The behavioural intelligence layer for SME lending

Your bank transaction data contains far more business intelligence than your models can see.

Undercurrent transforms raw bank transactions into explainable behavioural intelligence for better underwriting, portfolio management and relationship decisions. Go beyond categorisation and generic feature engineering to understand how businesses behave, adapt and evolve.

See hidden creditworthinessFind good businesses your current representation overlooks.
Respond as behaviour changesTrack business-model-specific risk and opportunity between model rebuilds.
Build reusable intelligenceUse the same explainable behaviour layer across models and decisions.

The problem

Every lender has transaction data. Very few truly understand it.

Millions of transactions are reduced to balances, turnover, averages and a handful of engineered features. Yet experienced underwriters can often recognise important patterns in the same statements because they interpret behaviour—not just numbers. Today’s models rarely capture that knowledge consistently.

The conventional pipeline

  • Transactions become disconnected features
  • Fixed windows assume unlike behaviours work alike
  • Business meaning is buried inside model-specific engineering
  • Underwriter knowledge remains difficult to scale

The missing intelligence layer

  • Business behaviour becomes a first-class object
  • Representations adapt to context, history and horizon
  • Metrics remain named, transparent and evidence-backed
  • The same behaviour layer supports many decisions

Cafe · POS driven

Lumpy revenue may signal strain

Money should arrive frequently. Broken settlement rhythm can be meaningful.

Ecommerce · Gateway paid

Lumpiness may be settlement timing

The signal must recognise gateway cycles before calling normal behaviour risky.

Tradie · Invoice driven

Three large payments may be healthy

The real behaviour may be debtor timing, recovery and supplier-payment hierarchy.

The missing layer

We do not build another credit model.

We build the adaptive behavioural intelligence layer between bank transactions and every lending decision. It gives models, analysts and underwriters a shared representation of how a business actually operates.

01 · EvidenceRaw bank transactions
02 · RepresentationAdaptive business behaviour
03 · IntelligenceValidated behavioural metrics
04 · DecisionsUnderwriting, monitoring and growth

Adaptive representation

The representation is often more important than the transformation.

Recurring payroll should not be analysed like seasonal revenue. Inventory purchasing should not be represented like customer receipts. Liquidity may require a different window from revenue stability. Undercurrent determines how each behavioural concept should be represented using available history, transaction density, business type, seasonality, lending context, prediction horizon and the pattern itself.

Define the behaviour

Start from the credit or portfolio question and the business concept that could explain it.

Explore representations

Test windows, rhythms, events, interactions and latent views appropriate to the signal.

Validate the evidence

Measure incremental contribution, stability, coverage, robustness and point-in-time integrity.

Reuse what earns a place

Deploy transparent metrics with lineage and retain the evidence for future decisions.

Most candidates fail. That is the evidence gate working.

The Business Behaviour Profile

From thousands of features to a measurable story of the business.

Undercurrent represents each business through transparent behavioural concepts. Every concept is supported by traceable metrics derived directly from transaction history—not a black-box score.

Ability

Can this business generate enough cash to service debt?

Stability

How predictable and resilient are its operations?

Willingness

Does it consistently demonstrate responsible payment behaviour?

Liquidity

Can it absorb an unexpected shock?

Growth quality

Is growth sustainable, cash-generative or fragile?

Operational discipline

How effectively does the business manage its financial obligations?

For human decision-makers

Understand what happened—and why.

Underwriters and strategy teams receive explainable behavioural evidence instead of manually interpreting hundreds of transactions.

For machine learning teams

Use ML-ready metrics, not another opaque score.

Validated behavioural metrics arrive with lineage, temporal context, coverage and explainability as a reusable representation layer.

Commercial value

One signal layer. Better decisions across the book.

Signals are measured once and validated wherever they earn a role. This creates a consistent behavioural vocabulary across the customer lifecycle instead of rebuilding disconnected features for every model.

Up to 10×faster portfolio analysis

Give strategy analysts rapid, repeatable feedback for monitoring the book, testing approval and pricing thresholds, and responding to emerging conditions. This matters between model rebuilds: the signals that reveal a macroeconomic shift can change quickly—and can look very different across business models.

Acquisition

Target businesses more likely to become profitable, performing customers.

Underwriting

Approve more creditworthy SMEs without proportionally increasing loss.

Pricing and limits

Match economics more closely to observed risk and resilience.

Portfolio strategy

Test score cut-offs and policy thresholds quickly, without waiting for a full model rebuild.

Early warning

Recognise business-model-specific changes in resilience before arrears become the first signal.

Collections

Prioritise accounts and choose treatment using current behavioural evidence.

Renewal

Recognise improving borrowers, protect good customers and manage declining ones.

Grow the performing book while improving risk-adjusted profitability.

Commercial value appears through additional performing-loan revenue, more accurate pricing, stronger renewal value, avoided credit losses and faster, lower-cost portfolio decisions.

incremental value =
performing-loan revenue
+ pricing uplift
+ renewal value
+ avoided credit losses
− implementation cost

Design-partner pilot

Prove the value on your own historical book.

Start with one decision and one agreed outcome. Undercurrent works with your existing data and models, validates incremental contribution against pre-agreed thresholds and leaves you with the evidence either way.

Built by Parisa Milne, PhD, drawing on more than fifteen years of applied data science across healthcare, retail, technology and SME lending—and a career spent finding trustworthy signal in noisy real-world systems.

01Choose a high-value decision and define the commercial success measure.
02Discover and validate behavioural signals against historical outcomes.
03Compare incremental lift, stability, coverage and explainability.
04Promote only evidence-backed signals into a controlled pilot.