

Rule 1 — Duplicate within 48 hours

Two transactions on the same account_id with the same merchant and the same amount_cents posted within duplicate_window_hours of each other.

• Severity: medium for amounts < $200, high for amounts ≥ $200.
• Common false positives: two coffees in a day, two gas-station fill-ups during a road trip — exclude food.coffee, food.restaurants, transportation.gas from this rule when amount is below $30.

Rule 2 — Statistical anomaly

For each (merchant, account_id) pair with ≥ 6 historical occurrences, compute mean and stddev. Flag any new transaction where |amount − mean| > anomaly_sigma × stddev AND |amount − mean| ≥ $50.

• Severity: medium baseline, high if > 5σ or amount > 4× mean.
• Common false positives: tip variations on restaurants, surge pricing on rideshare — bake a 25% leniency on food.restaurants and transportation.rideshare.

Rule 3 — First merchant high dollar

A transaction with a merchant that has zero prior occurrences in transactions_history AND |amount| ≥ first_merchant_high_dollar_cents.

• Severity: medium for $500–$2,000, high for ≥ $2,000.
• Common false positives: a planned major purchase (appliance, travel booking). Suppress when a recent category_override or user-confirmed plan exists.

Rule 4 — Geography or time anomaly

If description_raw includes a location string (state code, country code) and the inferred location differs by more than geo_radius_miles from the household's home location AND no travel context is set, flag.

If a transaction posts at an unusual time (between 2am–5am local) AND the merchant category is unusual for that hour (not gas station, not 24-hour pharmacy), flag.

• Severity: medium.
• This rule is noisy — use only when a location/time signal is clearly present.

Rule 5 — Unexpected fee

Any transaction with category: financial.fees is captured. Tag with fee_type (overdraft, foreign-transaction, monthly-maintenance, ATM, late-payment).

• Severity: low for fees < $10, medium for fees $10–$50, high for overdraft fees and late-payment fees regardless of amount.
• Always actionable: most fees are negotiable or avoidable.

Workflow

Anomaly Scan Progress:
- [ ] Step 1: Build merchant statistics from transactions_history
- [ ] Step 2: Run Rule 1 (duplicate within 48h)
- [ ] Step 3: Run Rule 2 (statistical anomaly)
- [ ] Step 4: Run Rule 3 (first-merchant high-dollar)
- [ ] Step 5: Run Rule 4 (geography/time)
- [ ] Step 6: Run Rule 5 (fees)
- [ ] Step 7: Deduplicate alerts (one tx may trigger multiple rules)
- [ ] Step 8: Rank by severity then dollar impact
- [ ] Step 9: Attach recommended actions

Deduplication

A single transaction may legitimately trigger multiple rules (e.g., a $5,000 charge at a brand-new merchant in another state — Rules 3 and 4). Combine into a single alert with triggered_rules: [...] and the highest severity.

Recommended actions

Each alert carries one of:

• call_bank_fraud_line — for high severity charges that the user does not recognize.
• freeze_card — for confirmed fraud or before calling the bank if multiple charges fired within hours.
• dispute_charge — for known-merchant overcharges or duplicate post-and-pending issues.
• negotiate_fee — for financial.fees (overdraft, monthly maintenance) that have a known negotiation playbook.
• monitor — for medium/low alerts where the user should confirm but no action is required if it's legitimate.

Always include a one-line evidence describing what triggered the rule, with concrete numbers.

Output contract

{
  "scanned": {
    "transactions_new": 142,
    "transactions_history": 1820,
    "merchants_with_stats": 218
  },
  "alerts": [
    {
      "id": "alert_20260424_001",
      "severity": "high",
      "triggered_rules": ["first_merchant_high_dollar"],
      "tx_ids": ["tx_20260424_011"],
      "merchant": "TECH-SHOP-ONLINE",
      "amount_cents": -245000,
      "account_id": "acc_cc_001",
      "evidence": "First charge ever from TECH-SHOP-ONLINE; amount $2,450 exceeds new-merchant high-dollar threshold of $500.",
      "suggested_action": "call_bank_fraud_line",
      "expected_user_response": "Confirm whether this purchase is yours; if not, freeze card and dispute."
    },
    {
      "id": "alert_20260424_002",
      "severity": "medium",
      "triggered_rules": ["statistical_anomaly"],
      "tx_ids": ["tx_20260424_023"],
      "merchant": "Whole Foods",
      "amount_cents": -38420,
      "evidence": "12-month avg $84.50, stddev $22.30; this charge ($384.20) is 13.4σ above mean.",
      "suggested_action": "monitor",
      "expected_user_response": "If you bought groceries for a party, ignore. Otherwise, check the receipt."
    },
    {
      "id": "alert_20260424_003",
      "severity": "high",
      "triggered_rules": ["fee_overdraft"],
      "tx_ids": ["tx_20260423_004"],
      "merchant": "Bank Name",
      "amount_cents": -3500,
      "evidence": "Overdraft fee of $35 charged on 2026-04-23. Account balance was negative for 2 days.",
      "suggested_action": "negotiate_fee",
      "expected_user_response": "Call the bank's customer line; first overdraft fees are typically waived on request."
    }
  ],
  "summary": {
    "alerts_total": 3,
    "high_severity": 2,
    "medium_severity": 1,
    "low_severity": 0
  }
}

Guardrails

• Never assert fraud. Use language like "unusual," "first-ever," "above threshold." The user confirms; this skill flags.
• Tune for false-positive cost. A medium-severity alert is read; a high-severity alert disrupts the user's day. Reserve high for truly unusual.
• Cite the math. Every alert states the comparator (mean, stddev, occurrence count, threshold). The user must be able to evaluate the reasoning.
• Honor history-context. A "first-ever merchant" trigger should not fire when the merchant exists in history under a different rendered string. Use normalized merchant names from the categorizer.
• Do not action without consent. Suggested actions are for the user to take; the household-finance system never executes financial actions.
• Retain the alert log. All emitted alerts append to reports/alerts/. False positives feed back into the threshold tuning.