A probability of default built on your own book.
The cashflow score is a probability of default for each applicant, from a model built on your book and your outcomes in the pilot, calibrated so a stated probability means what it says, validated before it is trusted, and run as a service afterwards, priced per score. Day one, before the history exists, is the behavioural assessment inside the memo.
- Built on your book
- Calibrated and validated
- Reasons in plain English
- Hosted service, or in your environment
Reasons, in plain English
- Receipts steady against the same months a year earlierSupporting
- Buffer covers about five weeks of outgoingsSupporting
- A new lender's repayments began in the last two monthsDetracting
Calibrated on your book: in this band, 3 to 4 in 100 defaulted over 12 months.
Data quality: 12 months, 2 accounts, balances reconcile, 1.4% of lines needing review. Confidence: high.
Why lenders choose it
Not a generic score
Decisions you can defend.
What you receive for every file
A probability of default and band, reasons in plain English each marked supporting or detracting, the affordability measures, data quality and confidence, an applicant explanation, and a versioned record.
Cashflow signals, read the way the business trades
The signals under the score are engineered for business lending: revenue, lender repayments, owner drawings, tax and super, supplier terms and customer concentration are read as a business's money, not as spending categories. They are engineered for each business type and read the operating rhythm: seasonality, payment cycles, weekday and month-end patterns. Each business is compared with its own history and the same months a year earlier before it is compared with anyone else, so the model learns from your book without judging every business by the same rule.
Day one, and the score
- Day one: the behavioural assessmentInside the memo, every applicant's statements rated on five behavioural measures with no model to train. See the assessment.
- Then the score, built on your bookA probability of default from your applications and outcomes, in the pilot, run as a service afterwards.
- One systemThe score decides the straightforward cases under your policy; the memo handles thin files, exceptions and referrals.
A separate score for renewals
The model learned across the book and the single customer's own history see different things, with gaps and overlaps. The renewal score is built separately on your renewals and their outcomes, and reads both. See renewals.
The explanation your customer can receive
The reasons in the notice come from the same signals as the score, in words the customer can act on.
Dear Ms Okafor,
Thank you for applying for equipment finance for Okafor Landscaping. We are unable to approve the application this time. We looked at twelve months of statements from the two business accounts you sent us. The main reasons were:
- Repayments to your existing lenders took a large share of the money coming in, and a new lender's repayments began in the last two months.
- The balance in your accounts was below two weeks of outgoings for most of the last quarter.
- Money coming in was lower than in the same months a year earlier.
You can ask for a person to review this decision, ask what information we used, or send us statements from any other account the business uses and we will look again.
Illustrative, synthetic. A fictional applicant.
Evidence, with its limit
Problem: a basic scorecard misses risk in how customers transact. What we did: engineered cashflow signals from each customer's transaction history and gave them to the same model on the same held-out set. Result: ranking power (AUC) rose from 0.7505 to 0.7768. Limit: anonymised consumer card research data; the business-lending score is measured on each lender's own book in the pilot.
How it reaches you
- An API for one fileSend the statements for one business, receive the scored result.
- Batch for a bookScore a whole portfolio in one run, for monitoring or for a pilot.
- File batch for evaluationsDrop in a folder of historical files and receive the scored results beside the outcomes.
- DeliveryDelivered as a hosted service for lenders and brokers, or in your environment for institutions; your files are processed for your decision and never used to train any model.
- Into your systemsThe score and reasons land in your loan origination system, CRM or decision engine.
- The same result in the memoThe underwriting memo uses the same scored result, so the memo and the score never disagree.
Governance shipped with the score
Not a separate purchase: what your model risk function needs to sign the score off and keep it signed off.
- A validation pack for your bookTesting on a later period than the one it learned from, calibration, ranking power, gaps between segments, fairness tests across groups with effect sizes and the mitigations tried, checks for deliberate faults and for leakage.
- MonitoringDrift in inputs, signals and outputs, on a cadence you agree, from daily feeds to a monthly review and a quarterly recalibration.
- Change controlEvery change versioned, with a rollback you can invoke.
- Model card and decision recordKept for your model risk function, and attached to each scored file.
- Australian regulatory framingWritten to be used within your responsible lending obligations, the Privacy Act including its transparency requirements for automated decisions, APRA CPS 230 and CPS 234 where you are a prudential lender, and the consumer data right rules where you use data feeds.
- The score informs; your policy and your people decideIt never approves or declines by itself. Your policy sets any straight-through path.
The pilot
The score is built on your book and run in shadow mode beside your incumbent scorecard, on deals you have already decided. You see:
- Agreed before it startsSuccess criteria and a named sign-off owner on both sides.
- The winner is namedEven when it is not us.
- Then per scoreTiered by volume and fixed for the contract year, as a service.
Cashflow score, bureau-only score and manual assessment compared
| Cashflow score | Bureau-only score | Manual assessment | |
|---|---|---|---|
| What it looks at | The business's own bank statements, every account | Credit history held by bureaus | Whatever the assessor reads |
| Businesses with little credit history | Scored from how money actually moves | Often too little to score | Possible, but slow |
| Reasons given | Plain-English reasons and an applicant explanation | A few standard reason codes | Notes, in the assessor's words |
| Calibrated to your book | Yes, built on it in the pilot | Generic to the bureau's population | Not applicable |
| Needs history | Your book and outcomes; until then the behavioural assessment inside the memo | No | No |
| Consistency across files | The same file always gets the same result | Consistent | Varies by person and by day |
| Record | Version, evidence and replay for every file | The score and its date | The file note |
| Best used | Beside your scorecard and your bureau data | As an input to the decision | For the cases that need judgement |
Questions about the cashflow score
What data does it need?
Bank statements for every account the business uses, as PDF, scanned PDF, CSV or XLSX. Application fields and a bureau file are optional and are used for the affordability measures.
Statements or data feeds?
Statements are enough to score a file. If you already receive transaction feeds through a data partner, the same score runs on them.
Do we need a book to start?
The score is built on your own applications and outcomes, so it needs your history. Until it exists, every applicant gets the behavioural assessment inside the memo, rated on the same signals with no model to train.
How is it calibrated?
On your own book, in the pilot, so a stated probability maps to the default rate you actually see. Recalibration is part of the monitoring cycle.
How are the explanations produced?
Each reason is read directly from the signals that moved the probability, then written in plain English. The applicant explanation and the underwriter's view come from the same signals, so they cannot disagree.
How is fairness tested?
The validation pack compares results across groups on your book, reports the size of any gap and records the mitigations tried and what they changed. The results go to your model risk function.
What happens when the data quality is low?
The score still comes back, with a lower confidence and the reasons: a short history, a missing account, balances that do not reconcile. Your policy decides whether such a file goes to a person.
How is it priced?
Per score, tiered by annual volume and fixed for the contract year, after a fixed-fee pilot. An optional platform fee covers the validation pack, monitoring and recalibration.
Where does it run?
Delivered as a hosted service for lenders and brokers, or in your environment for institutions; your files are processed for your decision and never used to train any model.
What you can buy
Start a pilot: the cashflow score
Who it is for
Business lenders with a book and recorded outcomes: heads of credit risk, decision science and model risk at lenders, finance companies and institutions.
What you get
- A probability of default for each applicant, from a model built on your own book in the pilot and calibrated so a stated probability means what it says
- Plain-English reasons with every score, and the explanation an applicant can receive
- The validation pack your model risk function will ask for: a later-period test, calibration, segment gaps, fairness tests, drift monitoring and change control
- A scoring API for one file and a batch for a book; the score and its reasons into your loan system
- Recalibration to your book on an agreed cadence, run as a service and priced per score
- The pilot report: ranking, approval rate at equal loss, loss at equal approval, calibration, segments, and a priced scenario under your own assumptions
Cashflow signals, read the way the business trades
In scoring, the cashflow signals feed a probability of default built on your book, so the model learns from your borrowers without judging every business by the same rule. The cashflow signals are engineered for business lending and for each business type, and read the business's operating rhythm: seasonality, payment cycles, weekday and month-end patterns. Each business is compared with its own history and the same months a year earlier before it is compared with anyone else. A separate renewal score sits beside it, because the book view and the single-customer view have gaps and overlaps.
How it arrives
A sandbox in your environment on a few of your own files while your privacy and security review runs, then the hosted service in Sydney: files in by upload or API, results back, nothing to install. Institutions whose policy requires it run it in their own environment.
The first step
A fixed-fee pilot on your own deals: the score is built on your book and run beside your incumbent scorecard, with success criteria agreed first. The winner is named even when it is not us. Until you hold the history, the behavioural assessment inside the memo is day one.
Score the cash flow. Explain the decision.
Start with a fixed-fee pilot on deals you have already decided.