Beyond what happened: predicting what comes next

How does Neumetria differentiate itself from what existing categories like Open Banking/data aggregation already claim to do?

Aggregation tells you what happened. Neumetria tells you what financial state a person is in now, and what is likely to happen next.

An aggregator connects the account, normalises the feed, and labels the merchant. That is real infrastructure, and we depend on it. But the output is a cleaner version of the same rows, and a category label is not an interpretation. Knowing someone spent 52 EUR at a supermarket tells you nothing on its own.

What sits above that is a continuously updated read of the person's financial state, with an explicit measure of how much evidence supports it. Two things follow from building it that way. The system abstains rather than guessing when evidence is thin, and every read is reconstructable later, on what data and under which version.

The harder standard we are building toward is calibration. Not a ranking, but a probability that means what it says. That property is earned from realised outcomes over time rather than asserted at launch, which is precisely why, once earned, it becomes difficult to replicate, and why I am wary of anyone claiming it on day one.

We are also a complement to the bureau, not a replacement. The bureau holds the long memory. We read the live state.

Walk us through what your signal actually looks like in practice, what goes in, what comes out, and how a bank or PSP acts on it?

Financial activity in, understanding out.

We ingest normalised financial activity from connected accounts rather than requiring a traditional customer file. Transactions and account context become the evidence from which the person’s live financial state is constructed.

The interpretation happens in between. A 52.40 EUR debit at a supermarket on a Saturday morning is read as an essential grocery purchase, one day after payday, consistent with this person's established rhythm. The same amount at the same merchant means something different on day 28 of a thinning month.

Out: the person's current financial state. Income rhythm and its stability. Liquidity horizon in days. Committed outflows and what remains after them. Observed debt service. Each carries a confidence, an as-of date, and the observation count behind it. Six observations and six hundred never look equally supported.

Then the institution decides what to do with it, and that is the part I want to be precise about. We do not author the credit policy. The lender writes its own conditions and thresholds, and we execute that published policy against the state we return, under the oversight arrangement they choose. Their rules, their outcome language, their accountability. For product surfaces, it is lighter again: readiness signals that tell a product whether this is a moment to prompt or a moment to stay quiet.

What is the smallest amount of transaction history you need before your signals become reliable for a given user, and how do you handle users whose income or spending is irregular, like gig workers?

There is no single threshold, and I would be suspicious of any vendor who gives you one.

Reliability is a property of the individual read, not the product. Confidence reflects how much evidence exists for this person on this question, so thin history produces a wider, less certain read rather than a confident wrong one. When it is too wide, we return insufficient data and name what is missing. Some reads firm up quickly because the pattern is dense. Others never do.

As a reference point, around 90 days is especially useful, because it gives you roughly three-monthly income cycles, six biweekly or thirteen weekly. A pattern has to repeat several times before you can tell a rhythm from a coincidence.

But that is an observation horizon, not a minimum. Different reads become useful at different points, and the confidence attached to each one tells you how much weight it deserves.

Irregular income is a separate question, and it is where most affordability tooling fails. Most systems treat irregularity as instability. An irregular shape can be entirely stable in its irregularity. A gig worker with variable weekly inflows and consistent behaviour around them is not the same risk as a salaried borrower whose income just became erratic. Where peer comparison is appropriate, we compare behaviour against relevant reference groups rather than assuming a single population baseline.

You're increasingly feeding signals to AI agents that can act autonomously. Where do you draw the line on what an agent should be allowed to do without a human in the loop?

We do not draw that line. The institution does, and we make the line enforceable.

If we decided which agent actions were acceptable, we would be authoring policy for a regulated firm, which is exactly the thing we refuse to do. What we provide is the mechanism: readiness signals a product reads before an agent acts, and an oversight setting the institution chooses.

Where I will state a view is on what the mechanism has to guarantee regardless of the setting.

An agent should never author or alter a rule. Policies are authored by the institution, validated, and versioned before use.

Consequence should be gated on reversibility. An agent surfacing context, or staying quiet, is cheap to undo. An agent producing an outcome that affects someone's access to credit is not.

And the route to a human has to survive the automation. Intervention, explanation, contest, and reassessment stay available regardless of the oversight setting. A right that can be disabled by configuration is not a right.

The failure mode I worry about is not an agent making one bad call. It is an agent making 10,000 consistent ones that nobody can reconstruct afterwards.

Regulators are more and more focused on alternative data underwriting. How do you make sure behavioural signals don't become a backdoor for disparate impact or de facto redlining?

By treating fairness through unawareness as insufficient, which it is. Removing protected fields does not remove proxies. It removes your ability to see them.

So, the discipline has to be testing rather than cleansing, and timing-based signals deserve the most scrutiny because they are the most seductive and the most dangerous.

The clearest example is late-night transaction activity. It looks intuitively useful, and it is trivially easy to compute, which is exactly why it deserves disproportionate scrutiny. It can reflect impulsive spending. It can equally reflect shift work or caregiving. The mistake is treating the timestamp as evidence of discipline, because then you have not measured discipline; you have penalised night nurses and single parents.

That does not automatically make the signal unusable. It means you cannot treat the correlation as the construct. A signal like that has to earn its place through validation and fairness testing, and it has to disappear if it fails.

Two things people miss.

Peer comparison is itself a fairness surface. If the way you group people correlates with a protected characteristic, the normalisation carries the bias rather than removing it. The grouping has to be audited on the same terms as the signals.

And selection bias is a fairness problem, not only a statistical one. You observe repayment only on loans that were written, so a model trained naively learns to reproduce whoever the existing policy already favoured. Accounting for that has to be part of validation, not an optional refinement.

If a bank or lender adopted this fully, what changes in 18 months: their product roadmap, their risk model, their org chart?

In that order, roughly, and with the org chart changing least.

The risk model moves first, and the interesting change is not accuracy on the population they already serve. It is which population becomes assessable. Cash flow evidence beside bureau history means thin-file and new-to-credit applicants stop being an automatic decline and start being an assessment with a confidence attached. Some of them are good risks the current model cannot see.

Pricing becomes more precise with it, because well-calibrated probabilities map directly to expected loss in a way that rank ordering alone does not.

The roadmap follows the risk model, not the other way around. Once a segment becomes assessable, products get built for it. Shorter terms, different limits, collection timed to when income actually arrives rather than to the first of the month. Also, the reverse: limits sized to what someone absorbs in a bad month rather than an average one.

The org chart change is smaller and specific. Someone has to own the feedback loop across applications, decisions, and realised outcomes. That sounds operational, but it matters enormously. If the learning loop only contains the borrowers the existing policy approved, the system risks learning that policy back rather than discovering where it was wrong.

About Amr Mohamed

Amr Mohamed is the Founder and CEO of Neumetria. His career spans more than two decades across banking, prime brokerage, wealth management, and fintech. He began at Banc of America Securities before co-founding Merlin Securities, where he built institutional prime brokerage infrastructure and scaled an organisation of more than 80 people. Merlin was acquired by Wells Fargo.

 

 

About Neumetria

Neumetria is the behavioural intelligence layer for financial services. It transforms financial activity into a live, context-aware understanding of each customer’s financial state, helping banks, fintechs, and AI agents understand behaviour, capacity, resilience, and intent to make better, more personalised decisions.