Loyalty glossary · 7. Data and analytics (31)
Probabilistic Matching
Probabilistic matching is the process of linking loyalty programme member records to transactions or other data when no unique identifier exists, by scoring the likelihood that two records refer to the same person. It tolerates imperfect data such as misspelled names or changed addresses.
Loyalty programmes rarely hold a single clean key across every touchpoint. A member books a stay, buys through a partner, and scans in store, but the transaction often carries no loyalty number. Exact matching on one field such as email fails whenever a member uses a work address, a typo, or a different surname. Probabilistic matching fills that gap by scoring records for likely identity rather than requiring equality.
The method works by assigning agreement weights to fields like surname, postcode, date of birth and phone number. Each pair of records gets a probability score, and a threshold decides which pairs count as the same member. The threshold is not a fact of nature, it is an operator choice. Too high and the programme misses real matches, leaving points unattributed. Too low and it links two different people, merging their histories. A false positive corrupts a member record in a way that a false negative does not.
Work the arithmetic on a realistic scale. A programme holds 1 billion points in transactions without a linked member ID. Exact matching on email alone links 40 percent, so 600 million points remain unattributed. Probabilistic matching using name, postcode and phone lifts the match rate to 70 percent, attributing an additional 300 million points. That is a 30 point improvement, and the residual 300 million points are still unassigned, which is 30 percent of the total. The numbers make the trade off visible: better matching recovers accrual liability, but the remaining gap is not small.
The trap is tuning the threshold for convenience rather than for error cost. A programme under pressure to lift its active member rate or to reduce unredeemed point liability may lower the threshold, accepting more false positives. That inflates matched transactions and makes the data look healthier than it is. The correct approach sets the threshold by comparing the cost of a wrong link to the cost of a missed link, and those costs are never equal. Awarding points to the wrong member is a financial error with a complainant, while missing a link merely leaves data incomplete.
Probabilistic matching produces links that are claims, not facts, but downstream systems routinely treat them as certain. A segmentation model, a fraud score or a personalised offer will consume a merged record as if the join were exact, discarding the probability. The result is confident decisions built on uncertain identity. The only defensible practice is to carry the match probability through the pipeline, not collapse it to a yes or no flag.
Because the link is a claim, it must be auditable. Keep the raw records, the scoring model, the threshold and the date of the match. When a member challenges a transaction or a point balance, the programme must explain why two records were joined. A probabilistic match with no audit trail is just a guess with a number attached, and it fails every dispute. Probabilistic matching is defensible only when the evidence behind each link can be shown.