Developers

Building a risk score from telecom signals

Telebase returns raw signals, not a score: carrier, country, number type, active status and, once live in your market, SIM swap recency. Most teams combine several of these into a single weighted number to feed a decision engine. Below is a starting framework, not a validated model. Treat the weights as a template to tune against your own fraud outcomes, not a fact about how risky each signal is.

Start with the signals, not a black box

A telecom signal is only useful once it is combined with the rest of your risk picture: device, IP, behavioural and document signals. Telebase does not publish a universal risk score, because the right weighting depends on your product, your fraud losses and your false-positive tolerance, all of which differ between a neobank, a crypto exchange and a lender. What follows is a worked example to adapt, not a number to copy.

A starting framework

The example below treats each signal as additive points towards an overall score, illustrative only:

• numberType is nonFixedVoip or voicemail: a candidate for the largest single weight in the example, since these types correlate with cheap, fast-provisioned numbers used for multi-accounting. See number type detection for fraud.
• active is false or null: a moderate weight in the example. A number that cannot be reached is a weak signal on its own but compounds with other flags.
• _meta.activeSource is VALID rather than LINE_STATUS: not a risk flag by itself, but a confidence discount. Treat an active result backed only by a format check as weaker evidence than one backed by a live line-status check, and weight it accordingly in whatever it feeds into.
• simSwap is SWAPPED and recent: the highest weight in the example once this field is live for your market, since a recent swap is the strongest single indicator of an in-progress account takeover attempt, and is exactly the signal to weight heavily immediately before a high-value transaction is authorised. SIM swap detection is launching for GB, DE, NL and FR and currently returns UNKNOWN for those markets; see the SIM swap API reference. Do not assign this weight based on an UNKNOWN result.

These are example weights for a template, not statistics derived from Telebase customer data or any published study. Telebase has not published, and does not claim, a validated accuracy figure for any scoring model built on its signals.

Tune it against your own outcomes

The only weights that matter are the ones validated against your own confirmed fraud and confirmed-good outcomes. Start with a simple additive model like the one above, log every decision and its eventual outcome, and adjust weights from there. A signal that looks high-risk in theory can behave differently in your specific customer base, so treat the starting framework as a hypothesis to test, not a conclusion.