Categorization

Purpose

Provides Claude with Danish merchant knowledge and categorization rules for classifying bank transactions into spending categories.

Categories

15 Danish spending categories with subcategories:

Merchant Pattern Database

Map raw transaction text patterns to normalized merchant names and categories. Patterns use * as wildcard.

Dagligvarer (Groceries)

Transport

Abonnementer (Subscriptions)

Restauranter (Dining)

Shopping

Bolig (Housing)

Sundhed (Health)

Underholdning (Entertainment)

Rejser (Travel)

Personlig pleje

Uddannelse (Education)

Normalization Rules

Before matching patterns, normalize the raw transaction text:

1. •Convert to uppercase for matching
2. •Replace Danish characters: Ø → OE, Æ → AE, Å → AA (for pattern matching only — preserve originals in raw_text)
3. •Collapse multiple spaces into one
4. •Trim leading/trailing whitespace

Transaction Type Detection

Use the transaction description prefix to identify the payment method:

Matching Order

When categorizing a transaction, check these sources in order — stop at the first match:

1. •learnings/categorization.md — Check for user corrections learned from previous sessions (confidence 1.0). These take precedence over all other rules.
2. •learnings/merchants.md — Check for merchant name aliases to normalize the raw text before pattern matching.
3. •merchant-overrides.csv — Check for a matching raw_pattern (confidence 1.0). These are learned from previous user corrections within the same session.
4. •Static pattern database — Check the merchant pattern tables above (confidence 1.0 exact / 0.8 partial).
5. •Intelligent classification — No pattern match, but Claude can infer category from transaction context (confidence 0.5–0.7).
6. •Fallback — Cannot determine category — assign to "Andet" (confidence 0.0).

Confidence Scoring

Assign a confidence score to each categorization:

When manual_override is TRUE for a transaction, always use the user's category regardless of what patterns suggest.

Learning from Corrections

When a user manually corrects a transaction's category, the correction should be saved to learnings/categorization.md so future sessions automatically apply the learned rule.

How to record a categorization learning:

1. •Open learnings/categorization.md
2. •Append a new row to the Learnings table with:
   • •Date: Today's date (YYYY-MM-DD)
   • •Pattern: The normalized raw text pattern to match (uppercase, wildcards allowed)
   • •Merchant: The normalized merchant name
   • •Category: The user-corrected category
   • •Subcategory: The user-corrected subcategory
   • •Context: Brief note about why this correction was made

Example learning entry:

| 2026-01-15 | *NETFLIX* | Netflix | Underholdning | Streaming | User: "Netflix er underholdning, ikke abonnement" |

Also update merchant-overrides.csv for the current session:

1. •Take the corrected transaction's raw_text from transactions.csv
2. •Normalize it: uppercase, trim whitespace, collapse multiple spaces
3. •Use this as the raw_pattern
4. •Use the user-corrected merchant, category, and subcategory from categorized.csv
5. •Set created_at to the current timestamp
6. •Only append if no row with the same raw_pattern already exists

Ambiguous Merchants

• •Netto: Always Dagligvarer, even though they sell non-grocery items
• •7-Eleven: Dagligvarer (most purchases are food/drink)
• •Amazon: Default to Shopping unless description contains "Prime" or "Kindle" (then Abonnementer or Uddannelse)
• •Normal: Shopping → Andet (sells mixed categories)
• •Matas: Sundhed → Apotek (primarily health/beauty)
• •Wolt: Restauranter → Takeaway unless the Wolt+ subscription charge (then Abonnementer)

Examples

Example 1: Clear Match

raw_text: "NETTO FO 1234 KØBENHAVN"
normalized: "NETTO FO 1234 KOEBENHAVN"
pattern match: *NETTO* → Netto
result: category=Dagligvarer, subcategory=Supermarked, merchant=Netto, confidence=1.0

Example 2: PBS Subscription

raw_text: "PBS FITNESS WORLD"
normalized: "PBS FITNESS WORLD"
pattern match: *FITNESS*WORLD* → Fitness World
result: category=Abonnementer, subcategory=Fitness, merchant=Fitness World, confidence=1.0, is_recurring=TRUE

Example 3: Intelligent Classification

raw_text: "RESTAURANT COFOCO KBH"
normalized: "RESTAURANT COFOCO KBH"
no pattern match — but "RESTAURANT" prefix suggests dining
result: category=Restauranter, subcategory=Restaurant, merchant=Cofoco, confidence=0.6

Example 4: Income

raw_text: "Løn fra Arbejdsgiver ApS"
prefix match: "Løn fra" → salary
result: category=Indkomst, subcategory=Løn, merchant=Arbejdsgiver ApS, confidence=1.0