Fraud Detection with Location and Property Signals

The costliest fraud evades credit scores and device fingerprints: synthetic identities and first-party bust-out. Global Mobility and Real Estate Property Data answer questions bureau files cannot: does this person live where claimed, and are they on the deed? This guide covers gaps in traditional stacks, residency patterns, documentary property checks, and the two-stage workflow banks use. Pair with identity graphs 101 and non-FCRA mortgage leads for marketing vs risk boundaries.

Where the Traditional Stack Falls Short

Bureau and device checks catch most stolen-identity and velocity fraud. What they miss concentrates in synthetic and bust-out fraud, which is expensive because a cultivated identity passes those gates. FinCEN advisories note signals are outside credit files attackers cultivated.

What Location Data Answers

MAID mobility at scale tests actual residence vs claimed address. Legitimate residents show:

• Evening/overnight presence at claimed address (10 PM-6 AM, 5+ nights/week).
• Weekend daily-life POIs within ~5 miles.
• Weekday commute or service POI patterns consistent with employment.
• Stable device set over 90+ days.

Synthetic patterns: no overnights, no nearby POIs, sometimes no US pattern. Join via MAID Feed and identity graph. See identity graphs 101.

What Property Data Confirms

Property file confirms owner/tenant of record: deed, tax, mortgage anchor. Tenancy weaker than ownership; combine with mobility. Ownership claim + no record + no device residency ≈ high-confidence flag. FBI mortgage fraud reports document occupancy misrepresentation; scale detection is what changed.

The Two-Stage Operational Pattern

Stage 1 application-time property lookup: sub-second, low false-positive, rarely sole decline. Stage 2 batch 24-72h mobility vs residence: catches owner-of-record overseas while device abroad. Loss reduction on synthetic and first-party tranches is the usual justification for this stack; size it against your own book rather than a vendor benchmark. Integrate via risk management and fraud and financial services.

Each signal alone over-triggers; combined score + investigation queue is operable. Document permitted use: fraud prevention vs marketing differs in DPA.

Procurement and Governance for Risk Use Cases

Risk buyers need sensitive-location exclusions, retention limits, and audit rights in contract: same RFP governance pillar as marketing mobility. Map fields to model cards; test deletion propagation when sources opt out.

Base Rates Decide Whether a Strong Signal Is Usable

Fraud signals are evaluated on accuracy and deployed against base rates, and the gap between those two facts is where most risk programs get into trouble. When genuine fraud is rare in a population, even a highly accurate rule produces far more false positives than true ones, because the small error rate applies to a very large honest majority. A rule described as rarely wrong can still generate a review queue in which most flagged cases are innocent.

Insist on precision at your operating threshold rather than an overall accuracy figure, and require it be measured on your own population rather than a vendor benchmark. Then convert it into the number that actually determines feasibility: how many reviews per confirmed case, and whether your team can work that volume without pushing the threshold somewhere unsafe. A signal that is analytically excellent and operationally unaffordable will be quietly switched off within a quarter, which is worse than never deploying it.

Fraud Signals Decay Because Someone Is Watching Them Too

Most data decays passively as the world moves on. Fraud signals decay adversarially, because the population being measured adapts to the measurement. A location or device pattern that separates fraud cleanly today will separate it less well once the technique is understood, and the adaptation shows up as a slow decline in precision that is easy to attribute to data quality rather than to a changing adversary.

Plan for that from the start. Monitor precision on a rolling basis rather than validating once at launch, keep a holdout that is not used for tuning so drift can be seen cleanly, and expect to retire and replace individual signals rather than maintaining a fixed rule set. Diversity matters more than in other domains: several independent signals are harder to defeat simultaneously than one strong signal, even where the single signal tests better in isolation.

A Decision You Cannot Explain Is a Decision You Cannot Defend

Risk decisions get challenged, by customers who were declined and by internal review, and a signal that cannot be explained in plain language becomes a liability at exactly the moment it matters. Before deploying a location or property signal, establish what you would tell a customer about why their application was routed for review, and confirm that explanation is both truthful and something you are willing to say out loud.

Keep the reasoning behind each decision, not just its outcome, so a case can be reconstructed months later. Also be deliberate about which signals are permitted to decline automatically and which may only route for human review, because the two carry very different obligations and very different consequences when wrong. Use in risk contexts frequently carries specific regulatory obligations depending on the decision being made, and the applicable requirements should be confirmed with counsel before deployment rather than after.

The Benign Lives That Produce the Synthetic Signature

The residency pattern above flags absence: no overnights at the claimed address, no daily-life POIs nearby, no stable device set. Several entirely legitimate populations produce that same absence. Recent movers carry a device history anchored to the prior address, and the bureau file usually lags the move too, so both signals agree and both are wrong. Seasonal residents split the year across two homes and can fail an overnight test at either. Students, traveling clinicians, long-haul drivers, and deployed service members generate the pattern by design. Adult children and caregivers who list a parent's address, households where the deed and the utility sit in different names, and people who left a home after a protective order all present a claim the property record will not confirm.

The exposure is not only fairness, though that is real and belongs in front of counsel before the rule ships. It is that these populations cluster by geography and life stage, so a rule that catches them will look like it is catching fraud in particular markets. Build the suppression logic explicitly instead of letting the model absorb it. A recent-move indicator, a second observed residence, a documented address relationship inside the household, or a device history shorter than the window the rule requires should route to a separate queue rather than adding weight to the score.

Then measure the flag honestly. You never observe the outcome of an application you declined, so the confirmed-fraud rate among flagged-and-declined cases is unobservable and any precision estimate built from approved cases alone runs optimistic. The remedy is a holdout that approves a sample of flagged applications and books the losses deliberately, sized to be affordable and reviewed by the committee that signs the model card. Without it you cannot separate a rule that catches synthetics from a rule that declines people who moved. Keep permitted use attached to this work, since your DPA should say whether the residency signal is fraud prevention or marketing, and privacy-safe location licensing does not travel automatically between lanes.