Core Glossary & Conversion Guide

Welcome — how to use this guide

This glossary and conversion guide collects concise, practical definitions and short examples for the core analytics and data terms used across this domain. Use it to onboard teammates, reduce miscommunication, check a formula quickly, or pick an appropriate statistical test. Each entry explains why the term matters, gives a short example, and includes cross-link suggestions to related learning decks or tools.

Scope & conventions

Entries are intentionally pragmatic: a plain definition, why it matters in decision making, a short example, and pointers to related resources. This is a living reference — teams should adapt definitions to local master-data standards before copying them into enterprise systems.

Core glossary entries

KPI vs metric

Definition: A metric is any measurable value (e.g., page views, units produced, defect count). A KPI (Key Performance Indicator) is a metric selected as strategically important and tied to a business objective.

Why it matters: Treating every metric as a KPI dilutes focus. KPIs are the measures you monitor and act on regularly.

Example: Daily active users is a metric; conversion rate may be a KPI when your goal is to grow revenue.

Data product

Definition: A data product is a reusable deliverable built from data (reports, models, APIs, datasets) designed for specific consumers and SLAs.

Why it matters: Thinking in products helps with ownership, quality, versioning, and user experience.

Example: A curated customer 360 API served to sales and support teams is a data product.

Event vs session

Definition: An event is a single recorded occurrence (button click, transaction). A session groups related events by a user during a continuous interaction window.

Example: A purchase is an event; the shopping visit including browsing, cart additions, and payment is a session.

Lineage

Definition: Lineage traces how a dataset or value was produced — sources, transformations, and destinations.

Why it matters: Lineage supports trust, debugging, impact analysis, and compliance.

Example: Knowing which ETL jobs and joins produced the weekly sales table helps diagnose sudden changes.

Model validation

Definition: The process of assessing whether a predictive model meets performance, stability, and fairness requirements for its intended use.

Why it matters: Validation ensures models generalize and remain appropriate when deployed.

Example: Splitting holdout data, calculating AUC and calibration plots, and running bias checks before deployment.

Drift

Definition: A change in the statistical properties of inputs, model predictions, or labels over time (covariate drift, concept drift, label drift).

Why it matters: Drift can reduce model performance and lead to poor decisions if not detected and addressed.

Example: A sudden change in customer behavior after a pricing update causing prediction errors.

Causal inference

Definition: Methods and reasoning used to estimate the effect of interventions (cause → effect), distinguishing causation from correlation.

Why it matters: Policy and investment decisions usually require causal evidence, not just correlation.

Example: Using randomized experiments or difference-in-differences to estimate the effect of a marketing campaign on sales.

Forecast horizon

Definition: The time window into the future for which a forecast is produced (short-, medium-, or long-term).

Why it matters: Forecast horizon drives model choice, feature selection, and evaluation metrics.

Example: Weekly demand forecast for the next 4 weeks has a different horizon and error tolerance than quarterly strategic forecasts.

Leading vs lagging indicators

Definition: Leading indicators tend to change before outcomes (predictive signs); lagging indicators confirm results after they occur.

Why it matters: Combine them: use leading indicators to act early and lagging indicators to verify outcomes.

Example: Website sign-ups (leading) vs. paid conversions (lagging).

Measurement frameworks

Definition: Structured approaches for choosing what to measure, how to instrument it, and how to interpret it (OKRs, AARRR, input–output–outcome models, etc.).

Why it matters: Frameworks align measurement with strategy and avoid vanity metrics.

Example: Using an input–output–outcome map to connect marketing spend to user acquisition and revenue.

Conversion & quick formulas

• Rate conversions: rate per 1000 = (count / population) × 1,000.
• Percent change: ((new − old) / old) × 100.
• Mean (average): sum(values) / n.
• Median: middle value when sorted (robust to outliers).
• Variance (sample): sum((x − mean)^2) / (n − 1).
• Standard deviation: sqrt(variance).
• Conversion between time units: 1 day = 24 hours = 86,400 seconds.
• Confidence interval (approx, large n): mean ± z*(sd / sqrt(n)). Use t-distribution for small samples.

Statistical-test quick guide

Choose a test by the question, data type, and assumptions. Below are common choices and the key assumptions to check.

• T-test (two-sample): Compares means of two groups. Assumptions: approximately normal distributions or large n; similar variances (or use Welch's t-test).
• Paired t-test: Compares means for paired observations (before/after). Assumptions: differences approximately normal.
• Chi-square test (independence): Tests association between categorical variables. Assumptions: expected cell counts not too small.
• ANOVA: Compares means across 3+ groups. Assumptions: normality and homogeneity of variances (alternatives: Kruskal–Wallis).
• Mann–Whitney (Wilcoxon rank-sum): Nonparametric alternative to two-sample t-test when distributions are skewed.
• Correlation (Pearson vs Spearman): Pearson for linear relationships with interval data and normality; Spearman for monotonic relationships or ordinal data.
• Regression: Use linear regression for continuous outcomes, logistic regression for binary outcomes; always check residuals, multicollinearity, and model fit.
• Multiple testing: Adjust p-values or control false discovery when running many tests (Bonferroni, Benjamini–Hochberg).

Checklist before applying a test

1. Define the hypothesis and estimand clearly.
2. Check data type and distribution.
3. Confirm independence or account for pairing/clustering.
4. Assess sample size and power; small samples may need nonparametric methods.
5. Report effect sizes and confidence intervals, not just p-values.

Practical examples & cross-links

Each glossary entry should link to relevant decks, datasets, and tools for deeper learning. Example cross-links: "See the Forecasting Primer deck for model horizons and metrics" or "See the A/B Testing deck for causal inference methods." Replace these placeholders with your team's local deck IDs or URLs when tailoring this domain.

How to adopt & adapt this resource

Recommended steps for teams:

• Review definitions and adapt wording to reflect local master-data and naming conventions.
• Add cross-links to your internal decks, data product pages, and lineage records.
• Create an owned glossary copy and version it so teams can evolve definitions without breaking others.

Notes & cautions

This guide is a practical reference, not a substitute for domain-specific standards, professional statistical consulting, or legally/compliantly certified procedures. Avoid using brief cheat sheets as the sole justification for complex analyses. Always validate assumptions, inspect data quality, and consult subject-matter experts for high-risk decisions.

Last updated: This is a living document. Add local examples, decks, and data product links to make it actionable for your organization.