Credit History Meets Financial Reality

A credit bureau brings together reported credit relationships and behaviour across institutions. The permitted data varies by country, but typically includes enquiries, defaults or adverse information, account information and repayment history. Some systems also include balances, credit limits and repayment amounts.

Its value is not simply the number of records.

The bureau matches information to people and businesses, applies common definitions and creates a shared reference point that many lenders can use. It is sometimes described as a "registry of truth". That is a reasonable characterisation. Sure there are errors and late reporting but they are very much in the minority given the scale and velocity of the data. There is also a common interest in data accuracy across all stakeholders to a credit reporting network.

It provides a structured record of obligations and outcomes across the credit system.

Transaction data has a different organising principle - activity in an account. A $300 payment to a lender tells us that $300 moved. On its own, it does not tell us whether that was the full repayment, whether it was made on time, what the outstanding debt is, or whether another account funded part of the obligation.

Conversely, transaction data can show salary, regular expenses, cash reserves, overdrafts and changes in spending long before those circumstances produce a missed credit payment. The differences extend beyond the information itself.

Traditional bureau data is generally obtained within a regulated credit-reporting framework after appropriate notice. Permissioned transaction data involves the customer actively agreeing to share particular accounts. A bank's own first-party transaction data is different again: it already sits within the customer's relationship with that institution.

That creates three quite distinct sources of information.

A bureau can have broad coverage of the credit system but little visibility of what is happening to cash flow today. A permissioned transaction feed can give an excellent view of current finances, but only for accounts the customer chooses to connect. A primary bank may have years of continuous transaction behaviour, while still knowing relatively little about credit held elsewhere.

Permissioned does not mean complete.

Thousands of accurately categorised transactions from a secondary account are still an incomplete picture if the main salary account or a material liability sits somewhere else.

The datasets are complementary not because one is old and one is new, but because they describe different economic facts and are collected in different ways.

The infrastructure for accessing transaction data is now large enough that this is no longer a theoretical debate.

In the UK, Open Banking passed 100 billion cumulative API calls, across the nine largest banks, in June 2026. Monthly traffic reached 2.94 billion calls in July, with around 80% generated by Account Information Services. Open Banking Limited lists uses including account aggregation, onboarding, identity, affordability and access to financial products.

That is compelling evidence of scale, but not of lending penetration. An API call is not a loan application, and the same customer can generate many calls. The more interesting point is that the plumbing required to move financial data securely into other applications is now operating at very large volumes.

India tells a different story. Its Account Aggregator system had passed 500 million fulfilled consents, more than 310 million customer-linked accounts and six billion data fetches by September 2026. More importantly for credit, Sahamati estimates that Account Aggregator-enabled lending reached INR 3.82 trillion across 36.8 million loans in FY26, representing 8.4% of retail and micro, small and medium enterprise (MSME) lending by value and 11.8% by volume. Around 18.2% of participating institutions' AA-enabled originations were to new-to-credit borrowers.

That is a materially different use case.

In a market with large numbers of consumers and small businesses whose traditional credit histories may be limited, transaction and cash-flow data can help establish creditworthiness in the first place. It can expand the lendable population rather than simply improve an already well-informed decision.

Brazil provides another variation. Its Open Finance framework allows customers to share account, card, credit, investment and foreign-exchange information, and the central bank explicitly points to cheaper credit, higher limits and easier access to products at institutions where the customer has no existing relationship - as intended benefits. Earlier this year, the Banco Central do Brasil (BCB) data already showed in excess of BZR 31bn of credit originated using Open Finance data sharing, with key operational stats more than doubling past 12 months - the BCB reported the ecosystem had passed 100 million customers or connected accounts and 154 million active consents.

Whilst a lot of the evidence focuses on account origination it's not yet clear if it has translated into measurable credit-performance improvements. This distinction informs us that infrastructure scale should not be confused with proven underwriting benefit.

Australia demonstrates something else again. Transaction-data lending predates the Consumer Data Right. Digital data collection (common name - screen-scraping) has been offered by various players since the early 2010s. Lender use took off after responsible lending obligations arrived in 2010-11. Take-up was strongest in the small-amount lender and non-bank lending sectors. The use was focused on retrieved bank statements and categorised income and expenditure information to reduce manual assessment.

The CDR therefore did not invent the lending use case. It provides a better regulated rail for obtaining information lenders were already learning to use. The presence of the CDR has expanded use cases and participants.

And then there is the United States, where traditional credit-bureau data is already deep and abundant. That makes Experian's October 2026 launch of a separate Cashflow Data Bureau particularly interesting.

The new consumer reporting agency operates under the Fair Credit Reporting Act and combines consumer-permissioned account connectivity, standardised transaction reporting, categorisation, cash-flow attributes, scoring and decisioning. Experian also plans a Credit and Cashflow Score combining traditional credit information with permissioned cash-flow data.

If cash-flow data were useful only where bureau information was missing, there would be less reason to build this in the world's deepest credit-information market.

The implication is fairly straightforward: the same data infrastructure solves different problems in different markets. In thin-file markets, it can make a credit decision possible. In mature markets, it can improve an existing one. That said, and despite recent politics, the US remains a market with plenty of immigration, with the need to supplement gaps in credit files.

The first use is mundane but valuable: verification. Income can be identified from account flows. Expenses can be categorised. Serviceability can be assessed without customers uploading piles of statements and payslips. The customer benefit is simple: digital documentation and a faster answer.

The second use is more interesting: decisioning. Cash-flow data can provide another look at applicants whose traditional credit record is incomplete or whose score sits just outside a lender's normal approval range. It can also be particularly useful in small-business lending, where the cash movements of the business may say more about repayment capacity than the owner's personal score.

Then some uses start to change the product itself: credit limits that respond to changing financial circumstances, working-capital facilities linked to turnover, automated renewals, pre-approved offers and earlier identification of hardship.

Nova Credit is a useful example of how the data-provider side has evolved alongside lenders.

Nova began with a very focused problem. New immigrants arriving in the US often had perfectly good credit histories overseas but effectively started again when they crossed the border. Nova built Credit Passport to translate international bureau information into a form American lenders could use. Then COVID arrived and immigration stopped.

According to a detailed account in Forbes, usage of Nova's Credit Passport fell about 75%. A conversation with SoFi about identifying which customers still had jobs led Nova toward bank-account data and income verification. It subsequently built Cash Atlas, which converts permissioned transaction histories into cash-flow attributes, reports and scores.

That is a genuine product pivot. Nova went from solving "this customer has credit history, but it is in another country" to helping solve "what can this customer's current financial behaviour tell us that the conventional credit file does not?"

Nova announced in October 2024 that SoFi would use Cash Atlas for consumer-permissioned bank data and cash flow analytics, while Atlanticus reports using cash-flow attributes alongside its existing models to approve 15% of a defined group of marginal declines without exceeding its expected risk. Whilst this is a vendor case study, it still clearly demonstrates the use case.

Nova's pivot is a useful example of the wider point: better access to transaction data does not just improve an existing credit process. It can give lenders new ways to assess customers, revisit declines and build products around cash flow rather than relying only on the traditional credit file.

Data in and of itself holds value if there is decision value.

FinRegLab's 2025 consumer research compared models using credit-bureau data, bank cash-flow information and combinations of both, and tested predictions against subsequent performance on new accounts.

The strongest model overall was a machine-learning model combining bureau and cash-flow data. It also generally produced the highest approval rates at given risk thresholds.

There is an important qualification. Adding cash-flow information improved prediction, but FinRegLab found that moving from traditional logistic regression to machine learning produced a larger improvement than simply adding the additional dataset.

That is a useful dose of reality. More data has a cost. It has to be obtained, categorised, interpreted, tested and maintained. Customers may fail to connect accounts. Connections break. Transactions are ambiguous. Lenders still have to determine whether the additional signal changes enough decisions to justify the work.

There is another important limitation. FinRegLab's modelling sample contained relatively few consumers with little or no traditional credit history, and excluded applicants without a credit score as well as those without sufficient usable bank-account history. FinRegLab itself says this prevents it from precisely estimating the impact on consumers who may have the most to gain from cash-flow underwriting. It believes the cash-flow results could have been stronger with more thin- or no-file borrowers, but says the available data cannot establish that definitively.

So the evidence supports cash-flow data. It does not support throwing away everything else.

It is worth being careful with claims that AI is suddenly changing credit scoring. It has been around for a while.

Researchers were experimenting with neural networks in credit scoring in the early 1990s. A 1992 paper tested a back-propagation neural network on credit applicants, and further studies through the 1990s compared neural networks with logistic regression and other conventional techniques. The results were interesting but hardly a clean victory for AI; traditional statistical models often remained very competitive.

Equifax commercialised its NeuroDecision neural-network credit-scoring technology in 2018, describing it as an explainable machine-learning approach designed for regulated credit decisions. Lenders have long had to explain declines, and that is a hurdle neural networks had to clear.

What may be changing now is not the existence of machine learning but what the models can learn from.

Revolut's PRAGMA project is a good example. Working with NVIDIA, Revolut is developing a transaction foundation model intended to replace multiple separate modelling systems across credit, fraud, engagement and recommendations with a shared behavioural intelligence layer. They reported early results on historical data included a 2.3x uplift in identifying credit default risk. (Note: these are internal comparisons against Revolut's existing systems, not external benchmarks or third-party verification). NVIDIA reports that the approach materially reduces the time spent engineering separate features and models for each problem.

The important idea is not simply using AI to produce a credit score. It is using AI to learn patterns in a customer's financial behaviour directly from long histories of transactions and other account activity, then reusing those patterns across different decisions.

That could increase the value of a long-running first-party banking relationship. Rather than designing a variable such as "average income volatility over six months", the model may discover useful patterns across sequences of salary, spending, balance and payment events that humans would struggle to specify in advance.

For lending, that could eventually support faster model development, more responsive limits, better identification of changing capacity, more tailored offers and new cash-flow products.

There are plenty of caveats. Reported model improvements need independent validation. A research benchmark is not the same thing as a production underwriting outcome. Credit decisions still need governance, reason codes and an ability to explain adverse decisions.

But the direction is worth watching. The better models become at understanding financial-event sequences, the more valuable rich transaction histories may become.

The commercial market increasingly looks less like bureau versus transaction data and more like different ways of putting the two together. In Australia alone, each of the consumer credit bureau players has solutions. Some are score-based, which necessitates categorisation for attributes; others go as far as becoming CDR-accredited, and there has been plenty of M&A along the way.

Different routes, but a similar conclusion. Useful data still has to be standardised, validated, approved and inserted into a lender's workflow before it becomes valuable. The market is not neatly organising itself around bureau data or transaction data. It is increasingly organising around who can turn both into trusted, decision-ready information.

For a lender, the sensible starting point is not "we need transaction data" - focusing on the problem applies to credit risk as well.

Too many applicants are being declined because their credit record provides insufficient evidence. An affordability assessment takes too long. A good customer is being priced badly. A small business has visible cash generation but very little conventional credit history. Deteriorating capacity is being identified only after repayments start failing.

Then ask whether another source of information improves that decision enough to be worthwhile.

Transaction data does not look like a wholesale replacement for the credit record across mainstream lending. It looks like another important source of financial evidence - sometimes supplementary, sometimes substitutive, and particularly valuable where the traditional record is weakest or least current.

The interesting opportunity is in the relationship between the two. The credit record provides an independent history of obligations and performance across the credit system. Transaction data provides a changing view of what is happening in the customer's financial life.

If lenders can combine those views well, the result should be quite practical: a better decision, made faster, with less paperwork, more appropriate pricing and greater access to credit for customers who deserve it.

That seems a more useful ambition than trying to declare one dataset the winner.