Every charge, sorted before you look at it.

Transactions reach feels.money through FeelConnect, the financial data platform we built underneath it. That covers banks, credit unions, payment apps, brokerages, and crypto exchanges and wallets. Connections are read-only: we can see balances and transactions, and there is no code path by which we could move money, initiate a payment, or change anything at your institution. The security page goes through that in detail.

If an institution is not covered yet, importing a CSV, OFX or QFX file gets you to the same place. Imported transactions are normalized into the same shape as synced ones, which means budgets, subscription detection, insights and the health score all behave identically. There is no second-class data path.

Once a transaction lands it is cached locally so that the app is fast and works even when an upstream provider is having a bad day. Amounts are stored as exact decimals and every calculation that matters runs through integer-cent arithmetic, because floating point quietly loses pennies and a personal finance app that loses pennies deserves to be uninstalled.

The primary categories are Income, Transfer, Food & Drink, Groceries, Shopping, Entertainment, Travel, Transport, Housing, Utilities, Subscriptions, Health & Fitness, Fees, and Other. Two of those — Income and Transfer — are not spending, and are excluded from budgets and spending analytics by design.

Food & Drink and Groceries are separate on purpose. Lumping the weekly shop together with three months of takeaway hides the single most actionable pattern in most people's spending, and a tracker that hides it is doing you no favours. The same logic splits Subscriptions out of Entertainment: one is a recurring commitment you can cancel, the other is a decision you make on a Friday night.

Categorization happens upstream, in FeelConnect, as transactions are ingested — so a transaction is already carrying a primary and a detailed category by the time it reaches your account. Nothing has to be classified while you wait, and nothing is left uncategorized until you next open the app.

The classifier itself is a table of explicit patterns — several hundred of them — tested against the description and merchant fields combined. The order of that table is the interesting part, because it encodes what actually matters:

1. Transaction-type descriptors first. Payroll, direct deposits, ATM fees, overdraft charges, wire transfers, credit-card payments, savings transfers. These outrank brand names deliberately — a card payment to a bank is a transfer, not a purchase from that bank, and getting this backwards is how budgets end up double-counting.
2. Then known merchant brands, most specific first, each carrying its own confidence value.
3. Then generic keyword fallbacks — lower confidence, and honest about it.

A model is never required for a category to exist, which is the whole point of building it this way round. Where FeelConnect's AI enrichment is used, it refines what the rules left ambiguous and never replaces them: a deterministic answer is computed first, and if the AI backend is slow, down, or returns nonsense, that deterministic answer ships anyway. Every enrichment response records which engine produced it — rules, ai, or ai+rules — so provenance is inspectable rather than mysterious. The same principle runs through the coach.

Anything a model returns is clamped back onto the taxonomy before it is stored. An invented category cannot enter your data: an unrecognized primary becomes Other, and an unrecognized subcategory falls back to the first valid one for its primary.

Bank descriptors are written for settlement systems, not for you. They arrive padded with store numbers, terminal ids, processor prefixes, city codes and stray punctuation. Before anything is grouped or displayed, each descriptor is normalized into a stable key: uppercased, runs of digits stripped, punctuation collapsed, whitespace squeezed, truncated to a sensible length.

That key does two jobs. It produces the clean, title-cased merchant name you see in the app, and — more importantly — it makes the same merchant collapse into one entity across months even when the store number changes every visit. Without that step, your top-merchants list would be forty variations of the same coffee shop.

Transactions are grouped by normalized merchant and direction, then tested on two axes. First, cadence: the median gap between charges has to fall inside a known window — weekly, biweekly, monthly, quarterly or annual — and most of the individual gaps have to sit near that period, not merely average out to it. Second, amount stability: the variation in the charge amounts has to stay under a ceiling, because a merchant you visit every week for wildly different sums is a habit, not a commitment.

Streams that pass both tests become recurring items, and the ones that are stable enough and cheap enough get classified as subscriptions rather than bills. The exact thresholds, and what the app does with the result, are laid out on the subscription tracker page.

Categorized transactions make a handful of deterministic detectors possible, and each one exists because it answers a question people actually ask about their own money:

• Lifestyle inflation — is the baseline drifting upward month over month, independent of any single big purchase?
• Impulse clusters — did several discretionary purchases happen inside the same short window?
• Dining trend — is restaurant and delivery spending moving in one direction over time?
• Largest purchases — the specific transactions behind an unusual month.
• Month-over-month comparison and top merchants — where the money concentrated.

None of these produce a score, and none of them scold. They are observations with the numbers attached, so you can decide whether a pattern is a problem or simply the shape of a particular month. If you want the interpretation, that is what the monthly review guide is for.