returnguard 1.2.0

Returns fraud detection for retail and eCommerce — wardrobing, serial returner, refund anomaly detection, behavioral fingerprinting, policy simulation

returnguard

AI-powered returns fraud detection for retail and eCommerce — score return requests, detect wardrobing, serial returners, bot patterns, and policy abuse. No $50K enterprise contract required.

The Problem

Returns fraud costs US retailers $101B/year. AI-generated fraud has "exploded overnight." Enterprise solutions (Happy Returns, Narvar) target only large retail — Shopify has zero built-in fraud scoring. Mid-market merchants are on their own.

Installation

pip install returnguard

Quick Start

from returnguard import FraudScorer, ReturnRequest, ReturnReason
from datetime import datetime, timedelta

scorer = FraudScorer(
    return_rate_threshold=0.30,
    high_value_threshold=150.0,
    velocity_limit=3,
)

request = ReturnRequest(
    return_id="RET-001",
    order_id="ORD-5521",
    customer_id="CUST-42",
    sku="SKU-JACKET-XL",
    quantity=1,
    reason=ReturnReason.CHANGED_MIND,
    order_date=datetime.utcnow() - timedelta(days=3),
    order_value=220.00,
    channel="shopify",
)

result = scorer.score(request)
print(result.risk_level)        # RiskLevel.HIGH
print(result.score)             # 0.6
print(result.signals)           # [FraudSignal.WARDROBING]
print(result.recommended_action)  # "Require photo evidence + manual review"

Fraud Signals Detected

Signal	Description
HIGH_RETURN_RATE	Customer's return rate exceeds threshold
WARDROBING	Use-and-return: changed_mind + < 7 days + high-value item
VELOCITY	Too many returns in a short window
SERIAL_RETURNER	Customer has been flagged multiple times
POLICY_ABUSE	Return submitted after policy window

Customer Profile Tracking

from returnguard import CustomerProfile

profile = CustomerProfile(
    customer_id="CUST-42",
    total_orders=20,
    total_returns=8,
    flagged_count=2,
)

result = scorer.score(request, profile=profile)
# Score accounts for historical return behaviour

Risk Levels & Actions

Risk Level	Score	Recommended Action
LOW	0.0–0.30	Auto-approve
MEDIUM	0.30–0.55	Flag for manual review
HIGH	0.55–0.75	Require photo evidence
CRITICAL	0.75–1.0	Block + escalate to fraud team

Batch Scoring

from returnguard import batch_score, abatch_score

# Sync
scores = batch_score(requests, scorer.score, max_workers=8)

# Async
scores = await abatch_score(requests, scorer.score, max_concurrency=8)

Advanced Features

Pipeline

from returnguard import FraudPipeline

pipeline = (
    FraudPipeline()
    .filter(lambda s: s.score > 0.5, name="high_risk_only")
    .map(lambda scores: sorted(scores, key=lambda s: -s.score), name="sort_by_risk")
    .with_retry(count=2)
)

high_risk = pipeline.run(all_scores)
print(pipeline.audit_log())

Caching

from returnguard import FraudCache

cache = FraudCache(max_size=1000, ttl_seconds=600)

@cache.memoize
def score_with_cache(request):
    return scorer.score(request)

print(cache.stats())

Validation

from returnguard import ReturnValidator, ReturnRule

validator = ReturnValidator()
validator.add_rule(ReturnRule("max_days", 60, "Return window expired"))
validator.add_rule(ReturnRule("max_order_value", 1000, "High-value item requires manual review"))

valid, errors = validator.validate(request)

Diff & Trend

from returnguard import diff_scores, RiskTrend

diff = diff_scores(previous_scores, current_scores)
print(diff.summary())  # {'added': 3, 'removed': 1, 'modified': 2}

trend = RiskTrend(window=20)
for score in historical_scores:
    trend.record(score.score)
print(trend.trend())       # "increasing"
print(trend.volatility())  # 0.12

Streaming & NDJSON

from returnguard import stream_scores, scores_to_ndjson

for score in stream_scores(results):
    process(score)

for line in scores_to_ndjson(results):
    file.write(line)

Audit Log

from returnguard import AuditLog

log = AuditLog()
log.record("scored", "RET-001", detail="risk=high")
log.record("blocked", "RET-001")
entries = log.export()

License

MIT