CDSP (DSP-110) key terms

Restating a stakeholder's request as a specific decision or outcome to be improved, with scope, constraints, and success criteria made explicit before any data work begins.

The measurable conditions agreed in advance that determine whether an analytical deliverable has solved the stated business problem, expressed in business units rather than model metrics alone.

The structured process of interviewing and workshopping with business sponsors to surface explicit asks, hidden assumptions, decision deadlines, and the constraints under which results will be used.

Breaking a broad business question into smaller answerable analytical questions that can each be matched to data, methods, and owners, so the overall effort can be sequenced and scoped.

An approach in which candidate explanations or expected effects are stated up front and then tested against data, focusing effort and preventing unfocused exploration of every available variable.

A precise, agreed specification of how a business metric is calculated, including numerator, denominator, population, time window, and exclusions, so that all parties measure the same thing consistently.

Translating a conceptual business idea into a computable field or formula using available data, resolving ambiguities such as event timing, attribution windows, and treatment of missing or cancelled records.

The pre-intervention or current-state value of a metric, captured before changes are made, which provides the reference point against which later performance and estimated impact are compared.

The gap between when an analytical result could inform a choice and when the decision is actually made; long latency means models must forecast further ahead or be designed for refresh.

A reference document describing each field in a dataset: its meaning, type, allowed values, units, source system, ownership, and known caveats, used to prevent misinterpretation across teams.

The documented path a data element travels from its originating system through transformations and joins to its final reporting location, used for impact analysis, debugging, and audit.

Systematically examining a dataset's structure and contents, including value distributions, distinct counts, null rates, and patterns, to understand its characteristics before analysis or modeling.

The standard axes along which data fitness is judged, typically completeness, accuracy, consistency, timeliness, validity, and uniqueness, each of which can be measured and monitored separately.

Absent values in a dataset that may arise from non-collection, system failure, or deliberate non-response, and whose pattern (random or systematic) affects which handling method is defensible.

Filling in missing values using estimated substitutes such as a mean, a model-based prediction, or the last known value, trading reduced data loss for introduced assumptions and possible distortion.

Identifying observations that deviate markedly from the bulk of the data, using rules, distance measures, or distributional thresholds, before deciding whether they are errors, rare events, or signals.

Distortion that occurs when the analyzed records differ systematically from the population of interest, for example because of how they were sampled, filtered, or retained in a system.

The procedure used to choose which records or units enter an analysis, such as random, stratified, or convenience sampling, which determines how far results can be generalized.

Two common data integration patterns: extract-transform-load applies transformations before loading into the warehouse, while extract-load-transform loads raw data first and transforms it inside the warehouse.

Creating new input variables from raw data, through aggregation, ratios, time windows, or encoding, to make patterns easier for a model to learn and to encode domain knowledge explicitly.

Including in model inputs information that would not be available at prediction time or that encodes the target, producing unrealistically strong training results that collapse in production.

Partitioning data into a portion used to fit a model and a held-out portion used to estimate performance on unseen cases, preserving the ordering of time when the data are temporal.

A resampling procedure that repeatedly splits data into training and validation folds and averages the results, giving a more stable performance estimate than a single holdout split.

Fitting a model so closely to the idiosyncrasies of training data that it captures noise rather than generalizable structure, showing strong training performance but weak performance on new data.

Adding a penalty on model complexity to the objective function, discouraging extreme coefficients or overly flexible fits and thereby improving generalization to unseen data.

The tension between models too simple to capture real structure (high bias) and models so flexible they fit noise (high variance), which guides choices about model complexity and data volume.

Searching over settings that control a model's learning process, such as tree depth or penalty strength, evaluated on validation data rather than the final test set to avoid optimistic estimates.

The distinction between predicting a discrete category, such as churn or not, and predicting a continuous quantity, such as revenue or time to event, which determines the algorithm and metric family used.

The combined considerations used to choose among candidate approaches, including predictive performance, interpretability, cost, runtime, maintainability, and fit to the decision the model will support.

The degree to which a person can understand why a model produced a given prediction or which inputs drive its behavior, often a requirement when decisions affect customers or must be justified.

A ranking of how much each input variable contributes to a model's predictions, useful for explanation and sanity checking but sensitive to correlation among inputs and to the method used.

A table of predicted versus actual class labels showing true positives, false positives, true negatives, and false negatives, from which accuracy, precision, recall, and related metrics are derived.

Precision is the share of predicted positives that are truly positive; recall is the share of actual positives that were found. They trade off against each other as the decision threshold moves.

Choosing the probability cutoff that converts a model score into an action, based on the relative business cost of false positives versus false negatives rather than on a default value.

A situation where one outcome is far rarer than another, which makes raw accuracy misleading and requires metrics and sampling or weighting strategies suited to rare events.

An unmeasured or uncontrolled factor that influences both the supposed cause and the observed outcome, creating a spurious association that can mislead causal interpretation of results.

The distinction between variables moving together and one actually producing change in the other; observational associations alone cannot establish cause without design or strong assumptions.

A comparison set that does not receive the treatment or intervention, used to separate the effect of the change from background trends and seasonality when estimating impact.

Re-running an analysis under alternative assumptions, subsets, or specifications to see whether the conclusion holds, which indicates how sensitive the finding is to analyst choices.

The tendency to seek, interpret, and remember evidence that supports existing beliefs, which can distort hypothesis testing, metric selection, and the reading of ambiguous results.

Expressing results with intervals, error bars, or probability statements rather than single point estimates, so decision makers understand the range of plausible values.

A determination that an observed result is unlikely to have arisen by chance under a stated null hypothesis, which says nothing by itself about the size or practical importance of the effect.

The magnitude of a difference or relationship in the units that matter to the business, reported alongside significance so stakeholders can judge whether the finding is worth acting on.

Choosing an appropriate chart form and encoding to make a pattern, comparison, or distribution immediately readable, matched to the audience and to the decision the chart supports.

A short opening section that states the question, the key finding, the recommended action, and the confidence level, written so a busy decision maker can act without reading the full analysis.

The deliberate ordering of an analytical presentation, typically situation, complication, question, answer, and recommendation, so evidence builds toward a clear decision rather than a tour of methods.