AI in Finance

References

A synthesis of selected conference talks and technical references. Citations link to the source material; they do not imply that every talk on this subject is included.

1. AASB 132 Financial Instruments: Presentation

   A financial instrument is a contract creating a financial asset for one entity and a financial liability or equity instrument for another. Financial assets include contractual rights to receive cash and equity instruments of another entity. The application guidance gives loans, bonds, notes and trade receivables/payables as examples of corresponding rights and obligations.

2. Basel Core Principles: Principle 26, Internal control and audit

   Principle 26 separates committing a bank, paying funds and accounting for assets and liabilities. Its criteria cover explicit delegation, approval limits, reconciliation, segregation of duties, cross-checking and dual control. Control functions need sufficient expertise, authority and resources relative to business activity. Independent internal audit assesses whether controls remain appropriate and effective, with access to relevant records and authority to assess outsourced functions.

3. Artificial Intelligence in UK Financial Services — 2024

   The Bank of England and FCA’s 2024 survey of 118 firms found AI uses across operations, retail banking, insurance, risk and other business areas. Internal-process optimization, cybersecurity and fraud detection were prominent applications. Respondents perceived current benefits especially in data analysis, anti-money-laundering and fraud work, and cybersecurity, while anticipating further operational-efficiency benefits. Their leading concerns included privacy, data quality and security. The survey separately classified application materiality and automation level: using AI did not necessarily mean granting autonomous decision authority.

4. FINRA Regulatory Notice 24-09

   FINRA identifies potential assistance with locating policies, summarizing research and obtaining issuer information from filings and earnings calls. It reminds member firms that existing technology-neutral obligations apply to GenAI, including third-party tools. Supervisory systems must be reasonably designed for the business; GenAI used in supervision calls for attention to governance, privacy, integrity, reliability and accuracy. Firms should evaluate tools before deployment, and public-communication standards apply whether content is generated by people or technology.

5. Microsoft Dynamics 365: Reconcile bank statements using advanced bank reconciliation

   The workflow compares bank-statement lines with internal bank transactions, maintaining matched and unmatched records. It supports grouped matches, date constraints and configured penny tolerances. Bank-originated fees or interest can be marked new and posted later; marking new is not itself posting. Crucially, a statement can be marked reconciled while unmatched items carry forward. Reconciliation can automatically post correction amounts, so its completion action can change accounting records. Documented reversal through a new transaction differs from editing an already reconciled statement.

6. CFPB: Chatbots in consumer finance

   The CFPB documents risks from inaccurate chatbot responses, failure to understand customer requests and obstacles to timely human assistance. A chatbot offering only generic policies may leave an account-specific problem unresolved. The report describes consumers encountering repeated automated interactions and difficult human-service offramps. These observations support evaluating whether a financial problem is resolved and whether escalation is usable, rather than counting a generated answer as successful service.

7. The Authorizer’s Assistant: A Knowledge-Based Credit Authorization System for American Express

   Dzierzanowski, Chrisman, MacKinnon and Klahr’s 1989 report describes American Express and Inference Corporation’s rule-based Authorizer’s Assistant, fully deployed for U.S. Personal and Gold cards in 1988. It addressed authorizers’ need to combine scattered account records with credit and fraud policies. The system both resolved selected transactions autonomously and advised human authorizers. Its screens combined account activity, recommendations, explanations and applicable authorization policy. Pilot evaluation, mainframe integration, acceptance testing and transfer of business control to operations were integral to deployment. Keeping much of its functionality outside the existing authorization system limited disruption to that critical system.

8. Credit Card Fraud Detection: A Realistic Modeling and a Novel Learning Strategy

   Dal Pozzolo, Boracchi, Caelen, Alippi and Bontempi describe a Worldline-informed workflow combining transaction-blocking rules, scoring rules, a learned model and investigators. Authorization checks precede later fraud investigation; an alert is not itself a fraud determination. Investigators inspect case records and contact cardholders, creating recent but selectively observed feedback. Other labels arrive later through disputes or an operational convention treating sufficiently old undisputed transactions as genuine. The paper evaluates precision among the limited number of cards investigators can examine, avoiding counting several transactions on one card as independent investigations. Separately learning from recent feedback and delayed records improved alert precision in its experiments.

9. David Durand: Credit-Rating Formulae

   In this 1941 NBER chapter, David Durand investigated statistically derived credit-rating formulae as an alternative to weights assembled through experience and intuition. Combining borrower attributes helped distinguish good and bad loans in the studied samples. Durand presented the formulae as supplements to investigation, not replacements for judgment. He explicitly warned that the data contained only previously approved loans and omitted important information such as payment history, limiting use on the broader applicant population.

10. When Is a Liability Not a Liability? Textual Analysis, Dictionaries, and 10-Ks

    Tim Loughran and Bill McDonald’s February 2011 paper examined whether general-purpose negative-word lists measured financial-document tone. Across 50,115 firm-year 10-Ks from 1994–2008, 73.8% of Harvard-list negative-word counts came from words typically not negative in financial usage. Terms such as liability and capital could describe ordinary accounting or an industry rather than adverse events. Their finance-specific negative list showed relationships with filing-period returns that simple proportional Harvard-list counts did not. The contribution was domain-appropriate measurement, not merely a larger dictionary.

11. GLEIF and ANNA: ISIN Issuer to LEI Mapping Table Factsheet

    An ISIN identifies a financial or referential instrument; an LEI identifies a legal entity participating in financial transactions. The mapping connects an instrument’s ISIN to its issuer’s LEI, preserving the distinction between issuer and instrument. This 2021 specification describes daily full files containing eligible active ISINs with available mappings from participating numbering agencies. Changes to identity fields trigger review; inactive instruments and specified entity-status changes can remove or replace mappings. Missing correspondence therefore does not establish that an instrument or entity does not exist.

12. DTC Distributions Service Guide

    DTC distinguishes announced distributions, calculated entitlements and actual allocations. Entitlement records include captured positions, rates, calculated amounts and subsequent adjustments. Interim accounting can transfer entitlement between sellers and buyers after the record date, including for many bond-interest payments. Participants must verify announcement information and report discrepancies. DTC allocates money after receiving funds from the paying agent, trustee or issuer; a scheduled payable date alone is not evidence of receipt. Later rate or position corrections can generate additional credits or debits. These distinctions support reconciling the instrument event, eligible position, expected amount and actual allocation separately.

13. SEC Beginners’ Guide to Financial Statements

    A balance sheet records assets, liabilities and shareholders’ equity at a point in time, with assets equal to liabilities plus equity. An income statement summarizes revenues and expenses over a period; a cash-flow statement tracks cash movements through operating, investing and financing activities. Profit and cash are therefore not interchangeable quantities. For financial AI, extracting a number also requires its statement, entity, period, unit and accounting context before comparing it or using it in a calculation.

14. XBRL International: xBRL-JSON tutorial and examples

    The published monetary fixture represents assets of EUR 1,230,000 with a fact identifier, entity identifier, concept, reporting instant and declared numerical precision. Its instant, 2020-01-01 at midnight, represents the end of December 31, 2019. The tutorial distinguishes instant from duration facts, supports additional dimensions and footnote links, and represents numerical values as strings to avoid precision loss. These fields preserve the meaning that a bare number cannot convey.

15. FinQA: A Dataset of Numerical Reasoning over Financial Data

    FinQA asks numerical questions over financial-report text and tables and annotates answers with reasoning programs. Its baseline first retrieves supporting facts, then generates an executable sequence of operations. This separates selecting the right evidence from performing arithmetic. A structurally valid program can still select the wrong numbers or operations, so answer checking must examine inputs, periods, units, and calculation semantics. The dataset illustrates why fluent financial prose is not equivalent to auditable numerical reasoning.

16. FRED real-time periods and historical information sets

    Economic observations have both the period they describe and a real-time period recording when the information was known before changing. FRED generally defaults to information available today, including revisions to historical observations. ALFRED-style queries change realtime_start and realtime_end to retrieve historically known information; both boundaries are inclusive. Setting both to a historical date requests that date's information set. Therefore, an observation describing an earlier quarter can still be unavailable at a decision date, and its later revised value must not replace its historical vintage in a replay.

17. IFRS Foundation: IAS 8 Basis of Preparation of Financial Statements

    Prior-period errors arise from omitted or misstated information caused by failing to use, or misusing, reliable information that was available. The IAS 8 overview describes correcting material prior-period errors retrospectively through restated comparative amounts, unless determining the effects is impracticable. A later report can therefore change the reported value for an earlier period.

18. SEC: EDGAR Application Programming Interfaces

    EDGAR identifies filers by CIK and exposes submissions and extracted financial facts. Company-concept responses distinguish taxonomy concepts and units, including different currencies. The aggregated XBRL APIs cover non-custom taxonomy facts applying to the whole filing entity. Frames select the last-filed fact best fitting a requested calendar period; company reporting calendars can differ. Thus a historical calendar frame is not inherently an information set available at that historical date, and these aggregates do not represent every custom or segment disclosure.

19. CME Group: Market Data Explanation and Disclaimer

    CME's website terms restrict redistribution, systematic extraction, derivative uses and other processing of covered market data, with specified uses requiring written authorization. They expressly prohibit using covered data and metadata for machine learning or AI, including training and generating outputs. Public visibility of a price therefore does not itself establish permission to ingest it into an AI application.

20. NIST Privacy Framework 1.0: lifecycle and minimized audit evidence

    The framework inventories data elements, processing purposes, actions, owners and flows. Policies define permitted uses and retention periods; the data lifecycle aligns with system development and operations. Authorizations must be maintained and revocable, access limited by least privilege, and deletion and destruction performed under policy. Audit records themselves must incorporate data minimization. Engineering application: define the decision evidence needed for review, its purpose, authorized readers, retention trigger and disposal method before logging. Retain the necessary decision, model and policy versions and relevant evidence without indiscriminately copying personal data into logs, prompts or backups. Where review requires sensitive evidence, constrain fields, access and retention rather than treating auditability as permission to keep everything. Assess removal and disclosure across downstream copies and service providers.

21. SEC: Selective Disclosure and Insider Trading

    The SEC's adopting release explains materiality through substantial likelihood that a reasonable investor would consider information important or that it would significantly alter the available information mix. Information is nonpublic when it has not been disseminated so investors generally can access it. The release rejects an exhaustive materiality checklist and discusses information barriers that prevent investment decision makers from receiving material nonpublic information.

22. How to Build Planning Agents Without Losing Control - Yogendra Miraje, FactSet

    The earnings-call preparation example separates natural-language tasks from executable function calls and passes task context through the workflow.

23. Trust, but Verify: High-Fidelity Reasoning in Agentic Workflows

    Distill document findings, then verify individual findings in separate calls before synthesizing them.

24. How Kepler Built Verifiable AI for Financial Services

    Use the model to decide what to compute and deterministic tools to extract and calculate the required values.

25. How Kepler Built Verifiable AI for Financial Services

    Derived numbers need replayable derivation histories spanning source values, calculations, and internal information, not just links to filings.

26. SEC: Non-GAAP Financial Measures Compliance and Disclosure Interpretations

    SEC staff explains that non-GAAP measures can mislead when adjustments are inconsistent between periods, when charges are excluded without comparable gains, or when labels conceal different calculations. A changed adjustment may require explanation and, depending on significance, recasting earlier measures. Similarly named non-GAAP measures are not necessarily comparable across companies. Research comparisons therefore need the measure's definition and adjustment history alongside its numerical value.

27. Trust, but Verify: High-Fidelity Reasoning in Agentic Workflows

    Make individual findings interrogable and expose an audit trail beyond the final report, allowing users to request detail where their judgment identifies a concern.

28. Double-entry bookkeeping and cash movement

    A journal entry posts amounts to accounts with total debits equal to total credits. Debits normally increase assets and expenses; credits normally increase liabilities, equity and revenue. Opposite entries decrease those balances. Revenue and expense balances ultimately affect equity, preserving assets = liabilities + equity. Borrowing $1,000 debits Cash and credits Notes Payable: assets and liabilities both increase. Providing a service on credit debits Accounts Receivable and credits Service Revenue, without receiving cash. Later collection debits Cash and credits Accounts Receivable; it exchanges one asset for another without recording revenue again. Thus a balanced posting records economic effects but need not represent a cash transfer.

29. Basel Committee: credit risk and contractual obligations

    Credit risk is the possibility that a borrower or counterparty does not meet agreed obligations. Loans are one source, but commitments, guarantees and settlement also expose banks. A failed settlement can involve credit, liquidity, market and operational risks together. Engineering implication: retain obligation, due-time, counterparty and settlement-state records rather than treating every financial loss as an incorrect prediction.

30. Microsoft Dynamics 365: Set up bank reconciliation matching rules

    Matching rules filter statement and internal transaction lines using specified criteria, and rule sets execute in sequence. By default, a rule matches the first qualifying bank transaction. A configuration option requires manual matching when multiple documents match on amount. This supplies a concrete rule-based baseline and shows why an apparently successful match can require ambiguity handling.

31. AI-Driven Multi-Document Correlation for Enterprise Financial Compliance and Fraud Detection

    Deployment must connect existing enterprise data systems and fit prioritized risk cases into the investigator's audit workflow.

32. DBS Empowers Its Customer Service Officers with a Gen AI-Powered Virtual Assistant

    DBS’s July 18, 2024 announcement describes an in-house assistant combining telephone speech recognition, real-time transcription and searches of the bank’s knowledge base. Customer-service officers receive query-specific information; the assistant also summarizes calls and prefills service-request fields. Pilots began in October 2023, with wider Singapore deployment planned before the end of 2024. The workflow targets searching and documentation around a human service interaction rather than replacing the entire customer-resolution process.

33. Your Agent Didn’t Fail. Your Harness Did.

    Delivery alone is insufficient: a named system of record must persist the fact and support replay into future work.

34. Chase: What are pending transactions and how long do they take?

    A pending credit-card transaction has not been fully processed or posted. Examples include restaurant tips, fuel purchases and merchants waiting for shipment. The amount can still change, and the account balance may not yet reflect it. Posting time depends on the merchant, issuer, network and transaction type.

35. Chase: Disputing a Charge

    Chase's operational FAQ distinguishes dispute routes by product and status. A credit-card charge must post before a dispute can open. A pending debit-card transaction can be disputed by telephone, while online debit disputes require posting. An assistant that applies the credit-card rule to every transaction would give incorrect service instructions.

36. CFPB Regulation E: § 1005.11 Procedures for resolving errors

    The official interpretation distinguishes checking whether an electronic transfer was credited or debited, without alleging an error, from communications that trigger error-resolution procedures. Reporting a lost access device triggers those procedures when the consumer also alleges possible unauthorized use. A properly asserted error still requires handling after account closure. Classifying the customer's request therefore affects the required servicing path.

37. SAS credit scorecards: grouped attributes and additive points

    A scorecard assigns points to borrower attributes, such as an age interval within an age characteristic. The total adds the applicable attribute points. SAS first groups predictor values into bins, fits logistic regression to those grouped variables, then scales its fitted parameters into score points. This connects the familiar additive scorecard to a statistical model rather than an unrelated collection of hand-written rules. Under the guide's convention, higher totals indicate lower modeled default risk. Grouping, fitted coefficients and scaling are separate choices; a decision policy then determines what action follows a score.

38. Conditional estimates versus decision thresholds

    predict_proba estimates P(Y=1|X=x); decision_function supplies scores that need not be probabilities. Ranking orders cases by score, while an action rule compares a score with a threshold. Changing the threshold leaves probabilities and ranking curves unchanged but changes predicted labels. The documented threshold tuner optimizes a chosen metric using cross-validation. Decision-theoretic derivation: with zero cost for correct decisions, false-positive cost C_FP and false-negative cost C_FN, positive action has expected cost C_FP(1-p), versus C_FN p for negative action. Choose positive when p > C_FP/(C_FP+C_FN), assuming positive total cost. Changing costs changes this threshold without changing p.

39. Calibration, reliability diagrams and discrimination

    Binary calibration means P(Y=1|p_hat=p)=p: among cases assigned approximately 0.8, approximately 80% should be positive. Reliability diagrams bin predictions and compare each bin's mean probability with its observed positive fraction; sample counts reveal thinly supported bins. Brier loss, mean((p_hat-y)^2), and log loss assess probabilistic predictions but combine calibration with discrimination and outcome uncertainty, so a lower total score does not isolate better calibration. Discrimination concerns ordering: a strictly increasing transformation can preserve ranking while changing probability values and calibration. Calibrators must use predictions from data independent of fitting the underlying classifier.

40. CFPB Regulation B: § 1002.9 Notifications

    For covered adverse-action statements, Regulation B requires specific principal reasons; merely citing internal policy or an insufficient qualifying score is inadequate. The official interpretation requires reasons to accurately describe factors actually considered or scored. For combined scoring and human judgment, reasons must come from the component responsible for the denial, or both components when both contributed through the described gray-band process. This makes the actual decision path essential: a plausible generated explanation cannot substitute for the reasons that produced the action.

41. Basel Committee: funding and market liquidity

    Funding liquidity concerns meeting expected and unexpected cash-flow and collateral needs without harming operations or financial condition. Market liquidity concerns the ability to offset a position near the market price despite trading depth and disruptions. A bank that appears solvent can still face a liquidity shortage. Credit, market and operational problems can affect liquidity, so cash availability and timing must be examined alongside asset value and expected losses. These concepts complement operational-risk diagnosis rather than forming mutually exclusive incident labels.

42. Basel Committee: sound management of operational risk

    Operational risk concerns losses caused by inadequate or failed internal processes, people or systems, or by external events. The Basel definition includes legal risk but excludes strategic and reputational risk, although management should consider those where appropriate. It is inherent across banking activities, rather than confined to a technology team. Technology failures, destructive cyber incidents and natural disasters can disrupt operations. Engineering illustration: an erroneous payment-routing process can create loss even when the intended borrower pays and market prices remain unchanged. The diagnostic question is which process or control failed, distinct from counterparty performance, price exposure or funding availability. Risk functions must coordinate because these loss channels can interact.

43. Basel Committee: Explanatory Note on Minimum Capital Requirements for Market Risk

    Market risk concerns losses from movements in market prices of positions, such as trading-book instruments. Value at risk describes a loss cutoff at a chosen percentile; expected shortfall considers the average loss in the tail beyond that cutoff. Equal cutoffs can therefore hide different severities of extreme losses. The Basel note also connects market risk to liquidity horizons: exiting or hedging a less liquid position under stress may take longer and affect prices. Sparse observations make some risks difficult to model, motivating separate stress treatment rather than treating missing history as evidence of no risk.

44. Federal Reserve: Credit loss-estimation methodologies

    Expected credit loss depends on probability of default, loss given default, and exposure at default. These components can have different drivers: borrower condition influences default, collateral affects recovery, and unused commitments can change exposure. The Federal Reserve’s historical discussion shows why a single default ranking is not a complete loss estimate. It also distinguishes scenario-conditioned estimates from long-run averages and explains why ordinary historical fit may fail to represent stressed conditions.

45. IT Admin for the AI Workforce — Sarthak Aggarwal, Decawork

    Normalize authenticated intent into a typed, logged plan before exposing execution to untrusted evidence.

46. IT Admin for the AI Workforce — Sarthak Aggarwal, Decawork

    Issue short-lived capabilities for approved actions rather than standing executor credentials, and record the delegation behind each action.

47. Commission Delegated Regulation (EU) 2018/389, Article 5

    Article 5 specifies dynamic linking when payment providers apply strong customer authentication under PSD2 Article 97(2). The payer must see the amount and payee; the authentication code must bind to those agreed details. Changing the amount or payee invalidates the code. The provision also protects the integrity of transaction details throughout authentication.

48. Your Agent Didn’t Fail. Your Harness Did.

    Approval must remain bound to one specific action and its scope, identity, arguments, and lifetime; expiration should terminate the approval path.

49. SEC Staff FAQs Concerning Risk Management Controls for Brokers or Dealers with Market Access

    SEC staff’s market-access explanation distinguishes preventing an order from entering a market from monitoring execution afterward. It describes aggregate customer and broker credit or capital thresholds, erroneous-order checks, restrictions on unauthorized users and restricted securities, and immediate execution reports to surveillance personnel. These controls cover manually entered as well as computer-generated orders. Threshold changes after blocking require evaluation under supervisory procedures and retained reasons. Required controls generally remain under the broker-dealer’s direct and exclusive control, subject to specified exceptions.

50. Morgan Stanley: Launch of AI @ Morgan Stanley Debrief

    Morgan Stanley describes a client-consented meeting workflow in which Debrief generates notes, identifies action items, summarizes discussion, creates an email for an advisor to edit and send at their discretion, and saves a note into Salesforce. This separates drafting a customer communication from the advisor's decision to send it while showing that another output can enter an operational record system.

51. Your Agent Didn’t Fail. Your Harness Did.

    Trace one real run from trigger identity through inherited state, authority, execution attempts, and surviving external evidence.

52. Investor.gov: Bonds — FAQs

    A bond is a debt security: its issuer borrows from investors and promises interest and repayment of principal under its terms. Principal is also called face or par value; maturity is when repayment comes due. Interest payments commonly follow a schedule. These promised payments are not guaranteed receipts: the issuer can fail to pay interest or principal on time. A bond’s market value can also differ from its face value.

53. Your Agent Didn’t Fail. Your Harness Did.

    Internal acceptance does not prove the intended result appeared at the user-visible boundary.

54. Resolving an ambiguous payment request

    A timeout can leave the client unable to tell whether Stripe received or executed a request. Stripe documents retrying with the same key and parameters until a server result is obtained, using backoff. An HTTP 500 remains indeterminate: side effects may exist even though the cached response stays unchanged. Stripe may reconcile partial mutations and emit webhook events for resulting objects. Supplying a local operation identifier in metadata lets the application correlate these objects with its own pending operation. Engineering consequence: preserve pending state until authoritative provider evidence resolves it; do not infer failure solely from a timeout.

55. Stripe retry keys and retention boundaries

    For a mutation, send an idempotency key and reuse that key with identical parameters when retrying an ambiguous connection failure. Stripe stores the first executed request's status and response body, including failures such as HTTP 500, and returns that result for repeated requests. Parameter mismatches produce an error. Keys may be pruned after they are at least 24 hours old; reuse after pruning creates a new request. Results are saved only after endpoint execution begins, so validation failures and conflicts with concurrently executing requests do not create a saved result.

56. Production Evals For Agentic AI Systems

    Evaluate failures across the workflow, including dependencies, planning, tool execution, and multi-agent coordination.

57. Production Evals For Agentic AI Systems

    Apply an SRE or production-engineering lens: assess delivered value, operational reliability, human burden, risk, user experience, scalability, and resilience.

58. FinanceBench: A New Benchmark for Financial Question Answering

    Pranab Islam and colleagues introduced FinanceBench in November 2023, pairing financial questions with reference answers and supporting passages. Their experiments manually assessed 2,400 outputs from 16 configurations on 150 cases, separately recording correct answers, incorrect answers and refusals. Supplying the correct evidence pages improved performance but did not eliminate reasoning errors. Qualitative review also found valid answers differing from the reference. The benchmark therefore separates information access, interpretation and answer assessment instead of treating successful retrieval or exact reference matching as sufficient.

59. The Selective Labels Problem: Evaluating Algorithmic Predictions in the Presence of Unobservables

    Observed outcomes can depend on earlier human decisions, so the labeled sample need not represent the people on whom a replacement model would act. This paper studies bail decisions: a release outcome is observed only for defendants released. Evaluating on that selected subset can misstate performance for a new decision policy. The same structural concern applies to underwriting when repayment outcomes are observed only for granted loans; that financial example is an inference from the mechanism, not an experiment in this paper.

60. AI-Driven Multi-Document Correlation for Enterprise Financial Compliance and Fraud Detection

    The speaker reports approximately 91% precision on an evaluation involving approximately three million financial records collected over five years across four jurisdictions.

61. Rolling-origin forecast evaluation

    Time-series cross-validation moves the forecast origin forward. At each origin T, the training set contains only earlier observations, and forecasts are scored against subsequent observations. Repeating this produces an out-of-sample error series; multi-step evaluation scores the horizon actually needed rather than only one-step predictions. Engineering consequence: preprocessing, model fitting and model selection must respect each origin's information boundary. Using a future observation anywhere in constructing its forecast defeats the chronological split.

62. Lookahead Bias in Pretrained Language Models

    The paper distinguishes historical prompt information from information induced by model pretraining. Its earnings-call experiment finds later pandemic information in risk descriptions generated from pre-pandemic calls. Instructions not to use later information do not eliminate the demonstrated leakage. Masking identifiers also does not guarantee de-identification: models can infer some firm identities and periods. Consequently, restricting retrieved documents to a historical cutoff does not by itself establish that the model's answer uses only historically available information.

63. CRSP10 historical security, distribution and delisting records

    CRSP retains active and inactive securities. PERMNO identifies a security; PERMCO identifies its company. Name-history rows associate tickers and other attributes with effective date ranges, unlike header values describing the latest available state. Distribution records identify dividends, splits and adjustment factors; merger records can link to a replacement security. Delisting records retain dates, reasons and post-delisting values or returns. Reconstruction implication: select securities eligible at the historical date, join identifiers within their effective intervals, and account consistently for distributions and exits. Filtering that universe to today's survivors conditions inclusion on a later outcome. It removes investments that were available then, potentially omitting failed firms and their losses and inflating evaluated returns. Delisting can also follow a merger, so neither every exit nor the bias's magnitude has one universal interpretation.

64. The Probability of Backtest Overfitting

    Testing many strategy configurations and choosing the strongest historical result makes selection itself a source of overfitting. Bailey and colleagues define the probability that an in-sample winner ranks below the median out of sample, then propose combinatorially symmetric cross-validation to estimate that quantity across partitions. The analysis concerns the strategy-selection process, not only parameter fitting inside one model. A single attractive backtest therefore omits evidence about how many alternatives were tried and how stable their rankings were.

65. Basel Committee: Principles for Operational Resilience

    The principles call for mapping critical operations to people, technology, information, processes and external dependencies; testing severe but plausible disruptions; and defining continuity triggers, responsibilities and succession of authority. Third-party contingency plans should consider substitution or bringing services in-house. Incident handling includes severity classification, recovery procedures, communication and lessons learned. Controls should be reassessed after component changes and incidents, and recovery procedures periodically tested and updated.

66. Right censoring and unresolved event outcomes

    Let T be event time and C the end of observation. A right-censored record stores U=min(T,C) and an indicator delta showing whether the event was observed. When delta=0, the record establishes only that the subject remained event-free through C; the later outcome is unknown. Survival models use this partial information rather than treating C as an event time. S(t)=P(T>t) describes survival beyond t. Financial application: for a default-within-H label, an account observed without default only until C<H is unresolved, not a negative example. A negative requires sufficient event-free follow-up through H.

67. Stripe card-dispute timing and lifecycle

    A dispute becomes observable to a merchant when Stripe notifies it through configured channels or the API, after the account owner contests a payment. Cardholders usually have 120 days to initiate a dispute, with exceptions; future-service windows may start at the service date. Early fraud warnings and inquiries can precede formal disputes, but warnings can also arrive afterward. A case can await evidence or issuer review before its status becomes won or lost; rare late wins can revise a loss. An inquiry closed without escalation is not an explicit issuer win. Label-design implication: preserve payment time, observation time, reason and evolving status. At a chosen cutoff, an unresolved case remains unresolved, and no recorded dispute means only that none has been observed by then, not confirmed non-fraud.

68. Stripe dispute categories and contested fraud claims

    Disputes include different reasons, such as duplicate processing, missing products and fraud. Stripe defines the fraudulent category as a cardholder's claim that the payment was unauthorized, and explicitly notes that the cardholder may instead have failed to recognize a legitimate charge. Thus a fraud-coded dispute is an allegation with a particular operational meaning, not infallible confirmation of fraud. For supervised learning, define whether the target is a reported fraud dispute, an adjudicated case outcome or separately investigated fraud. These targets require different evidence and should not be silently substituted for one another.

69. Federal Reserve SR 26-2: Revised Guidance on Model Risk Management

    The guidance describes clear accountability across development, validation and monitoring, supported by policies, model inventories and documentation tracking recommendations, responses, exceptions and remediation. Effective challenge requires expertise, sufficient independence for objectivity and influence; validation quality depends on rigorous review rather than a prescribed organizational structure. Validation generally precedes first use, with controls and stakeholder awareness when urgent use precedes completion. Conceptual review and outcomes analysis assess suitability; change scope and materiality inform validation timing. Monitoring considers changing exposures, clients, data and markets, with adjustment, recalibration or redevelopment when warranted. Footnote 3 excludes generative and agentic AI while including traditional quantitative and non-generative, non-agentic AI models.

70. The IIA’s Refreshed Three Lines Model

    The IIA distinguishes three responsibilities: first-line management owns day-to-day risks and operates controls; second-line specialists provide support, challenge and monitoring; internal audit supplies independent assurance and advice, reporting functionally to the board. An accounts-payable team checking duplicate payments exemplifies a first-line control, not independent audit. Coordination does not eliminate these distinctions, and internal audit remains responsible for its conclusions when relying on another provider’s work. An assurance map records which providers cover which risks, exposing duplication and omissions.

71. The Build-Operate Divide: Bridging Product Vision and AI Operational Reality

    Prioritize human-in-the-loop review at consequential decision points in high-risk, high-trust workflows.

72. Build AI Systems for Discernment, Not Approval - Angel Ortmann Lee, Duolingo

    The When Machines Mislead case study inserted fake copy-typing alerts into legitimate historical exam sessions and found that skilled proctors accepted half of those alerts.

73. Mission-Critical Evals at Scale: Learnings from 100,000 Medical Decisions

    Combine evaluator confidence with contextual risk to prioritize cases for expert review.

74. NIST AI RMF 1.0: accountability, appeals and override

    NIST calls for documented responsibilities and communication, defined human oversight, feedback channels that let affected people report problems and appeal outcomes, and monitoring that includes appeal and override. It also calls for documenting responses to errors; its human-AI discussion identifies override frequency and rationale as useful evidence. Engineering application: attach a dispute to its decision ID and original input/model/policy versions; route it to an accountable reviewer empowered to inspect evidence and authorize correction or override. Record the reviewer, evidence, rationale, revised disposition and downstream correction status, and communicate the result to the affected person. Feed recurring errors into evaluation and remediation. A trace alone does not provide this correction path.

75. Production Evals For Agentic AI Systems

    Use simulated workflow scenarios and score both task completion and execution quality.

76. IT Admin for the AI Workforce — Sarthak Aggarwal, Decawork

    Agents with authority and side effects need an operational lifecycle beyond model behavior testing.

77. IT Admin for the AI Workforce — Sarthak Aggarwal, Decawork

    Separate planning from execution so that reasoning over content does not itself confer authority to invent and execute actions.

78. Moving away from Agile: What's Next?

    Automation can move the bottleneck into human collaboration and manual review while increasing code complexity.

79. The Build-Operate Divide: Bridging Product Vision and AI Operational Reality

    Treat the prototype-to-reliability transition as a recurring operational iteration loop, described here as crossing a quality chasm.

80. SEC: Electronic Recordkeeping Requirements, Release 34-96034

    The 2022 amendments provide an audit-trail alternative to non-rewriteable, non-erasable storage. The alternative preserves timestamped modifications and deletions, action times, applicable individual identities and information needed to recreate original and intermediate records. The final text accommodates automated actions and permits individual identity through a unique identifier. Its purpose is reconstructable, authentic records, not merely a log declaring that an operation succeeded.

81. Agents Need Feature Flags

    Track mitigation effectiveness and record flag changes with enough context to reconstruct an incident.

82. Agents Need Feature Flags

    Ship agent-wide and per-tool kill switches first, and ensure in-flight work checks them at the next decision point.

83. FSB: Financial stability implications of artificial intelligence

    The FSB’s November 2024 assessment identifies third-party dependencies, correlated market behavior, cyber risk, and model risk as channels through which AI adoption can amplify financial vulnerabilities. It distinguishes operational uses, such as document summarization and retrieval, from the broader consequences of common infrastructure and similar decisions across institutions. A system that is useful to one organization can still share failure dependencies with many others, so evaluation should include provider concentration and contingency planning alongside local model quality.

84. Event time and actual availability require separate handling

    This first-party issue reports that created_timestamp was used for deduplication rather than a decision-time availability filter, allowing later backfills to replace historical values. Its proposed predicate requires created_timestamp <= entity_timestamp, independently of event-time TTL. Engineering inference: retain event time, actual availability time and immutable versions; at decision time T, first exclude versions unavailable at T, then select the appropriate event and version. A correction to yesterday's event published tomorrow must not enter today's replay. Availability should represent when the serving system could use the value, not merely the business event's timestamp.

85. How to Build Planning Agents Without Losing Control - Yogendra Miraje, FactSet

    Insert a natural-language blueprint before low-level planning to provide a high-level task breakdown and restrict the planner's available tools.

86. FICO Celebrates 70 Years of Innovation and Says “Hello, Future” to Applied Intelligence

    FICO’s corporate history connects several distinct developments: William Fair and Earl Isaac founded the company in 1956; commercial credit-risk models followed in the late 1950s; application-processing software launched in 1972 to address growing application volumes and slow, uneven manual reviews. The FICO Score, introduced in 1989, supplied a standardized cross-bureau credit-risk measure. The account dates Falcon’s neural-network payment-fraud system to 1992. These distinguish estimating repayment risk, processing applications and detecting suspicious payments rather than treating financial AI as one forecasting task.

87. BloombergGPT: A Large Language Model for Finance

    Shijie Wu, Ozan Irsoy and colleagues introduced BloombergGPT in March 2023 as a 50-billion-parameter language model trained on mixed financial and general-purpose text. The aim extended beyond one sentiment classifier to financial language tasks including entity recognition, entity linking and question answering while retaining general capabilities. Its entity-linking evaluation required identifying company mentions and returning their stock tickers; performance varied by document domain. The paper reports financial and general benchmark results, rather than demonstrating an autonomous investment or transaction workflow.