Recurring Charge Detector

Recurring Charge Detector

Table of Contents

• Overview
• Input contract
• Cadences
• Workflow
• Status transitions
• Output contract
• Guardrails

Overview

A subscription audit is mostly the same question asked many ways: which charges arrive on a clock? This skill clusters historical transactions by merchant + cadence + typical amount, requires ≥ 3 confirming occurrences before promotion, and tracks each cluster's lifecycle (active → suspected_dormant → cancelled).

It computes annualized cost for each active recurring entry — the single most useful number in a subscription audit ("you are spending $191 a year on Netflix").

Input contract

The caller provides:

• transactions — array with {id, account_id, date, merchant, amount_cents, category, subcategory}.
• existing_recurring — current recurring.json state.
• lookback_days (default 540, ~18 months) — only transactions within this window are clustered.
• today — ISO date for status calculations.

Cadences

If the inter-arrival times don't fit any of the above with the listed tolerance, mark cadence: "irregular" and do not promote.

Workflow

Recurring Detection Progress:
- [ ] Step 1: Group transactions by (merchant, account_id)
- [ ] Step 2: For each group with >= 3 transactions, sort by date
- [ ] Step 3: Compute inter-arrival deltas; classify cadence
- [ ] Step 4: Check amount stability (±10% or fixed)
- [ ] Step 5: Promote to active if cadence + amount + occurrences pass
- [ ] Step 6: Compute next_expected_date and annualized cost
- [ ] Step 7: Diff against existing_recurring; emit transitions
- [ ] Step 8: Detect dormant (missed expected date by > tolerance)

Step 1 — Group

Group by (merchant, account_id). Subscriptions can split across accounts; treat each account as its own series. (One Netflix charge that moved from credit card A to credit card B will appear as two clusters; the agent layer reconciles them.)

Step 2 — Sort and threshold

Drop groups with < 3 transactions. They cannot be promoted yet — surface them as candidates[] with evidence_count so the user knows what's accumulating.

Step 3 — Cadence classification

Compute deltas d_i = date[i+1] - date[i] in days. Classify cadence using the table above. Acceptance rule:

• ≥ 80% of deltas fall within the cadence's tolerance band.
• Allow up to one large delta (> 2× nominal) only if it's followed by re-anchoring at the original cadence — this is normal for "skipped a month" patterns.

If multiple cadences are plausible, pick the one with tighter fit (smallest standard deviation of deltas).

Step 4 — Amount stability

Compute mean amount μ and standard deviation σ of the cluster's amounts.

• σ / |μ| ≤ 0.05 → "fixed" (e.g., Netflix $15.99 every month).
• σ / |μ| ≤ 0.20 → "variable_low" (utilities, where amount drifts seasonally).
• σ / |μ| ≤ 0.40 → "variable_high" (e.g., grocery delivery — flag but allow).
• σ / |μ| > 0.40 → reject as recurring (probably a misclustered general merchant).

Record amount_cents_typical = round(μ) and amount_cents_variance = round(σ).

Step 5 — Promotion

Promote to status: "active" if:

• occurrences ≥ 3.
• cadence is one of the named cadences (not irregular).
• amount stability is fixed or variable_low (or variable_high with a warning).

Step 6 — Forward projection

next_expected_date = last_seen + cadence_interval. Update with each new occurrence.

annualized_cost_cents:

annualized_cost_cents = amount_cents_typical × multiplier. For inflows (paychecks), this is annualized income; for outflows, annualized cost.

Step 7 — Diff against existing

For each existing recurring.json entry, find the matching new cluster (by merchant + account_id + cadence). Update fields:

• last_seen, occurrences, amount_cents_typical, next_expected_date.
• If amount drifts > 15% from prior amount_cents_typical → emit event: "amount_changed" with old and new.

For new clusters not in existing → emit event: "new_recurring".

Step 8 — Dormant detection

For each active entry where today > next_expected_date + tolerance and no matching transaction exists in the window:

• First miss → status: "suspected_dormant", event: "missed_expected".
• Two consecutive misses → propose status: "cancelled" with event: "likely_cancelled".

The agent layer confirms before flipping to cancelled.

Status transitions

candidate (< 3 occurrences)
   ↓ [3rd occurrence on cadence]
active
   ↓ [missed expected date]                 ↓ [user confirms]
suspected_dormant ─ [resumed] → active     paused
   ↓ [missed twice in a row]
likely_cancelled
   ↓ [user confirms]
cancelled

Output contract

{
  "active": [
    {
      "id": "rec_netflix",
      "merchant": "Netflix",
      "account_id": "acc_cc_001",
      "category": "entertainment.streaming",
      "amount_cents_typical": 1599,
      "amount_cents_variance": 0,
      "cadence": "monthly",
      "first_seen": "2024-03-14",
      "last_seen": "2026-01-14",
      "occurrences": 23,
      "next_expected_date": "2026-02-14",
      "annualized_cost_cents": 19188,
      "status": "active"
    }
  ],
  "candidates": [
    {
      "merchant": "Blue Bottle Coffee",
      "account_id": "acc_cc_001",
      "evidence_count": 2,
      "needed": 3,
      "reason": "below promotion threshold"
    }
  ],
  "events": [
    { "type": "new_recurring", "id": "rec_chatgpt", "evidence_count": 3 },
    { "type": "amount_changed", "id": "rec_netflix",
      "old_cents": 1499, "new_cents": 1599, "delta_pct": 6.7 },
    { "type": "missed_expected", "id": "rec_amazon_prime",
      "expected_date": "2026-01-10", "today": "2026-01-20",
      "suggested_status": "suspected_dormant" }
  ],
  "summary": {
    "active_count": 27,
    "annualized_cost_total_cents": 487200,
    "candidates": 4,
    "dormant": 1
  }
}

Guardrails

• Three is the minimum. Two same-merchant charges a month apart are not enough — the second could be a one-off subscription trial.
• Cluster by merchant, not by description. Use the merchant field from the categorizer, not description_raw. Different rendering of the same merchant string would otherwise split a real recurring into two phantom ones.
• Inflows count too. Paychecks are recurring income.salary; track them. The cash-flow forecaster needs both sides.
• Variable utilities. Allow variable_low for utilities. Do not propagate variance noise into the forecaster — use amount_cents_typical (the mean).
• Tax-loss-harvesting / rebalancing trades can superficially look recurring (same security, regular pattern). Treat brokerage account transactions with category: savings_investment.* as exempt from recurring promotion unless the cadence is annual or longer.
• Dormant ≠ cancelled. Surface dormant; let the user confirm cancellation. Some merchants skip a month legitimately.