Feature store lookup and feature calculation

Plain language meaning

Feature store lookup and feature calculation convert approved banking data into reusable model inputs, while preserving definitions, versioning, lineage, freshness and point-in-time correctness.

This topic is about controlled banking features for credit, fraud and compliance decision support. It is not about storing random data fields because they are technically available.

In a real bank, this is not a loose technology idea. It is a controlled operating step where customer facts, banking policy, model behaviour, human authority, legal obligations and retained evidence must line up. AI and ML can improve speed, consistency and detection quality, but the bank must still prove why the process was fair, explainable, secure, monitored and fit for purpose.

Where it sits in the banking AI journey

This card belongs to The Complete Pipeline. The working flow is Approved data, Feature definition, Point-in-time lookup, Feature calculation, and Model input set.

Read the flow as a banking control journey. Each stage needs a source system, a decision purpose, a failure mode, a control owner, a fallback path, a customer-impact view and retained evidence. Without those elements, the bank may have automation, but it does not yet have a bank-grade AI process.

Banking data and evidence

The important data points are income stability, balance trend, repayment history, utilisation ratio, transaction velocity, KYC risk flag, delinquency marker, and feature timestamp. These items matter because they influence lending eligibility, affordability, fraud risk, compliance treatment, operational queueing, regulatory reporting, customer explanation and audit traceability.

The evidence pack should include feature definition, calculation log, feature snapshot, version record, test result, model input file, and approval note. A strong bank can replay the case from source data to feature values, model output, control result, human review, final outcome and monitoring result. A weak bank only knows that a system produced an answer.

Controls that make AI adoption safe

The core controls are feature owner, definition version, calculation test, freshness rule, point-in-time lock, reuse approval, and lineage retention. These controls make the topic bank-grade because they tie technical output to approved policy, legal obligations, model governance, operational resilience and management accountability.

AI can help compare records, detect anomalies, retrieve policy, summarise case evidence, prioritise work, highlight weak signals and improve investigator consistency. It should not invent missing facts, ignore failed checks, bypass authority, hide uncertainty, decide material customer outcomes without approval or create explanations that cannot be tied back to approved sources.

Regulatory and governance lens

For banking use cases, model risk, fair lending, adverse-action explanation, credit-risk governance, data lineage, operational resilience, AML/CFT risk-based controls, sanctions compliance, fraud information sharing and auditability can meet in the same workflow. The practical design must therefore be narrower and more disciplined than a generic AI design.

The practical test is simple: if a reviewer asks why the bank used the data, why the model output was trusted, why the customer received that action, why an alert was cleared, why an exception was approved, or why a regulatory record was prepared, the evidence must already exist.

Diagram walkthrough

Read the diagram from left to right as Approved data, Feature definition, Point-in-time lookup, Feature calculation, and Model input set. The diagram is a control map, not decoration. It shows the minimum route by which data, AI or ML output, human action and audit evidence should connect.

Use it as a 30-minute study method. For each box, ask what system produces the data, what can go wrong, what control detects the weakness, who reviews the case, what customer or regulatory impact could arise and what record proves closure.

Most important mistake to avoid

The common failure is treating a feature as a simple column while different teams define it differently or calculate it using future data that was not available at the decision time.

The correction is to slow down the thinking, not necessarily the process. A well-designed banking AI process can be fast, but every fast step must still leave behind source lineage, control evidence, decision reason, human accountability, monitoring data and issue ownership.

Source anchors for accurate study

Federal Reserve SR 26-2, dated 17 April 2026, supersedes SR 11-7 and SR 21-8 for traditional model risk management and clarifies that generative and agentic AI need governance through broader risk-management controls.

NIST AI RMF 1.0 uses Govern, Map, Measure and Manage functions for AI risk management, and NIST AI 600-1 adds generative-AI risk actions for content provenance, hallucination, data protection, cybersecurity and human oversight.

U.S. Regulation B, 12 CFR 1002.9, requires specific principal reasons for adverse action in covered credit decisions, including when a creditor uses an AI model. CFPB Circular 2022-03 was withdrawn on 12 May 2025; do not cite it as current guidance. Primary sources: https://www.consumerfinance.gov/rules-policy/regulations/1002/9 and https://www.consumerfinance.gov/compliance/guidance/withdrawn-guidance/.

The EU AI Act treats AI systems used to evaluate creditworthiness or establish credit scores for natural persons as high-risk, except certain fraud detection and prudential capital contexts.

The Basel Framework IRB standards require banks to estimate and validate PD, LGD and EAD using relevant data, meaningful risk differentiation and ongoing governance.

FFIEC BSA/AML examination guidance expects suspicious activity monitoring systems to be risk-based, explainable by management, periodically reviewed and independently validated where appropriate.

Federal Reserve SR 26-3 and FinCEN's 12 June 2026 Section 314(b) materials clarify fraud-related information sharing under the USA PATRIOT Act safe-harbor framework for participating financial institutions.

OFAC's Framework for Compliance Commitments describes sanctions compliance programme components including management commitment, risk assessment, internal controls, testing and auditing, and training.

A feature value with an evidence trail

For a fictional application at 10:00, define income stability as the number of qualifying salary credits in the prior six complete calendar months. The feature contract names the eligible transaction codes, treatment of reversals, currency conversion policy, account ownership rule and minimum observation period. A same-day credit at 10:05 cannot appear in the 10:00 score, even if a later batch assigns it an earlier business date. The online service records feature version, input event IDs, calculation time, availability cutoff and missing-data reason. An offline training job must reproduce that value from the same rules.

The lookup result should distinguish zero qualifying credits from no account history and from a feed outage. These cases may require different model treatment and customer follow-up. If two customers are merged in the master system, historical features must use the relationship state known at the original decision, while later corrections remain available for investigation. The feature owner approves a definition change, compares old and new values on a fixed sample, and coordinates any model validation or threshold review it triggers. A release test should replay a real decision timestamp and compare each online value with the independently reconstructed value, not merely compare the names of columns.

Banking practice note: customer purpose

For feature store lookup and feature calculation, customer purpose is not a side detail. It decides whether the bank can connect the AI or ML output to a real banking purpose, a real customer or regulatory outcome and a real accountable owner. Study the topic as a banking process first and a model process second.

Trace one item from income stability to feature definition. Then ask which control from feature owner proves the item was complete, current, authorised, relevant and fit for use. If that trace cannot be shown without manual guessing, the process is not yet bank-grade.

AI can assist by comparing records, detecting unusual patterns, retrieving approved policy, summarising weak evidence, prioritising exceptions and preparing review notes. The bank should not allow a generated explanation, a confident score or a convenient dashboard to replace validation, consent, human judgement, customer communication, regulatory judgment or issue closure.

A strong implementation records the source event, data timestamp, consent or lawful basis, model or prompt version, feature values, score or generated output, threshold, reason code, user action, exception status, monitoring result, owner review and final outcome. That record lets risk, compliance, audit, technology and operations speak from the same facts.

Banking practice note: consent and lawful use

For feature store lookup and feature calculation, consent and lawful use is not a side detail. It decides whether the bank can connect the AI or ML output to a real banking purpose, a real customer or regulatory outcome and a real accountable owner. Study the topic as a banking process first and a model process second.

Trace one item from balance trend to calculation log. Then ask which control from definition version proves the item was complete, current, authorised, relevant and fit for use. If that trace cannot be shown without manual guessing, the process is not yet bank-grade.

Banking practice note: source lineage

For feature store lookup and feature calculation, source lineage is not a side detail. It decides whether the bank can connect the AI or ML output to a real banking purpose, a real customer or regulatory outcome and a real accountable owner. Study the topic as a banking process first and a model process second.

Trace one item from repayment history to feature snapshot. Then ask which control from calculation test proves the item was complete, current, authorised, relevant and fit for use. If that trace cannot be shown without manual guessing, the process is not yet bank-grade.

Banking practice note: KYC and identity

For feature store lookup and feature calculation, KYC and identity is not a side detail. It decides whether the bank can connect the AI or ML output to a real banking purpose, a real customer or regulatory outcome and a real accountable owner. Study the topic as a banking process first and a model process second.

Trace one item from utilisation ratio to version record. Then ask which control from freshness rule proves the item was complete, current, authorised, relevant and fit for use. If that trace cannot be shown without manual guessing, the process is not yet bank-grade.

Banking practice note: account behaviour

For feature store lookup and feature calculation, account behaviour is not a side detail. It decides whether the bank can connect the AI or ML output to a real banking purpose, a real customer or regulatory outcome and a real accountable owner. Study the topic as a banking process first and a model process second.

Trace one item from transaction velocity to test result. Then ask which control from point-in-time lock proves the item was complete, current, authorised, relevant and fit for use. If that trace cannot be shown without manual guessing, the process is not yet bank-grade.

Banking practice note: feature freshness

For feature store lookup and feature calculation, feature freshness is not a side detail. It decides whether the bank can connect the AI or ML output to a real banking purpose, a real customer or regulatory outcome and a real accountable owner. Study the topic as a banking process first and a model process second.

Trace one item from KYC risk flag to model input file. Then ask which control from reuse approval proves the item was complete, current, authorised, relevant and fit for use. If that trace cannot be shown without manual guessing, the process is not yet bank-grade.

Banking practice note: point-in-time correctness

For feature store lookup and feature calculation, point-in-time correctness is not a side detail. It decides whether the bank can connect the AI or ML output to a real banking purpose, a real customer or regulatory outcome and a real accountable owner. Study the topic as a banking process first and a model process second.

Trace one item from delinquency marker to approval note. Then ask which control from lineage retention proves the item was complete, current, authorised, relevant and fit for use. If that trace cannot be shown without manual guessing, the process is not yet bank-grade.

Banking practice note: model version

For feature store lookup and feature calculation, model version is not a side detail. It decides whether the bank can connect the AI or ML output to a real banking purpose, a real customer or regulatory outcome and a real accountable owner. Study the topic as a banking process first and a model process second.

Trace one item from feature timestamp to feature definition. Then ask which control from feature owner proves the item was complete, current, authorised, relevant and fit for use. If that trace cannot be shown without manual guessing, the process is not yet bank-grade.

Banking practice note: decision threshold

For feature store lookup and feature calculation, decision threshold is not a side detail. It decides whether the bank can connect the AI or ML output to a real banking purpose, a real customer or regulatory outcome and a real accountable owner. Study the topic as a banking process first and a model process second.

Trace one item from income stability to calculation log. Then ask which control from definition version proves the item was complete, current, authorised, relevant and fit for use. If that trace cannot be shown without manual guessing, the process is not yet bank-grade.

Banking practice note: reason code

For feature store lookup and feature calculation, reason code is not a side detail. It decides whether the bank can connect the AI or ML output to a real banking purpose, a real customer or regulatory outcome and a real accountable owner. Study the topic as a banking process first and a model process second.

Trace one item from balance trend to feature snapshot. Then ask which control from calculation test proves the item was complete, current, authorised, relevant and fit for use. If that trace cannot be shown without manual guessing, the process is not yet bank-grade.

Banking practice note: human review

For feature store lookup and feature calculation, human review is not a side detail. It decides whether the bank can connect the AI or ML output to a real banking purpose, a real customer or regulatory outcome and a real accountable owner. Study the topic as a banking process first and a model process second.

Trace one item from repayment history to version record. Then ask which control from freshness rule proves the item was complete, current, authorised, relevant and fit for use. If that trace cannot be shown without manual guessing, the process is not yet bank-grade.

Banking practice note: fraud control

For feature store lookup and feature calculation, fraud control is not a side detail. It decides whether the bank can connect the AI or ML output to a real banking purpose, a real customer or regulatory outcome and a real accountable owner. Study the topic as a banking process first and a model process second.

Trace one item from utilisation ratio to test result. Then ask which control from point-in-time lock proves the item was complete, current, authorised, relevant and fit for use. If that trace cannot be shown without manual guessing, the process is not yet bank-grade.

Banking practice note: sanctions control

For feature store lookup and feature calculation, sanctions control is not a side detail. It decides whether the bank can connect the AI or ML output to a real banking purpose, a real customer or regulatory outcome and a real accountable owner. Study the topic as a banking process first and a model process second.

Trace one item from transaction velocity to model input file. Then ask which control from reuse approval proves the item was complete, current, authorised, relevant and fit for use. If that trace cannot be shown without manual guessing, the process is not yet bank-grade.

Banking practice note: AML control

For feature store lookup and feature calculation, AML control is not a side detail. It decides whether the bank can connect the AI or ML output to a real banking purpose, a real customer or regulatory outcome and a real accountable owner. Study the topic as a banking process first and a model process second.

Trace one item from KYC risk flag to approval note. Then ask which control from lineage retention proves the item was complete, current, authorised, relevant and fit for use. If that trace cannot be shown without manual guessing, the process is not yet bank-grade.

Banking practice note: fair lending

For feature store lookup and feature calculation, fair lending is not a side detail. It decides whether the bank can connect the AI or ML output to a real banking purpose, a real customer or regulatory outcome and a real accountable owner. Study the topic as a banking process first and a model process second.

Trace one item from delinquency marker to feature definition. Then ask which control from feature owner proves the item was complete, current, authorised, relevant and fit for use. If that trace cannot be shown without manual guessing, the process is not yet bank-grade.

Banking practice note: regulatory reporting

For feature store lookup and feature calculation, regulatory reporting is not a side detail. It decides whether the bank can connect the AI or ML output to a real banking purpose, a real customer or regulatory outcome and a real accountable owner. Study the topic as a banking process first and a model process second.

Trace one item from feature timestamp to calculation log. Then ask which control from definition version proves the item was complete, current, authorised, relevant and fit for use. If that trace cannot be shown without manual guessing, the process is not yet bank-grade.

Banking practice note: operational exception

For feature store lookup and feature calculation, operational exception is not a side detail. It decides whether the bank can connect the AI or ML output to a real banking purpose, a real customer or regulatory outcome and a real accountable owner. Study the topic as a banking process first and a model process second.

Trace one item from income stability to feature snapshot. Then ask which control from calculation test proves the item was complete, current, authorised, relevant and fit for use. If that trace cannot be shown without manual guessing, the process is not yet bank-grade.

Banking practice note: customer harm

For feature store lookup and feature calculation, customer harm is not a side detail. It decides whether the bank can connect the AI or ML output to a real banking purpose, a real customer or regulatory outcome and a real accountable owner. Study the topic as a banking process first and a model process second.

Trace one item from balance trend to version record. Then ask which control from freshness rule proves the item was complete, current, authorised, relevant and fit for use. If that trace cannot be shown without manual guessing, the process is not yet bank-grade.

Banking practice note: audit trail

For feature store lookup and feature calculation, audit trail is not a side detail. It decides whether the bank can connect the AI or ML output to a real banking purpose, a real customer or regulatory outcome and a real accountable owner. Study the topic as a banking process first and a model process second.

Trace one item from repayment history to test result. Then ask which control from point-in-time lock proves the item was complete, current, authorised, relevant and fit for use. If that trace cannot be shown without manual guessing, the process is not yet bank-grade.

Banking practice note: data quality

For feature store lookup and feature calculation, data quality is not a side detail. It decides whether the bank can connect the AI or ML output to a real banking purpose, a real customer or regulatory outcome and a real accountable owner. Study the topic as a banking process first and a model process second.

Trace one item from utilisation ratio to model input file. Then ask which control from reuse approval proves the item was complete, current, authorised, relevant and fit for use. If that trace cannot be shown without manual guessing, the process is not yet bank-grade.

Banking practice note: privacy minimisation

For feature store lookup and feature calculation, privacy minimisation is not a side detail. It decides whether the bank can connect the AI or ML output to a real banking purpose, a real customer or regulatory outcome and a real accountable owner. Study the topic as a banking process first and a model process second.

Trace one item from transaction velocity to approval note. Then ask which control from lineage retention proves the item was complete, current, authorised, relevant and fit for use. If that trace cannot be shown without manual guessing, the process is not yet bank-grade.

Banking practice note: committee reporting

For feature store lookup and feature calculation, committee reporting is not a side detail. It decides whether the bank can connect the AI or ML output to a real banking purpose, a real customer or regulatory outcome and a real accountable owner. Study the topic as a banking process first and a model process second.

Trace one item from KYC risk flag to feature definition. Then ask which control from feature owner proves the item was complete, current, authorised, relevant and fit for use. If that trace cannot be shown without manual guessing, the process is not yet bank-grade.

Banking practice note: reconciliation

For feature store lookup and feature calculation, reconciliation is not a side detail. It decides whether the bank can connect the AI or ML output to a real banking purpose, a real customer or regulatory outcome and a real accountable owner. Study the topic as a banking process first and a model process second.

Trace one item from delinquency marker to calculation log. Then ask which control from definition version proves the item was complete, current, authorised, relevant and fit for use. If that trace cannot be shown without manual guessing, the process is not yet bank-grade.

Banking practice note: exception ownership

For feature store lookup and feature calculation, exception ownership is not a side detail. It decides whether the bank can connect the AI or ML output to a real banking purpose, a real customer or regulatory outcome and a real accountable owner. Study the topic as a banking process first and a model process second.

Trace one item from feature timestamp to feature snapshot. Then ask which control from calculation test proves the item was complete, current, authorised, relevant and fit for use. If that trace cannot be shown without manual guessing, the process is not yet bank-grade.

Banking practice note: monitoring cadence

For feature store lookup and feature calculation, monitoring cadence is not a side detail. It decides whether the bank can connect the AI or ML output to a real banking purpose, a real customer or regulatory outcome and a real accountable owner. Study the topic as a banking process first and a model process second.

Trace one item from income stability to version record. Then ask which control from freshness rule proves the item was complete, current, authorised, relevant and fit for use. If that trace cannot be shown without manual guessing, the process is not yet bank-grade.

Banking practice note: closure evidence

For feature store lookup and feature calculation, closure evidence is not a side detail. It decides whether the bank can connect the AI or ML output to a real banking purpose, a real customer or regulatory outcome and a real accountable owner. Study the topic as a banking process first and a model process second.

Trace one item from balance trend to test result. Then ask which control from point-in-time lock proves the item was complete, current, authorised, relevant and fit for use. If that trace cannot be shown without manual guessing, the process is not yet bank-grade.

Banking practice note: policy retrieval

For feature store lookup and feature calculation, policy retrieval is not a side detail. It decides whether the bank can connect the AI or ML output to a real banking purpose, a real customer or regulatory outcome and a real accountable owner. Study the topic as a banking process first and a model process second.

Trace one item from repayment history to model input file. Then ask which control from reuse approval proves the item was complete, current, authorised, relevant and fit for use. If that trace cannot be shown without manual guessing, the process is not yet bank-grade.

Banking practice note: investigator feedback

For feature store lookup and feature calculation, investigator feedback is not a side detail. It decides whether the bank can connect the AI or ML output to a real banking purpose, a real customer or regulatory outcome and a real accountable owner. Study the topic as a banking process first and a model process second.

Trace one item from utilisation ratio to approval note. Then ask which control from lineage retention proves the item was complete, current, authorised, relevant and fit for use. If that trace cannot be shown without manual guessing, the process is not yet bank-grade.

Banking practice note: model drift

For feature store lookup and feature calculation, model drift is not a side detail. It decides whether the bank can connect the AI or ML output to a real banking purpose, a real customer or regulatory outcome and a real accountable owner. Study the topic as a banking process first and a model process second.

Trace one item from transaction velocity to feature definition. Then ask which control from feature owner proves the item was complete, current, authorised, relevant and fit for use. If that trace cannot be shown without manual guessing, the process is not yet bank-grade.

Banking practice note: control attestation

For feature store lookup and feature calculation, control attestation is not a side detail. It decides whether the bank can connect the AI or ML output to a real banking purpose, a real customer or regulatory outcome and a real accountable owner. Study the topic as a banking process first and a model process second.

Trace one item from KYC risk flag to calculation log. Then ask which control from definition version proves the item was complete, current, authorised, relevant and fit for use. If that trace cannot be shown without manual guessing, the process is not yet bank-grade.

Banking practice note: customer purpose

Trace one item from delinquency marker to feature snapshot. Then ask which control from calculation test proves the item was complete, current, authorised, relevant and fit for use. If that trace cannot be shown without manual guessing, the process is not yet bank-grade.

Banking practice note: consent and lawful use

Trace one item from feature timestamp to version record. Then ask which control from freshness rule proves the item was complete, current, authorised, relevant and fit for use. If that trace cannot be shown without manual guessing, the process is not yet bank-grade.

Banking practice note: source lineage

Trace one item from income stability to test result. Then ask which control from point-in-time lock proves the item was complete, current, authorised, relevant and fit for use. If that trace cannot be shown without manual guessing, the process is not yet bank-grade.

Banking practice note: KYC and identity

Trace one item from balance trend to model input file. Then ask which control from reuse approval proves the item was complete, current, authorised, relevant and fit for use. If that trace cannot be shown without manual guessing, the process is not yet bank-grade.

Banking practice note: account behaviour

Trace one item from repayment history to approval note. Then ask which control from lineage retention proves the item was complete, current, authorised, relevant and fit for use. If that trace cannot be shown without manual guessing, the process is not yet bank-grade.

Banking practice note: feature freshness

Trace one item from utilisation ratio to feature definition. Then ask which control from feature owner proves the item was complete, current, authorised, relevant and fit for use. If that trace cannot be shown without manual guessing, the process is not yet bank-grade.

Banking practice note: point-in-time correctness

Trace one item from transaction velocity to calculation log. Then ask which control from definition version proves the item was complete, current, authorised, relevant and fit for use. If that trace cannot be shown without manual guessing, the process is not yet bank-grade.

Banking practice note: model version

Trace one item from KYC risk flag to feature snapshot. Then ask which control from calculation test proves the item was complete, current, authorised, relevant and fit for use. If that trace cannot be shown without manual guessing, the process is not yet bank-grade.

Banking practice note: decision threshold

Trace one item from delinquency marker to version record. Then ask which control from freshness rule proves the item was complete, current, authorised, relevant and fit for use. If that trace cannot be shown without manual guessing, the process is not yet bank-grade.

Banking practice note: reason code

Trace one item from feature timestamp to test result. Then ask which control from point-in-time lock proves the item was complete, current, authorised, relevant and fit for use. If that trace cannot be shown without manual guessing, the process is not yet bank-grade.

Banking practice note: human review

Trace one item from income stability to model input file. Then ask which control from reuse approval proves the item was complete, current, authorised, relevant and fit for use. If that trace cannot be shown without manual guessing, the process is not yet bank-grade.

Banking practice note: fraud control

Trace one item from balance trend to approval note. Then ask which control from lineage retention proves the item was complete, current, authorised, relevant and fit for use. If that trace cannot be shown without manual guessing, the process is not yet bank-grade.

Banking practice note: sanctions control

Trace one item from repayment history to feature definition. Then ask which control from feature owner proves the item was complete, current, authorised, relevant and fit for use. If that trace cannot be shown without manual guessing, the process is not yet bank-grade.

Banking practice note: AML control

Trace one item from utilisation ratio to calculation log. Then ask which control from definition version proves the item was complete, current, authorised, relevant and fit for use. If that trace cannot be shown without manual guessing, the process is not yet bank-grade.

Banking practice note: fair lending

Trace one item from transaction velocity to feature snapshot. Then ask which control from calculation test proves the item was complete, current, authorised, relevant and fit for use. If that trace cannot be shown without manual guessing, the process is not yet bank-grade.

A lookup at 10:03

For Sravanthi's application, the service requests a dated income measure and repayment history. A feature response includes value, unit, source and transformation versions, observation window and validity. A previous loan's late payment posted after the application cutoff cannot appear in the original score vector. An unavailable repayment source is not the same as a verified zero missed payments.

Compare the online lookup with an offline point-in-time reconstruction. Test a source correction, stale cache and ambiguous account relationship. A cached value within its approved age can be used; an old value outside the limit should trigger referral or fallback. Preserve the served vector in protected decision evidence and make a later corrected reconstruction separately available for impact review.