subscription-detection

name: subscription-detection description: Rules for detecting recurring charges and subscriptions from transaction history. Reference this when identifying services the user pays for regularly.

Subscription Detection

Purpose

Provides rules for detecting recurring charges (subscriptions) from categorized transaction data. Identifies services the user pays for on a regular basis.

Before Detection — Check Learnings

Before applying automatic detection, read learnings/subscriptions.md:

1. Confirmed Subscriptions — If a merchant is listed here, always flag it as a subscription (skip automatic detection)
2. Not Subscriptions — If a merchant is listed here, never flag it as a subscription (even if it appears recurring)

Only proceed with automatic detection for merchants not found in either list.

Detection Criteria

A series of transactions qualifies as a subscription when all five criteria are met:

Frequency Detection

Determine the frequency by measuring the average interval between transactions:

If the interval doesn't match any frequency, the charges may be irregular purchases rather than a subscription.

Amount Tolerance

Allow ±5% variation in the recurring amount to account for:

• Price adjustments
• Currency conversion fluctuations
• Tax changes
• Service tier changes

Calculate: |amount_new - amount_avg| / amount_avg <= 0.05

If the amount changes by more than 5%, flag it as a potential price increase rather than disqualifying it as a subscription.

Annual Cost Calculation

Once a subscription is detected:

• Monthly: amount × 12
• Quarterly: amount × 4
• Yearly: amount × 1
• Weekly: amount × 52

Store in the annual_cost field of subscriptions.csv.

Subscription Status

Examples

Netflix Detection

Given these transactions in categorized.csv:

2025-11-01  Netflix  -149.00  Abonnementer
2025-12-01  Netflix  -149.00  Abonnementer
2026-01-01  Netflix  -149.00  Abonnementer

Step 1: Group by merchant

• Merchant "Netflix" has 3 transactions

Step 2: Check minimum occurrences

• 3 occurrences ≥ 3 minimum — passes

Step 3: Measure intervals

• Nov 1 → Dec 1 = 30 days
• Dec 1 → Jan 1 = 31 days
• Average = 30.5 days — matches monthly (28–31 days ±5 days)

Step 4: Check amount tolerance

• All amounts are -149.00 — 0% variance — passes

Step 5: Check recency (assuming today is Feb 1, 2026)

• Last charge: Jan 1, 2026
• Expected next: ~Feb 1, 2026
• Within expected interval + 7 days — passes

Result:

subscription_id: "sub-netflix-001"
merchant: Netflix
category: Streaming
amount: 149.00
frequency: monthly
annual_cost: 1788.00
first_seen: 2025-11-01
last_seen: 2026-01-01
status: active

Edge Cases

Variable Subscription Amounts

Some services (e.g., electricity, phone bills) have variable amounts:

• If the merchant is a known subscription (from the categorization skill) AND has regular frequency, detect it even if amounts vary more than 5%
• Store the most recent amount, not the average

Free Trial Conversions

A single charge from a service with no prior history is not a subscription yet. Wait for the minimum 3 occurrences before detecting.

Annual Subscriptions

Annual subscriptions may only have 1 occurrence within the data window. If the merchant is a known subscription service (e.g., Adobe CC annual plan), flag it as a potential subscription with frequency: yearly even with fewer than 3 occurrences.

Merged Households

If a service appears from multiple accounts (e.g., both lønkonto and budgetkonto), treat each account's charges separately. Don't merge cross-account charges into one subscription.

Stopped Subscriptions

If a previously detected subscription has no charge in the last expected interval + 7 days:

• Change status to paused
• Do not delete the row — the user may want to see it was detected

Learning from Corrections

When a user corrects subscription detection, record the correction in learnings/subscriptions.md:

If user says something IS a subscription (false negative):

1. Open learnings/subscriptions.md
2. Add a row to the "Confirmed Subscriptions" table:
   • Date: Today's date (YYYY-MM-DD)
   • Pattern: Normalized merchant pattern (e.g., *MERCHANT*)
   • Merchant: Normalized merchant name
   • Frequency: Expected frequency (monthly, yearly, etc.)
   • Note: Why this is a subscription

If user says something is NOT a subscription (false positive):

1. Open learnings/subscriptions.md
2. Add a row to the "Not Subscriptions" table:
   • Date: Today's date (YYYY-MM-DD)
   • Pattern: Normalized merchant pattern
   • Merchant: Normalized merchant name
   • Reason: Why this isn't a subscription (e.g., "one-time purchase", "irregular spending")

Example corrections:

## Confirmed Subscriptions
| 2026-01-15 | *HEADSPACE* | Headspace | monthly | User confirmed meditation app subscription |

## Not Subscriptions
| 2026-01-15 | *AMAZON* | Amazon | Regular purchases, not a subscription |

Related Skills

• See skills/categorization/SKILL.md for how merchants are identified
• See skills/data-schemas/SKILL.md for the subscriptions.csv structure
• See learnings/subscriptions.md for user corrections to subscription detection