How SUBSIGHT Works
A completely deterministic, local heuristic engine engineered to uncover recurring commitments without cloud models or bank scraping.
End-to-End Processing Flow
Every step performs deterministic transformations and logs measurable telemetry.
Transaction Data
Step 1Parses statement CSV or PDF tables, normalizes dates, signs, and debits.
Data Cleaning
Step 2Discards unparseable entries, scrubs noise tokens, and dedupes identical transactions.
Merchant Normalization
Step 3Strips city suffixes, matches regex aliases, and maps raw descriptors to canonical brands.
Pattern Analysis
Step 4Groups by merchant and clusters amounts using adaptive tolerance (max 20, 5% median).
Recurring Payment Detection
Step 5Measures delta regularity across weekly, monthly, quarterly, and annual horizons.
Subscription Identification
Step 6Synthesizes composite confidence scores (0.0–1.0) and assigns quality bands.
Financial Insights
Step 7Generates annual burdens, monthly equivalents, and surfaces review candidates.
Detection Signals & Weighting Model
Confidence (0.0 to 1.0) is a weighted linear composite of five deterministic vectors.
Merchant Regularity
Weight: 10%Ratio of transaction tokens mapping cleanly to known canonical entities versus raw strings.
Amount Consistency
Weight: 25%Variance across payments in the cluster, with automated price-revision merge for stepped jumps ≤ 25%.
Interval Regularity
Weight: 35%Standard deviation of day deltas relative to the median interval, tolerating single skipped months.
Occurrence Count
Weight: 20%Progressive confidence scaling with observation history (reaches full score at ≥ 6 payments).
Recency Horizon
Weight: 10%Validates that the last payment occurred within 1.6× of the expected cadence before decaying.
Clean intervals, tight amount spread, multiple occurrences, and verified canonical matching.
Consistent cadence with minor amount jitter or fewer recorded historical occurrences.
Possible recurring payment with higher interval irregularity. Surfaced with a review badge.
Probabilistic Nature of Detection
Detection is probabilistic. Raw statement narration varies widely across banks, merchants, and aggregators. SUBSIGHT never asserts that a detected charge is definitively an intentional subscription or that you have forgotten it.
Low-confidence and long-tenure results are always surfaced for your review. You maintain full control to confirm, ignore, or mark any detected candidate as “Not a subscription” at any time.