Why Bayesian match scoring beats rule-based reconciliation
Rule-based reconciliation gives match / no-match. Bayesian scoring gives confidence with rationale — review the uncertain, auto-approve the rest.
Most reconciliation tools in the Indian market are built on the same foundation: rule-based exact-key matching. If the order ID in the marketplace report equals the order ID in the bank statement, it's a match. If not, it's an exception. This is simple, fast, and catastrophically brittle.
The brittleness shows up the moment reality deviates from the rule. Order IDs get truncated. Bank references are reformatted. A settlement lands with a slight amount difference because of a rounding rule you didn't know about. A missing leading zero turns a match into a non-match. Rule-based systems surface all of these as exceptions that a human has to review one at a time.
Bayesian probabilistic matching — specifically the Fellegi-Sunter framework that ReconPe's ACRE engine is built on — takes a fundamentally different approach. Instead of a binary match/no-match verdict, every candidate pair gets a confidence score from 0 to 100. The score is computed by summing weighted agreement on each field: order ID, amount, date, customer reference. Each field's weight is learned from your data, not hard-coded.
The practical consequence is that the 'exceptions' queue shrinks dramatically. Pairs with score above 90 are auto-approved — they're either a clean match or have only the kind of variance that indicates a legitimate small difference you've already decided to tolerate. Pairs below 40 are also auto-classified — they're clearly non-matches, and the system looks for better candidates. Only the 40-90 middle band gets surfaced for human review.
In practice, teams moving from rule-based to probabilistic reconciliation see their review queue drop by 70–90%. Not because the matching is less rigorous, but because the matching is more honest about uncertainty. A rule-based tool pretends every non-match is equally confusing; a probabilistic tool tells you exactly which cases deserve a human look.
This matters more than it sounds. Reconciliation bottlenecks at human review. If your tool surfaces 3,000 exceptions per month and only 300 of them actually need attention, you're paying analyst hours to clear noise. Probabilistic matching is the structural fix — not a bigger team, not faster analysts, but a system that only asks for help when it's actually uncertain.