How well it reads
Measured on every release, on the same test sets, for every category.
- Whole line right
- 85.5%
- Category right
- 86.6%
- Merchant right
- 95.1%
- Precision
- 99.2%
- Coverage
- 87.3%
- F1
- 92.5%
On 963 transactions it never saw, measured 5 Nov 2026, under the balanced setting. Each category is held to 95% precision from 20 answers on a test set, and one below it is marked; under 20, its precision is shown as a count. Under the complete setting, which gives every line a category as services that categorise everything do, 92.4% of all lines are right.
The setting
A lender chooses how far to trust Statement. Strict gives only what a rule reads. Balanced, the default, adds the bank’s merchant code where a connection sent one, the business a register of Australian places lists in the line’s suburb, and a modelled reading where it keeps its category at 95% precision or better. Broad holds each category to 90%. Complete gives every line a category, guessing where nothing read it, so Statement can be compared with services that categorise everything; its guesses are marked and it has no precision floor. The thresholds were set on lines kept apart from these test sets.
| Test set | Precision | Coverage | All lines right | F1 | Weakest category | Modelled answers |
|---|---|---|---|---|---|---|
| Strict The rules alone. Every answer is a name or a kind of business a rule read. | ||||||
| Transactions it never saw | 99.4% | 74.1% | 73.7% | 84.7% | Shopping 96.4% | — |
| Real payee names | 100.0% | 59.8% | 59.8% | 74.8% | Transport and fuel 100.0% | — |
| Balanced (default) The rules, the bank’s merchant code where a connection sent one, a register of Australian places, and a modelled reading where it keeps its category at 95% precision or better. | ||||||
| Transactions it never saw | 99.2% | 87.3% | 86.6% | 92.5% | Health 97.1% | 24 of 26 right |
| Real payee names | 100.0% | 70.1% | 70.1% | 82.4% | Transport and fuel 100.0% | 12 of 12 right |
| Broad More answers. The bank’s merchant code where a connection sent one, a register of Australian places, and a modelled reading where it keeps its category at 90% precision or better. | ||||||
| Transactions it never saw | 99.0% | 89.6% | 88.7% | 93.5% | Health 97.1% | 44 of 48 right |
| Real payee names | 99.4% | 71.3% | 70.9% | 82.7% | Entertainment and subscriptions 94.9% | 21 of 26 right |
| Complete Every line gets a category. Read as broad reads, and where nothing read a line, the model’s best guess, marked as one: money out as a spending line, money in as a transfer. Never income, a commitment or gambling on a guess, and not held to a precision floor. | ||||||
| Transactions it never saw | 92.4% | 100.0% | 92.4% | 92.4% | Eating out and takeaway 83.3% | 44 of 48 right |
| Real payee names | 75.3% | 100.0% | 75.3% | 75.3% | Eating out and takeaway 43.0% | 21 of 26 right |
- Transactions it never saw
- Real payee names
The line, its category and its merchant
A line is two answers: what the money was for, and who was paid. Each is measured on its own, and the line as a whole.
Transactions it never saw: each line wholly right, then its category and its merchant
| Setting | Whole line right | Category right | Merchant right |
|---|---|---|---|
| Strict | 72.6% | 73.7%99.4% precise | 95.1%99.2% precise |
| Balanced (default) | 85.5% | 86.6%99.2% precise | 95.1%99.2% precise |
| Broad | 87.5% | 88.7%99.0% precise | 95.1%99.2% precise |
| Complete | 91.1% | 92.4%92.4% precise | 95.1%99.2% precise |
Whole line: the category right and, where the line has a merchant, the merchant named right, of all 963 lines. Category: of every line, a line it declines counted as wrong. Merchant: of the 366 lines whose label names a business. Precise: of the answers it gives, the share right.
Real payee names: each line wholly right, then its category and its merchant
| Setting | Whole line right | Category right | Merchant right |
|---|---|---|---|
| Strict | 43.1% | 59.8%100.0% precise | 43.8%98.8% precise |
| Balanced (default) | 52.2% | 70.1%100.0% precise | 52.9%98.7% precise |
| Broad | 52.2% | 70.9%99.4% precise | 52.9%98.7% precise |
| Complete | 52.3% | 75.3%75.3% precise | 52.9%98.7% precise |
Whole line: the category right and, where the line has a merchant, the merchant named right, of all 1,143 lines. Category: of every line, a line it declines counted as wrong. Merchant: of the 1,143 lines whose label names a business. Precise: of the answers it gives, the share right.
Counted by household
779 of the 963 transactions it never saw come from one source, so a figure counted by line is mostly that source. Counted by household, each source weighs at most 20 lines and that one is 16% of the figure. Every line still counts.
| Setting | Precision by line | Precision by household | Coverage by line | Coverage by household |
|---|---|---|---|---|
| Transactions it never saw 14 sources, each counted as at most 20 lines | ||||
| Strict | 99.4% | 97.7% | 74.1% | 81.6% |
| Balanced (default) | 99.2% | 97.7% | 87.3% | 85.3% |
| Broad | 99.0% | 97.7% | 89.6% | 86.5% |
| Complete | 92.4% | 88.1% | 100.0% | 100.0% |
Test sets
| Test set | Precision | Coverage | Recall | F1 | F1 since first release |
|---|---|---|---|---|---|
| Transactions it never saw | 99.2% | 87.3% | 86.6% | 92.5% | +24.5 pts |
963 labelled transactions from public datasets, held out while the rules were written. The labels follow this taxonomy; 41 made before four of its decisions were brought into line, as the bank’s-words set’s were. Payee named on 54.7%. Measured 5 Nov 2026. | |||||
| Real payee names | 100.0% | 70.1% | 70.1% | 82.4% | +50.8 pts |
1,143 Australian businesses from public registers and map data, each with a right answer. Payee named on 52.9%. Measured 5 Nov 2026. | |||||
| The bank’s own words not independent | 99.8% | 81.2% | 81.0% | 89.5% | +35.5 pts |
680 lines from 14 banks. The rules were written from them, so this checks nothing was lost; it does not measure accuracy. 46 lines that should not be read were left unread, except 3. Measured 5 Nov 2026. | |||||
| BPAY billers not independent | 98.7% | 65.9% | 65.0% | 78.4% | +71.9 pts |
6,308 billers’ short names as a bank prints them, against their full registered names. Agreement, not correctness. Measured 5 Nov 2026. | |||||
- Transactions it never saw 92.5%
- Real payee names 82.4%
- The bank’s own words 89.5%
- BPAY billers 78.4%
By category
F1 and the number of lines in each test set. Open a category for a few of its lines as statements write them, for precision, coverage and recall, and for what the wrong answers were labelled.
| Category | Transactions it never saw F1 · lines | Real payee names F1 · lines | The bank’s own words F1 · lines | |||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Salary or wages | too few · 6 | — | 92.6% · 29 | |||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
Examples, precision, coverage and recall
| ||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
| Government benefit | — | 90.9% · 24 | too few · 9 | |||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
Examples, precision, coverage and recall
| ||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
| Other income | too few · 8 | — | 66.7% · 64 | |||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
Examples, precision, coverage and recall
| ||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
| Loan or advance received | — | — | too few · 2 | |||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
Examples, precision, coverage and recall
| ||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
| Rent | too few · 4 | too few · 26 | too few · 15 | |||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
Examples, precision, coverage and recall
| ||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
| Mortgage repayment | too few · 3 | 97.6% · 21 | too few · 2 | |||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
Examples, precision, coverage and recall
| ||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
| Strata and body corporate | — | — | — | |||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
Not measured: Owners corporations are in no public register, so there is nothing independent to test against. The words were checked against 715 strata billers’ names, the same names they were written from: a check, not a measurement. Examples
| ||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
| Utilities | too few · 9 | 88.7% · 84 | too few · 14 | |||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
Examples, precision, coverage and recall
| ||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
| Groceries | 96.3% · 169 | 91.2% · 31 | 98.0% · 25 | |||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
Examples, precision, coverage and recall
| ||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
| Transport and fuel | 95.8% · 136 | 78.7% · 225 | 97.2% · 37 | |||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
Examples, precision, coverage and recall
| ||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
| Insurance | too few · 8 | 85.2% · 66 | 95.2% · 22 | |||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
Examples, precision, coverage and recall
| ||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
| Health | 90.4% · 39 | 71.9% · 41 | too few · 19 | |||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
Examples, precision, coverage and recall
| ||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
| Education and childcare | too few · 3 | too few · 18 | too few · 12 | |||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
Examples, precision, coverage and recall
| ||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
| Phone and internet | too few · 4 | too few · 6 | 100.0% · 20 | |||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
Examples, precision, coverage and recall
| ||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
| Loan repayment | — | 70.1% · 113 | too few · 9 | |||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
Examples, precision, coverage and recall
| ||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
| Credit card payment | — | too few · 4 | too few · 16 | |||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
Examples, precision, coverage and recall
| ||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
| Buy now pay later | too few · 3 | said 2 | too few · 9 | |||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
Examples, precision, coverage and recall
| ||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
| Child support paid | — | — | — | |||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
Not measured: One payee, Services Australia’s Child Support program, so there is no register of payees to test against; the line is read by the program’s own name. Checked on every corpus here for a collision and found none. Examples
| ||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
| Eating out and takeaway | 91.7% · 253 | 85.5% · 190 | 93.1% · 31 | |||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
Examples, precision, coverage and recall
| ||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
| Shopping | 80.4% · 120 | 86.1% · 193 | 81.7% · 55 | |||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
Examples, precision, coverage and recall
| ||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
| Entertainment and subscriptions | 92.2% · 62 | 86.1% · 49 | 96.2% · 27 | |||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
Examples, precision, coverage and recall
| ||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
| Travel | too few · 11 | too few · 8 | too few · 13 | |||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
Examples, precision, coverage and recall
| ||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
| Donations and charity | — | — | — | |||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
Not measured: Charities are in a public register, the ACNC’s, and that register is where the names this reads them by come from, so it cannot also measure them; no other corpus here names a charity a household paid. Added on 29 October 2026. Examples
| ||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
| Gambling | — | too few · 1 | too few · 3 | |||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
Examples, precision, coverage and recall
| ||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
| Cash withdrawal | 98.5% · 33 | 69.7% · 43 | 90.3% · 34 | |||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
Examples, precision, coverage and recall
| ||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
| Transfer | 96.8% · 80 | — | 87.5% · 140 | |||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
Examples, precision, coverage and recall
| ||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
| External transfer | — | — | said 23 | |||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
Not measured: People are in no public register, by design. Read from the bank’s own wording and a given name, and checked on the bank-wording lines and on transfers between common Australian names. Examples
| ||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
| Bank fees and interest | too few · 6 | — | 95.7% · 60 | |||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
Examples, precision, coverage and recall
| ||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
| Tax, rates and government charges | too few · 6 | — | too few · 13 | |||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
Examples, precision, coverage and recall
| ||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
Releases
Registers read again 5 Nov 2026 · ff86718
The public registers read again for November. Every figure held.
Transactions it never saw F1 92.5% +0.0 ptsReal payee names F1 82.4% +0.0 pts
Registers read again 5 Nov 2026 · 2b2015a
The public registers, places, names, web addresses and bank logos read again for November. Every figure held.
Transactions it never saw F1 92.5% +0.0 ptsReal payee names F1 82.4% +0.0 pts
Bank formats read through 29 Oct 2026 · 868fd65
HSBC’s terminal number and country written onto a suburb, and ING’s and Bendigo’s purchase words before a merchant’s name, no longer hide the business from the register of places.
Transactions it never saw F1 92.5% +0.0 ptsReal payee names F1 82.4% +0.0 pts
Every free register, and a donations line 29 Oct 2026 · 8ff7816
Foursquare’s places and seventeen state and Commonwealth registers — liquor and gaming licences, schools, hospitals, care services, fuel sites, chain stores — join the register of places. Charities, councils, health funds, universities and chains are read by name. Donations and charity is a line. A company suffix written onto a name is split off.
Transactions it never saw F1 92.5% +0.4 ptsReal payee names F1 82.4% +2.3 pts
Web addresses and words for food 29 Oct 2026 · a0bf0e5
A shop named by its own web address, from Wikidata. Words that name a place to eat, from the register of places. An app bank’s own covers and forwards, and the holder’s own name mid-line.
Transactions it never saw F1 92.0% +0.2 ptsReal payee names F1 80.1% +0.5 pts
Own money moving, and chains on a campus 28 Oct 2026 · cover-1028
An app bank’s “Cover to Spending” and “Cover from” lines are the holder’s own money covering a purchase, whatever purchase they name, and are read as transfers. A university’s name is read after the chains’ names, so a chain’s shop on a campus is the chain. Both came from real statements, where coverage rose from 82.1% to 83.7%; the test sets here hold few such lines, and one more payee is named right.
Transactions it never saw F1 91.8% +0.0 ptsReal payee names F1 79.6% +0.0 pts
A setting that answers every line 28 Oct 2026 · complete-1028
Complete reads as broad reads and then gives every line nothing read the model’s best guess, marked as one: a spending line for money out, a transfer for money in, never income, a commitment or gambling. It is for comparing with services that categorise every line and quote accuracy over all of them, which is recall here, now published for every setting. Not held to a precision floor, and never the default. Published with it: the merchant on its own, on the lines whose label names a business, and the line as a whole.
Transactions it never saw F1 91.8% +0.0 ptsReal payee names F1 79.6% +0.0 pts
A register of Australian places 28 Oct 2026 · places-1028
A card line no rule reads is now read from the business a register of 724,000 Australian places (Overture Maps) lists by those words in the suburb the line names, matched by the suburb’s postcodes from G-NAF, and named. Only where every place that matches says the same line of a budget, and the line’s own words do not point elsewhere. Balanced and broad; never strict, and never money in. Most of the rise is the synthetic account’s own shops, which are real and listed: on held-out lines from shops the development half never had, the default went from 67.4% to 70.9%, at 97.1% precision.
Transactions it never saw F1 91.8% +5.9 ptsReal payee names F1 79.6% +3.8 pts
Every set measured again 28 Oct 2026 · measured-1028
The held-out set was rebuilt from its public sources, line for line, and the BPAY directory fetched again; both were measured with the reader of 23 October. No reading changed. The billers’ figures moved because the directory did.
Transactions it never saw F1 86.0%Real payee names F1 75.8% +0.0 pts
Lines as the bank writes them 23 Oct 2026 · rules-1023
An accented letter, a payment terminal’s code and a date or time written into a line no longer hide the payee, so four brands a rule was written for are read, and a monthly payment written with its date is one arrangement. Pay, dividends and workers’ compensation read from the words of a credit. Measured on the payee names and the bank’s words only: the held-out set and the billers are kept outside the repository and keep the figures of the release before.
Real payee names F1 75.8% +0.3 pts
Rules tidied 4 Oct 2026 · no-lenders-1004
One rule and one test line taken out. No other reading changed.
Transactions it never saw F1 86.0% +0.0 ptsReal payee names F1 75.5% +0.0 pts
More read by a rule 4 Oct 2026 · rules-1004
Chiropractors, orthodontists and psychologists, which a rule meant to read and never did. Wineries, cellars, chicken shops, nail bars and shoe shops by the word, from a larger map. The model’s thresholds set again on what the rules now leave: strict reads more, and the default reads what it did, more of it by a rule.
Transactions it never saw F1 86.0% +0.0 ptsReal payee names F1 75.5% +0.3 pts
A model, and a setting 3 Oct 2026 · 6772eda
A model fitted here reads lines no rule reads. A lender chooses strict (the rules alone), balanced (the default, each category held to 95%) or broad (90%). The figures from this release are the default’s.
Transactions it never saw F1 86.0% +2.3 ptsReal payee names F1 75.2% +2.2 pts
Car makes, benefits and account conduct 3 Oct 2026 · 3dc6fd7
Every car and motorcycle make on Queensland’s register of new vehicles. Government payments named with the Act they are paid under. Dishonours, overdraws and late fees as the banks write them. A loan received and child support paid are lines of their own.
Transactions it never saw F1 83.7% +0.0 ptsReal payee names F1 73.0% +2.4 pts
Every category over 95% 30 Sept 2026 · bf28028
A transfer’s reference no longer reads as a power bill. The held-out labels follow the taxonomy for liquor, gyms, councils and bakeries, and every release is re-scored on them.
Transactions it never saw F1 83.7% +0.1 ptsReal payee names F1 70.5% +0.0 pts
Forecourts, cash machines and agencies 30 Sept 2026 · d11e7a4
Fuel retailers from two state price registers, cash-machine operators and agency groups, each named from a public register.
Transactions it never saw F1 83.6% +0.1 ptsReal payee names F1 70.5% +5.2 pts
Three categories and BPAY 27 Sept 2026 · a72888b
Tax and government, paid to a person, and strata. BPAY billers by code and short name. Car finance.
Transactions it never saw F1 83.4% +4.0 ptsReal payee names F1 65.3% +3.3 pts
Kinds of business 27 Sept 2026 · 9c4b485
Kinds of business mined from real venue names. A transfer under the holder’s own name is their own money.
Transactions it never saw F1 79.5% +6.1 ptsReal payee names F1 62.0% +3.6 pts
The bank’s own words 27 Sept 2026 · 6c6f733
Fees, cash, transfers, pay and benefits, read from how 14 banks write them. Licensed bookmakers and debt collectors.
Transactions it never saw F1 73.3% +3.7 ptsReal payee names F1 58.4% +3.7 pts
Australian brands 27 Sept 2026 · 8e4b79f
Brands from a national index of shop names, one name for each merchant.
Transactions it never saw F1 69.6% +0.8 ptsReal payee names F1 54.6% +11.6 pts
Enrich 20 Sept 2026 · 67c74e1
Each reading names the payee and says how sure it is. Energy retailers and lenders named from public registers.
Transactions it never saw F1 68.9% +0.9 ptsReal payee names F1 43.1% +11.4 pts
First release 19 Sept 2026 · 413245b
A bank export read into income, commitments and spending.
Transactions it never saw F1 68.0%Real payee names F1 31.6%
What the figures mean
- Precision
- Of the lines Statement put in a category, the share that were right.
- Coverage
- The share of lines Statement put in any category. The rest it declines.
- Recall
- The share of all lines put in the right category, a line Statement declines counted as wrong: accuracy over all lines, the figure services that categorise every line quote as their accuracy. Under the complete setting, which answers every line, it is the same as precision.
- F1
- Precision and recall together, as their harmonic mean.
- Payee named
- The share of lines where the business was identified by name.
- Merchant right
- Of the lines whose label names a business, the share where Statement named that business. Beside it, of the names it gave, the share that were right.
- Whole line right
- Of all lines, the share with the category right and, where the line has a merchant, the merchant named right too.
AUC is not reported: it measures how well a score ranks, and Statement answers or declines rather than ranking. The same figures, for a program: /v1/statement/accuracy.