How Matching Works

Bank Reconciliation compares statement rows with transactions in one account. This reference explains automatic-match requirements, scores, safeguards, and settings.

Read it when a match needs explaining or before you change a matching setting.

Automatic-match requirement

The amount decides whether a pair exists. Everything else only decides how good it is.

Two rows can only be matched automatically if their amounts are exactly equal. Text never creates a pair - a perfect merchant name against a different amount produces nothing. Text and dates rank pairs the amount has already admitted.

The matcher is deterministic: the same statement, account state, and settings produce the same result. A missed match remains visible for review instead of being matched automatically.

The stages, in order

1. Loading the window

Before matching, the account is read for a window around the statement’s period - the statement’s own dates, plus the date tolerance, plus any extra padding. Transactions outside it are never considered.

Anything loaded purely as headroom, and never matched, is not shown to you at all. Transactions inside the window but outside the statement’s own dates appear as Outside period and are kept by default, because your statement makes no claim about them.

2. Pairs fixed by a reference

If your statement carries a bank transaction ID - OFX files do, as FITID - and a transaction in Actual was imported with that same ID, the two are pinned together before anything else runs. No scoring, no competition, confidence exact.

The statement’s own reference or authorisation number is tried the same way. Both are attempted, in that order: an OFX ID your account has never seen does not stop the row’s reference from being looked up.

This is the only tier that bypasses scoring entirely, because an identifier the bank issued is not evidence about a pairing - it is the pairing.

3. Finding candidates

For every remaining statement row, the matcher looks for transactions where both:

• the amount is exactly equal, and
• the date is within the date tolerance (7 days by default, either side).

Text is not consulted here at all. A transaction that fails either test is not a candidate, whatever its description says.

A second amount is also tried. When your bank prints the original amount of a foreign purchase in the description - ... EUR53.45 beside a converted charge - that figure is looked up too. A purchase abroad posts as the converted amount, while a transaction created by an SMS or automation often carries the original one, so the posted figure never matches and the original one matches to the cent.

This is not a tolerance. It is an exact match against a second figure the bank itself stated. It is scored lower than a posted-amount match, and it is rejected outright unless the merchant text also agrees - without that, a VAT ON SERVICE CHARGES EUR41.00 fee row would claim an unrelated EUR41.00 purchase purely by arithmetic.

4. Scoring

Every candidate gets a score out of 100:

Signal	Points	Notes
Amount exactly equal	50	Always awarded - a candidate cannot exist without it
Same day	25
1 day apart	20
2-3 days apart	14
4-7 days apart	7
Merchant text	0-25	The similarity score, scaled

A statement reference found verbatim in the transaction’s notes takes the full 25 text points rather than competing for them. For accounts built by automation the bank’s reference often sits in the notes, which makes it near-identity evidence.

An original-currency match has 12 points withheld, so a posted-amount match always outranks one.

The resulting score becomes a confidence label: high at 85 or more, medium at 65 or more, low below that. A pair scoring under the auto-match floor (60) is never matched automatically, whatever else is true about it.

5. Comparing the text

Which field in Actual holds your merchant text depends on how your transactions got there, so it is a setting rather than a guess:

Where your merchant text lives	How the transactions were created
The payee you curate	Entered by hand, or imported through Actual
The imported payee	A bank-sync or import path
The notes	SMS, n8n, Shortcuts or similar automation

Notes are compared by containment, in both directions. A note may legitimately hold the bank’s text plus your own words - NORTHGATE FUEL 933 #fuel | paid for Dad - and those additions must not cost anything. It also runs the other way: an automation that captured a truncated name (NORTHGATE FUEL 933) still fully matches a statement that spells it out in full (NORTHGATE FUEL 933 WESTPORT). Tags are stripped before comparing, since a tag is never bank text.

Payee fields are compared symmetrically instead, blending character-level and word-level similarity, because both sides are short.

A description too common to mean anything is rejected before it can match. A statement row reading FEE or PAYMENT would otherwise be contained in a large share of your notes and match almost anything. Two guards: very short single-word descriptions are refused outright, and so is any description whose every word is common across the loaded window. The second only applies once there are at least ten transactions to judge frequency against.

6. Assigning the pairs

Every candidate pair from every row competes in one sorted pass, best first. A transaction can be claimed only once.

This matters: matching each row in turn against its own best option lets the first row take a transaction a later row needed more. One global pass gives each transaction to the row with the strongest claim on it.

The order is best score first, then the closest date, then the strongest text - so ties never fall to whichever row happened to be listed first.

Cases that require your decision

The matcher stops rather than guess. Three distinct situations, each shown differently:

Several candidates

Two plausible candidates whose scores are within 8 points of each other are both offered, and neither is chosen. A 94 and a 91 are not a decision anyone should make silently.

Two refinements stop this being over-cautious:

• When the text plainly separates them, the better one is taken. Two fee rows differing only in the amount they quote can score 100 and 95 while their text scores 1.00 and 0.79 - the evidence separates them clearly, and the scores are close only because text carries a quarter of the budget. This applies only when nothing else contradicts: a candidate whose text agrees better while its date agrees worse is genuinely ambiguous, and stays your decision.
• When rivals are interchangeable - identical score, identical date distance, identical text - asking which one it is has no answer, because either choice produces the same reconciliation. One is taken and the rest fall to other rows.

Likely duplicates

When the rivals are near-identical to the winner rather than merely close - within 3 points, both matched on the posted amount, dates within a day - the loser is not a competing match, it is the same transaction recorded twice.

This is decided after the whole assignment is known, because it is a count mismatch rather than a resemblance. Three statement rows against three identical transactions is a clean correspondence, not triplication. A rival still unclaimed at the end is surplus; one that another row took was simply another real transaction.

Nothing confident enough

When a row’s best candidate falls below the auto-match floor, the plausible few are offered so you can choose - at most five, and none scoring below 60. Below that a candidate carries nothing but a coincidence of amount and a date in the window, and offering it would waste your attention.

If nothing clears even that, the row is left alone and offered as Not in Actual, which is a better answer than a bad guess.

The leftover passes

Once every amount-based avenue is spent, what remains on both sides is compared again - this time allowing the amounts to disagree. Nothing here is ever matched automatically at any size. The amounts differ, and only you can say which figure is right.

Tier 1 - amounts that could plausibly be the same transaction. The text must agree strongly (0.75 or better) and the amounts must be within 25% of each other. This is the tier that catches a conversion done wrong before it reached your budget.

Tier 2 - same merchant, same day, amounts unrelated. Text must agree very strongly (0.80) and the dates must be within one day, but the amounts are ignored entirely.

Neither tier reaches further than your date tolerance. They may relax the amount - that is what they are for - but the window is yours. Set it to 0 and nothing is offered across a day, in any tier; set it to 3 and the second tier still stops at its own one day, because that is narrower.

Ignoring the amount is only defensible once everything else is spent, and “spent” is the load-bearing word. Both sides of a tier-2 pairing must be untouched by tier 1. A row tier 1 could pair with a plausible amount must never also be offered an implausible one: Rivera Cafe EUR129.00 posting −131.34 was once offered a −13.71 transaction purely because both said “Rivera Cafe” on the 21st - while −13.71 was the match for the Rivera Cafe EUR14.00 row two lines away, at the same conversion rate.

One pairing, or a group

Both tiers work on connected components - the set of rows and transactions related to each other, directly or through a shared neighbour. That is exactly the set a decision on any one of them affects, and its size decides what you are shown:

• One row and one transaction → a pairing. Nothing else could be meant, so relating them costs nothing and finding it by hand would cost you.
• Anything larger → a group needing pairing. Every row sees its own candidates, closest amount first, and settling one frees the rest.

Within a leftover pair, closeness of amount is scored: identical-but-for-rounding ranks near the top, a pair at the 25% limit ranks at the bottom. Before this, three CAREEM RIDE rows competing for one −9.74 transaction scored 50, 50 and 45 while their gaps were 3.8%, 20.0% and 12.5% - the true pairing ranked no higher than a row five times further away.

What it never does

• It never bridges an amount gap with text. No merchant name, however perfect, creates an automatic match against a different amount.
• It never matches two transactions to one statement row, or one transaction to two rows.
• It never chooses between candidates it cannot separate. That is your decision, and the tool says so rather than guessing.
• It never writes anything. Matching produces a comparison. Every write comes from a decision you took and approved at Review.
• It never overwrites your decisions. Re-running the matcher rebuilds the comparison; the choices you made are a separate layer on top.

What you can change

Five controls, under Matching on the Reconcile screen. They are saved per account as a reconciliation profile, and changing one takes effect when you choose Re-run.

Control	Default	What it does
Where the merchant text is	All, best match	Which Actual field is compared - the payee you curate, the imported payee, the notes, or whichever of them agrees best.
Match transactions within (days)	7	How far apart a statement row and a transaction may be posted and still pair. Actual’s own importer uses the same 7 days.
Look beyond the statement period (days)	3	How far outside the statement’s dates transactions are loaded. They are shown separately and never counted as missing.
Ignore #tags in notes	On	Workflow tags such as #API are not the bank’s text, so comparing them weakens every score.
Match foreign transactions on their original amount	On	Try the foreign-currency figure the bank printed, as a second exact amount.

The last two sit under Advanced matching options.

What is fixed

Everything else in this page is built in. It is documented so a pairing can be explained, not because there is a control for it - if a figure below surprises you, the answer is a decision on the row rather than a setting to change.

Behaviour	Value
Auto-match floor - the score a pair must reach to be taken without asking	60
Ambiguity margin - how close a runner-up must be for the matcher to refuse to choose	8 points
Text separation margin - how much better the winner’s text must be to settle a close score	0.15
Duplicate margin - how near-identical a rival must be to be called a duplicate	3 points
Candidates offered when nothing clears the floor	at most 5, none below 60
Amount mismatch review - text floor	0.75
Amount mismatch review - widest gap offered	25%
Either review tier - widest date gap offered	your date tolerance
Same merchant and date - text floor	0.80
Same merchant and date - date tolerance	1 day, or your window if it is narrower
Original-amount match - points withheld	12
Original-amount match - text corroboration required	0.60

Loosening any of these would widen what you are offered, never what is decided for you: the exact-amount rule and the auto-match floor govern that, and neither leftover tier can match anything on its own.