What Creditcrest Enrich gets right
Measured on 5 November 2026, against names nobody here chose.
Two things are measured and they are not the same thing. Coverage is how many payees it categorises at all. Accuracy is how many of those it gets right. A single number combining them would hide whichever half is worse, so they are never multiplied together on this page.
What it was run over
- OpenStreetMap branded Australian outlets (ODbL), judged once per brand, over 28,777 sites.
- The Consumer Data Right Register: all 84 energy retailers a household can hold a bill from, all 117 banking brands, and all 131 non-bank lenders.
- Every insurer registered to sell private health cover in Australia, from the Private Health Insurance Ombudsman (54 names, including the brands one fund underwrites for another company).
- Every Commonwealth payment the Department of Social Services reports recipients for, taken from the column headings of its own release (24 names, read as money in).
- Forty thousand real venue trading names from three states’ liquor and fuel registers (CC-BY), which carry no right answer and so measure what this SAYS rather than whether it is right.
The curated list in this repository — the one the merchant rules were written from — scores full marks and appears in no figure here. A list that was used to write the rules measures the rules against themselves.
Accuracy, per confidence band
| Band | Names | Claimed a category | Right | Accuracy | What it means |
|---|---|---|---|---|---|
| named | 499 | 499 | 499 | 100.0% | A rule matched the merchant’s name in the description. |
| kind | 185 | 185 | 185 | 100.0% | A rule matched a word for a kind of business inside a name this has never seen. It puts the money on the right line and it does not identify the payee. |
| paired | 0 | — | — | — | The holder’s own money moving between their accounts: an app’s split of a payment into a saver, or that share arriving in the saver the same day for the same amount, whatever name it carries. Not yet measured on labelled statements. |
| coded | 0 | — | — | — | No rule matched the words. The bank sent a code that settles the line: the card scheme’s code for the kind of business, or its own word for a fee or interest. Not yet measured on labelled statements. |
| listed | 106 | 106 | 106 | 100.0% | No rule matched the words. A public register names the business and says what kind it is: Wikidata by the web address on the line, a register of organisations (charities, councils, health funds, universities, chains) by their name, or a register of Australian places (Overture Maps, Foursquare and state registers) by these words in the suburb the line names. Measured on public registers and sample statements, not yet on lenders’ statements. |
| modelled | 12 | 12 | 12 | 100.0% | No rule matched. A model fitted in this repository read the words; the confidence is the share of its answers at this score, in this category, that were right on lines held apart to measure it. |
| inferred | 0 | — | — | — | No rule matched. The size and direction of the money were used instead, which is a guess and is marked as one. |
| guessed | 0 | — | — | — | Nothing read the line, and the complete setting gives every line a category: this is the model’s best guess, a spending line for money out and a transfer for money in. Most such guesses are wrong; it is marked as one. |
| declined | 10 | — | — | — | The name is recognised and is deliberately not categorised, because the name does not settle what the payment was. |
| unknown | 482 | — | — | — | Nothing in the description matched a rule. |
A figure this high means "two wrong answers in 684", not "almost never wrong". It is measured over the brand names in the lists above, and the long tail below is 39,548 real venue names of which four in five are not read at all. That is why the denominator is in the table and why coverage is printed beside accuracy everywhere on this page: either figure on its own is a number somebody will quote wrongly.
It got there by fixing wrong answers rather than by adding merchant names. Every collision the audit found was one brand trading in two industries — Costco sells fuel and groceries, Harcourts is a real-estate agency and a pub, Arcline is RACV’s electricity brand and RACV is an insurer. No rule in a list fixes a fact about the world. Where a word in the name settles the question the rules use it; where nothing settles it the name is recognised and deliberately not categorised, which is what the declined band is.
declined and unknown carry nought, which is not a low opinion of them: they are the two answers that make no claim about a category at all, and nought is the correct confidence in a claim nobody made. inferred is the one band no payee list can measure — it is a guess from the size and direction of a credit, and measuring it would take a real statement with the income labelled on it, which this repository does not have and will not invent.
Accuracy and coverage, per category
Every category has to be right at least 95% of the time it answers, one at a time rather than on average — an average passes while one line of a household's assessment is wrong two times in five, and the line that is wrong is somebody's petrol or somebody's rent. Coverage is printed beside it, always. The two trade against each other: a category can reach full marks by answering once, so a figure with no coverage next to it says nothing.
| Category | Names | Categorised | Right of its answers | Third party named | Clears 95% |
|---|---|---|---|---|---|
| Transport and fuel | 225 | 64.9% | 100.0% | 50.7% | yes |
| Shopping | 193 | 75.6% | 100.0% | 73.1% | yes |
| Eating out and takeaway | 190 | 74.7% | 100.0% | 50.0% | yes |
| Loan repayment | 113 | 54.0% | 100.0% | 43.4% | yes |
| Utilities | 84 | 79.8% | 100.0% | 77.4% | yes |
| Insurance | 66 | 74.2% | 100.0% | 43.9% | yes |
| Entertainment and subscriptions | 49 | 75.5% | 100.0% | 55.1% | yes |
| Cash withdrawal | 43 | 53.5% | 100.0% | 30.2% | yes |
| Health | 41 | 56.1% | 100.0% | 26.8% | yes |
| Groceries | 31 | 83.9% | 100.0% | 67.7% | yes |
| Rent | 26 | 42.3% | 11 of 11 | 30.8% | too few |
| Government benefit | 24 | 83.3% | 100.0% | 83.3% | yes |
| Mortgage repayment | 21 | 95.2% | 100.0% | 4.8% | yes |
| Education and childcare | 18 | 72.2% | 13 of 13 | 22.2% | too few |
| Travel | 8 | 87.5% | 7 of 7 | 62.5% | too few |
| Phone and internet | 6 | 83.3% | 4 of 4 | 83.3% | too few |
| Credit card payment | 4 | 100.0% | 5 of 5 | 100.0% | too few |
| Gambling | 1 | 100.0% | 1 of 1 | 100.0% | too few |
| Buy now pay later | 0 | — | 2 of 2 | — | too few |
Right of its answers is counted by the category it gave: of the names put on that line, the share that belonged there. Under 20 answers one wrong one moves it by more than five points, so it is shown as a count and not judged.
credit card payment, 4 names; gambling, 1 name. A share of four names is not a measurement, and the row is here because the register says so rather than because the figure means anything. Asking OpenStreetMap for bookmakers, for instance, gets one Australian brand carrying a brand tag, so whether the rules read it says nothing about gambling as a whole.
Coverage is weakest at rent. Of 26 names the corpus puts on that line, 42.3% are categorised at all. That is a fact about how many independent businesses there are rather than about the rules: nobody writes a rule for every forecourt in the country.
No category answered 20 times or more has a wrong answer on these names. That says it has not yet been wrong here, not that it cannot be. The wrong answers it has given on these lists were each a brand that trades in two industries: Costco sells fuel and groceries, Harcourts is a real-estate agency and a pub, La Trobe is a university and a mortgage lender. Where a word in the name settles which — INSURANCE, ELECTRICITY — the rules use it; where nothing settles it, the name is recognised and deliberately not categorised, which costs coverage and is why that column is printed beside this one.
Under all of it is one number: four out of five real Australian venue names are not recognised at all. Everything else on this page rests on that being said out loud.
The names it must not answer for
151 bank names and bank brands, where the right answer is no answer — a payment to a bank does not say whether it was a mortgage, a card, a fee or the person moving their own money. None of them was categorised. 2 of them are in the rules with no category against them, on purpose, so that no broader rule can guess and so the file can say the name was looked at.
The long tail
39,548 real Australian venue trading names from three states' liquor and fuel registers. There is no right answer to compare against, so this is what it says rather than whether it is right:
| Band | Names | Share |
|---|---|---|
| named | 2,837 | 7.2% |
| kind | 11,300 | 28.6% |
| listed | 1,913 | 4.8% |
| modelled | 3,339 | 8.4% |
| unknown | 20,141 | 50.9% |
7.2% of real Australian venue names have their third party named by a rule. A further 28.6% are categorised from a word for a kind of business sitting inside a name this has never seen, which puts the money on the right line and is not merchant recognition and is not sold as it. 4.8% more are read and named from a register of Australian places that lists the business in the suburb the line names. 8.4% more are read by the model under the default setting; it was kept apart from these names while it was fitted. The rest say unknown.
What is not measured at all
11 of the 29 categories have no independent corpus. Silence here is not a pass.
| Category | Why there is no list to measure it against |
|---|---|
| Strata and body corporate | 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. |
| Tax, rates and government charges | Tax offices, councils, revenue offices and fines offices are few, and are named rather than looked up. They are checked on the bank-wording lines instead. |
| External transfer | 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. |
| Salary or wages | An employer is any of two and a half million registered businesses. The only complete list is the business register, and a household is paid by one row of it. |
| Other income | A dividend, a refund, a private sale. Nobody publishes a register of the people who send a household money. |
| Buy now pay later | The providers are in the Consumer Data Right’s non-bank lending register and their names do not say the product — "Zip", "humm" — so the register settles that they are credit providers and not which kind. They are measured as a commitment rather than split out, because splitting them would be this repository deciding the answer the corpus declined to give. |
| Transfer | Not a payee at all. A transfer is recognised from the words a bank writes for its own rails. Those words are held to the 95% floor on 719 lines from 14 banks, but the rules were written from those lines, so that is a floor that must hold rather than a measurement. |
| Bank fees and interest | Not a payee either. A bank names its own fees, so the wording belongs to the bank rather than to a company a register could list. The same 719 lines of bank wording hold fees to the floor, for the same reason not as a measurement. |
| Child support paid | 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. |
| Loan or advance received | Money a lender pays in. The lenders are on the registers the commitments are measured on, but those name who a household repays, read as money out; nobody publishes which credits were a loan drawn down. Held to the floor on the bank-wording lines that say so — a drawdown and a pay advance — as a check, not a measurement. |
| Donations and charity | 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. |
A reason against each rather than a list, because a list invites the reader to assume nobody looked. Somebody did, for every one of them, and the reason is a different reason each time: a register that does not exist, a register that exists and blocks fetchers, and a payee that is not an entity at all. Saying so is the only honest thing available.
How to check any of this
Every figure on this page is recorded with the release it was measured on, and the test suite compares them on every change: a change to the merchant rules that moves a number fails until the figure here is updated. The same figures, as data: /v1/enrich/accuracy.