liveBookmanning.com

What you will learn in Financial AI in Practice

A playbook for credit, fraud, and investment systems

by Taehun Kim

Read it on liveBook Get the book at manning.com

Chapter by chapter

Concepts explained in this book

• fraud detectionThe process of checking activity or applicants for indicators of fraud.
• large language modelA large language model is a language model with many learned parameters, typically trained on extensive text data to process or generate language.
• pipelineAn ordered sequence of data-processing or modeling steps that transforms inputs into outputs.
• graph analyticsA graph-based approach represents entities and their relationships as a graph and analyzes that structure to identify patterns or make predictions.
• class imbalanceA condition in which a classification dataset has substantially different numbers of observations in its classes.
• supervised learningSupervised learning trains models on labeled examples to predict outputs for new inputs.
• retrieval-augmented generationA method that retrieves relevant external information and supplies it to a generative model when producing an output.
• graph neural networkA neural network that learns from graph-structured data by aggregating information across connected nodes.
• community detectionThe identification of groups of graph nodes that are more densely connected to one another than to the rest of the graph.
• defense in depthDefense in depth layers distinct protective controls so that one control can compensate for another's failure or limitations.
• interpretabilityThe degree to which people can understand the basis of a model’s predictions or decisions.
• unsupervised learningUnsupervised learning is a machine-learning approach that identifies structure or patterns in data without relying on labeled outcomes.
• semi-supervised learningSemi-supervised learning uses a limited amount of labeled data together with a larger pool of unlabeled data.
• graphA structure representing entities as nodes and relationships between them as edges.
• AI agentA software system that perceives context, makes decisions, and takes actions to pursue specified goals.
• financial DNAFinancial DNA is the example application's precomputed user profile represented by a vector containing semantic and behavioral dimensions.
• discretizationBinning is the technique of grouping continuous or categorical values into discrete intervals or categories.
• rule-based systemA rules-based system is a computational system whose behavior is determined by explicitly specified if-then rules.
• signalA signal is an output from an analytical model that conveys information for a subsequent financial decision or action.
• precisionPrecision is the proportion of positive classifications made by a system that are correct.
• weight of evidenceWeight of evidence is a transformation that expresses the strength of association between a predictor category and a binary outcome using event and non-event proportions.
• rules engineA software component or system that manages and executes predefined rules against data or events.
• independent variableA measurable input variable or derived attribute supplied to a statistical or machine-learning model.
• thresholdA cutoff value used to divide observations or trigger an action based on a measured value.
• prompt engineeringPrompt engineering is the deliberate design and refinement of instructions supplied to generative models to elicit desired outputs.
• embedding modelAn embedding model is a model that maps content, such as text, to numerical vector representations that encode semantic relationships.
• anomaly detectionThe identification of observations or patterns that substantially differ from expected behavior, including without labeled training examples.
• resamplingChanging a dataset's composition by sampling its observations, often to address class imbalance.
• graph-based pattern discoveryAn analytical technique that represents entities and their relationships as a graph to reveal patterns such as shared connections.
• cost-sensitive learningCost-sensitive learning incorporates differing costs of prediction errors into model training or decision-making.
• survival analysisA statistical approach to analyzing the time until an event occurs while accounting for observations whose event time is unknown or incomplete.
• false-positive rateThe proportion of actual negative cases that a classifier incorrectly identifies as positive.
• monitoringThe ongoing observation of systems, signals, and outcomes to identify changes, failures, or control issues.
• feature engineeringFeature engineering is the process of constructing informative input variables from raw data for statistical or machine-learning models.
• reconstruction lossA reconstruction error is the difference between an input and the output produced when a model attempts to reconstruct that input.
• centralityThe value assigned to a node by a centrality measure.
• guardrailA guardrail is a constraint or control that limits model behavior to reduce unwanted or unsafe outputs.
• four-layer architectureA book-specific architecture framework that organizes a financial AI system into data asset, modeling, application, and strategy and monitoring layers.
• concept driftA change over time in the relationship between input data and the target a model is intended to predict.
• service layerThe architectural layer that integrates analytical outputs into user-facing or operational financial workflows.
• LLMOpsLLMOps is the set of practices for developing, deploying, monitoring, and maintaining systems based on large language models.
• edgeA connection between vertices in a graph, representing a relationship or interaction.
• nodeA vertex in a graph representing an entity or object.
• parallelismParallel processing is the simultaneous execution of multiple computational tasks across processing resources.
• missing valueData values that are absent from a dataset for some records or fields.
• loan lifecycleThe sequence of stages through which a loan proceeds from origination to closure.
• accuracyAccuracy is the proportion or degree of predictions that a model gets correct relative to the evaluated cases or reference outcomes.
• MCPMCP is a protocol for standardizing how AI models or agents interact with external tools and data sources.
• vintage analysisVintage analysis groups loans by a shared origination period and compares their performance over time.
• AI in financeAn application of artificial intelligence to a financial process, product, or decision.
• strategy and monitoring layerThe decision-system layer that interprets model signals against business objectives, policies, regulations, and monitored outcomes.
• fraudFraud is intentional deception intended to secure an unfair or unlawful gain or cause another party a loss.
• time-based analysisAn analytical approach that represents credit outcomes in relation to the time elapsed since loan origination or observation.
• champion-challenger frameworkChampion-challenger testing compares an incumbent model or decision strategy with one or more candidates to evaluate potential improvements or replacements.
• embeddingAn embedding is a numerical vector representation of content that preserves properties useful for similarity comparison and retrieval.
• communityA group of graph nodes more densely connected to one another than to the rest of the graph.
• Louvain algorithmA heuristic algorithm for detecting graph communities by optimizing modularity.
• autoencoderAn autoencoder is a neural network trained to reconstruct its input, typically by encoding it into a lower-dimensional representation and decoding it.
• recallThe proportion of all fraudulent cases found within a selected group of model-ranked cases, such as the top n percent.
• alternative dataAlternative data is information from nontraditional sources used to analyze markets, businesses, or economic conditions.
• information valueInformation value is a measure of how well a variable separates two outcome classes, calculated by summing weighted differences across its categories or bins.
• target driftTarget drift is a change over time in the distribution or proportion of a model’s target variable.
• human-in-the-loopHuman oversight is the involvement of people in reviewing, approving, or supervising system decisions and operation.
• directional accuracyDirectional accuracy is the proportion of evaluated cases in which a prediction correctly identifies the direction of an observed change.
• fundamental dataCompany- and economy-related information derived from financial reports and other measures of underlying financial condition.
• market dataData describing financial markets, including prices, trading volumes, and market activity.
• underwritingCredit underwriting evaluates a prospective borrower's creditworthiness to determine whether and on what terms to extend credit.
• area under the curveThe area under a receiver operating characteristic curve, summarizing a classifier’s ability to distinguish between classes across thresholds.
• neural message passingA graph-learning mechanism that updates node representations by exchanging and aggregating information along graph edges.
• market narrativeA market narrative is a recurring interpretation of events, expectations, fears, or developments that influences how financial participants understand markets.
• data layerAn architecture layer containing governed data assets and the processes that prepare them for analysis and modeling.
• data explorationExploratory data analysis is the examination of data to understand its structure, quality, and distributions before further analysis or modeling.
• personalizationPersonalization is the adaptation of a product, service, or interaction to an individual customer's characteristics or preferences.
• retrainingModel retraining is the process of training an existing model again, typically with updated data, to refresh its learned parameters.
• population stability indexThe population stability index quantifies changes between distributions of a variable or model population across time or samples.
• hallucinationA model hallucination is generated content that is unsupported or misleading but presented as though it were reliable.
• KSThe maximum difference between two cumulative distributions, often used to measure how well a model separates groups.
• chunkingChunking divides documents or other content into smaller segments for processing or retrieval.
• qualitative engineA qualitative engine is a system component that analyzes unstructured textual information to generate structured interpretations and market signals.
• agent orchestrationOrchestration is the coordination and scheduling of multiple tasks, processes, or services in a workflow.
• shadow modeShadow mode is a deployment approach in which a model processes live inputs without controlling decisions, allowing its outputs to be evaluated safely.
• return on investmentReturn on investment is the ratio of gain or loss from an investment to its cost, typically expressed as a percentage.
• parameter-efficient fine-tuningParameter-efficient fine-tuning adapts a pretrained model by training a small subset of its parameters or adding lightweight trainable components.
• joinA join operation combines records from separate data sources according to a relationship between their fields.
• precision at top n percentThe proportion of positive cases among the highest-ranked n percent of cases.
• sentinel valueA designated data value used to represent a special condition, such as missing or inapplicable information, rather than a valid measurement.
• data ingestionData ingestion is the process of collecting and importing data from source systems into a system for storage or processing.
• recommender systemA software system that generates recommendations from data or predictive models.
• graph convolutionA neural-network layer that transforms node features and combines information from neighboring nodes in a graph.
• AI engineeringAI engineering is the discipline of designing, building, deploying, and operating artificial-intelligence systems in production settings.
• ParquetParquet is a column-oriented file format for storing structured data with efficient compression and schema metadata.
• importance scoreFeature importance is a measure of the contribution or influence of an input feature on a model’s predictions.
• generative artificial intelligenceGenerative AI is artificial intelligence that produces new content such as text, images, audio, code, or other data.
• ScorecardMonitoringScorecardMonitoring is an OptBinning class for evaluating the stability of scorecard models across populations.
• credit risk modelA statistical or machine-learning model that estimates a borrower’s credit risk from borrower attributes and behavior.
• semantic matchingSemantic matching is a retrieval technique that compares representations of content according to meaning rather than exact wording.
• imputationThe replacement of missing data with estimated or designated values.
• reason codeA reason code is an explanation assigned to a credit decision, such as a code stating why an application was rejected.
• anomalyAn outlier is a data value that is unusually distant from other observations in a dataset or distribution.
• data qualityData quality is the degree to which data is accurate, complete, consistent, timely, and fit for its intended purpose.
• datamartA subject-focused repository of organized data for analysis and decision-making.
• auditabilityAuditability is the property of a system or process that allows its data, decisions, actions, and controls to be examined and verified.
• cross-asset relationshipA cross-asset interconnection is a relationship between the behavior or prices of different asset classes or financial instruments.
• signal-to-noise ratioA measure comparing the strength of informative signal to the amount of background noise in data or observations.
• policy tonePolicy Tone is a score representing the expressed or inferred stance and sentiment of a policy-related news source or statement.
• supply riskSupply Risk is a score representing the degree of disruption or threat to the availability of goods, inputs, or production capacity described in relevant information.
• credit scoringThe process of estimating a borrower’s creditworthiness to support lending decisions.
• model selectionModel selection is the process of choosing among candidate predictive models according to performance and other requirements.
• hyperparameter tuningHyperparameter tuning is the process of selecting configuration values for a model’s learning procedure to improve its performance.
• retentionRetention is the proportion or number of users who continue using a product or service over a specified period.
• Kaplan–Meier estimatorA nonparametric estimator of a survival function that accounts for right-censored observations.
• right censoringAn observation is right-censored when the event of interest has not occurred by the end of the observation period, leaving its later time unknown.
• survival probabilityThe probability that an entity has not yet experienced the event of interest by a specified time.
• false positiveA false positive is an alert or classification that incorrectly identifies legitimate activity as suspicious or problematic.
• points-to-double-odds scaleA score scale in which a specified increase in points corresponds to a doubling of the odds.
• machine learningA computational approach in which systems learn patterns from data to make predictions or decisions.
• scorecardA credit-scoring model that assigns point contributions to applicant characteristics to produce a credit score.
• explainability frameworkAn explainability framework is a set of methods and controls for making a model’s outputs and decision process understandable.
• feature storeA centralized repository for storing, managing, and serving validated features for machine-learning workflows.
• overfittingOverfitting occurs when a model captures patterns specific to its training data that do not generalize reliably to new data.
• graph densityGraph density is the proportion of possible edges that are present among a graph's nodes.
• cross-validationCross-validation is a model evaluation technique that partitions data into subsets and repeatedly trains and evaluates a model on different subsets.
• Apache AirflowApache Airflow is an open-source platform for authoring, scheduling, and monitoring workflows programmatically.
• adjacency matrixA matrix whose entries represent whether or how pairs of graph vertices are connected.
• cost efficiencyThe degree to which the value or benefit of a data source justifies its total acquisition and operating costs.
• reporting lagThe delay between the period covered by a company's financial report and the time that report becomes available.
• survivorship biasBias in analysis caused by considering only entities or data that survived to be observed, while excluding those that did not.
• timelinessThe degree to which data is delivered soon enough to support an intended decision or action.
• look-ahead biasThe use of future information that would not have been available at the time of a simulated decision or prediction.
• charge-offThe accounting recognition of a debt as a loss when it is considered unlikely to be collected.
• debt restructuringThe modification of a borrower's debt terms to address repayment difficulty or improve repayment prospects.
• fine-tuningFine-tuning adapts a pretrained model by continuing training on task-specific or domain-specific data.
• false negativeA false negative is a negative classification that incorrectly labels an actual positive case as negative.
• average revenue per userAverage revenue per user is the mean revenue generated per user over a specified period.
• directed acyclic graphA directed acyclic graph is a directed graph with no directed cycles, often used to represent task dependencies.
• promptA prompt is an instruction or input supplied to a generative model to guide its output.
• cachingA cache is a storage layer that retains previously computed or retrieved results for faster reuse.
• application fraudApplication fraud is fraud committed by providing deceptive information when applying for a financial product or service.
• identity theftIdentity fraud is fraud involving the deceptive use or impersonation of a person's identity.
• internal fraudInternal fraud is fraud committed by people within an organization, such as employees misappropriating funds or stealing customer data.
• market manipulation fraudFraudulent activity intended to deceive market participants by improperly influencing financial-market prices or trading.
• unstructured dataInformation that does not conform to a predefined schema for direct machine-readable interpretation.
• evaluation datasetA portion of data kept separate from model training for evaluation.
• Evidently AIEvidently is an open-source tool for evaluating and monitoring machine-learning data and models.
• labelThe outcome value associated with an example that a predictive model is trained to estimate.
• batchingThe execution of data-processing work on accumulated records in groups rather than as each record arrives.
• risk tierA category representing a level of assessed risk, such as low, medium, or high.
• watchlistA watchlist is a collection of people or entities flagged for screening against specified risk or compliance concerns.
• logitThe logarithm of the ratio of an event’s probability to the probability of its complement.
• privacyThe protection of personal information through appropriate controls on its collection, use, sharing, retention, and access.
• data cleaningThe process of detecting and correcting or removing errors, inconsistencies, and irrelevant data to improve data quality.
• f-scoreThe F1 score is the harmonic mean of precision and recall for a classification model.
• transparencyTransparency is the quality of making a model's processes, inputs, or outputs sufficiently understandable to support scrutiny and accountability.
• LIMEA model-agnostic explanation method that approximates a model locally with an interpretable model around a prediction.
• SHAPA feature-attribution method that explains a model prediction by assigning each input feature a contribution based on Shapley values.
• device fingerprintThe technique of identifying or characterizing a device from attributes such as its software, hardware, and settings.
• receiver operating characteristicA curve that plots a classifier’s true positive rate against its false positive rate across decision thresholds.
• exchange-traded fundAn investment fund whose shares trade on an exchange and that typically tracks a basket of assets, an index, or another exposure.
• XGBoostXGBoost is an open-source software library for gradient-boosted decision trees.
• securityThe protection of systems, information, and operations against unauthorized access, disruption, or harm.
• random forestA random forest is an ensemble model that combines predictions from multiple decision trees trained with randomized data and feature selection.
• confusion matrixA table that summarizes classification results by comparing predicted classes with actual classes.
• complianceRegulatory compliance is adherence to applicable laws, regulations, supervisory requirements, and prescribed financial controls.
• artificial intelligenceArtificial intelligence is the capability of computational systems to perform tasks that typically require human intelligence.
• real-time risk decisionThe process of evaluating a transaction and determining whether to approve, decline, or challenge it while it is being initiated.
• clusteringClustering is the grouping of observations according to similarities in their measured attributes.
• real-time processingNear-real-time processing handles incoming data with short delays rather than waiting for a scheduled batch.
• model architectureThe structural design of a predictive model, including its components and how they are organized.
• synthetic dataThe creation of artificial data that reproduces selected properties of real data without directly using the original records.
• fairnessThe principle that decisions and outcomes should treat affected people or groups equitably and avoid unjustified discrimination.
• NetworkXNetworkX is a Python library for creating, manipulating, and studying complex networks.
• foundational modelA foundation model is a large, broadly trained model that can be adapted or guided for many downstream tasks.
• metadataMetadata is descriptive information about data that identifies its attributes, context, structure, provenance, or meaning.
• transaction costTrading expenses, including fees, slippage, and spreads, that reduce returns.
• domain knowledgeSubject-matter understanding that guides the interpretation and transformation of data into meaningful variables.
• provenanceInformation describing the origin, ownership, and changes to data as it moves through systems.
• routingModel routing is the process of directing requests to different models according to their characteristics or requirements.

These guides are generated from the text of Manning books by Manning's knowledge graph and link to the sections they draw on. The full books are available on liveBook and at manning.com.