Categorization precedence
Big Picture Money applies categorization sources in order. The first confident match wins.
Layer order
- Your rules — explicit patterns you created
- Built-in transaction-type patterns — payroll, transfers, credit card payments, and similar
- Your learned patterns — categories you assigned manually, stored as weighted histograms per merchant pattern
- Embedding neighbor — on-device similarity to your past corrections (when enabled)
- Merchant database — spine extraction plus token-scored lookup against the local merchants database. When no merchant alias matches, the app hashes the payee (and description when present) and looks up residual crowd-consensus rows in the same database — still below your own rules and learned patterns
- MCC — merchant category code from the bank file, when present
- Local classifier — on-device model trained from your history (when enabled)
Confidence notes
| Source | Typical confidence |
|---|---|
| Manual / rule | 100% |
| Learned pattern (dominant) | Share of your past choices (needs ≥ 80% of weight) |
| Merchant database | Token overlap score (needs ≥ 60%); crowd residual uses consensus confidence |
| MCC | Shipped default for that code |
| Embedding / classifier | Model score; low scores may enqueue review |
Ambiguous merchants
When you categorize the same merchant pattern under more than one category — for example Walmart as Groceries on some rows and Shopping on others — the app keeps both signals. The category with the highest weight wins when its share is at least 80%. When the top two choices are close, the app may show a lower-confidence suggestion.