[ML02] Business Analytics Series.
Updated Jul 20, 2026

Tags: machine learning data science business analytics

Business Analytics Cheatsheet Series

Mục đích: Ghi chú cá nhân + chia sẻ kiến thức về nền tảng phân tích và xử lý dữ liệu của Data science trong các bài toán kinh doanh.

Định dạng: Intuition → Công thức → Code Python thực tế.

Đối tượng: Data science/engineer, BA.

Topics: business-analytics crisp-dm glossary kpi series-index

How to use this series

Each file is a lookup table, not a tutorial. Three columns throughout:

Column What it holds
Term The technique, method, theorem, or definition.
Type Where it sits in the data science toolkit — Framework, Process, Metric, Test, Algorithm, Technique, Diagnostic, Artifact, Concept, Anti-pattern, Governance.
Definition / Usage The concept, formula, role, and do/don’t. Optimized for recall, not first-time learning.

Series index

# Bài viết Covers
00 Business Analytics Cheatsheet Series This page: reading guide + shared glossary.
01 Business Analytics Foundations & Types of Analytics What BA is, BA vs DA vs DS, evolution, 6-step BA process, four analytics types.
02 CRISP-DM, KPI Tree & Problem Framing CRISP-DM six phases, KPI Tree, problem framing, A/B basics, templates.
03 Data Quality & Feature Engineering Quality dimensions, missingness, leakage, encoding, scaling, feature selection, pipelines.
04 EDA & Visualization for Decision-Making Univariate/bivariate/multivariate EDA, correlation vs causation, chart choice, chartjunk.
05 Linear Regression (Simple & Multiple) SLR/MLR, least squares, r², LINE assumptions, residual analysis, inference, MRA.
06 Classification Evaluation: ROC/PR, Thresholding, Cost-Sensitive, Calibration Confusion matrix, ROC/PR, thresholding, cost-sensitive decisions, calibration.
07 Segmentation & Dimensionality Reduction k-means, GMM/EM, PCA, RFM, hybrid segmentation pipeline.

Reading order


Shared glossary

Terms used across multiple decks. Deck-specific terms live in their own file.

Business & product metrics

Term Type Definition / Usage
AOV — Average Order Value Metric (output) Σ(price × qty) / #orders. Revenue lever independent of traffic. Levers: bundles, upsell, free-shipping threshold. Watch: returns and order edits change the denominator.
CR — Conversion Rate Metric (output) #conversions / #sessions (or users). Decompose into stage-wise conditional probabilities before optimizing. Always report the window and the unit (session vs user).
CAC — Customer Acquisition Cost Metric (input) Total acquisition spend / new customers. Compare against LTV, not against revenue.
LTV — Lifetime Value Metric (output) Expected discounted margin over a customer’s lifetime. Sensitive to churn assumptions — state the horizon.
MRR / ARR Metric (output) Monthly/Annual Recurring Revenue. MRR = #Customers × ARPA × (1 − Churn).
ARPA Metric (output) Average Revenue Per Account. Pricing/packaging lever in the MRR tree.
Churn Metric (output) Share of customers lost in a period. Define the event precisely (contract end vs inactivity threshold) — the definition is the model’s label.
Retention Metric (output) 1 − churn for the cohort. Report as a curve by cohort age, not a single number.
Cohort Concept Group sharing a start event (signup month, first purchase). Cohorting isolates mix-shift from real behavior change.
North Star Metric (framework) Single metric that best proxies delivered customer value; the root of the KPI Tree.
Vanity metric Anti-pattern Metric that moves without informing any decision (raw pageviews, cumulative signups). Drop unless tied to a lever.

Metric design

Term Type Definition / Usage
Leading KPI Metric class Moves before the outcome; usable for steering (add-to-cart rate, activation rate). Noisier but actionable.
Lagging KPI Metric class Confirms the outcome after the fact (revenue, churn). Authoritative but too slow to steer. Pair one leading with one lagging.
Input vs Output metric Metric class Input = what the team controls (emails sent, latency). Output = what the business feels (revenue). Set targets on outputs, run sprints on inputs.
Health vs Performance metric Metric class Health = system doesn’t break (latency, payment error rate). Performance = system produces value. Ship neither alone.
Guardrail Metric class Hard constraint that a winning variant must not violate (e.g., “overall CR must not drop >1pp”). Defined before the experiment, not after.
Funnel Framework Ordered stage sequence (session → view → add_to_cart → checkout → purchase). Multiply stage rates to get end-to-end CR; the smallest stage rate is rarely the biggest absolute loss — compute absolute drop-off.
Uplift Metric Incremental effect of treatment vs control. Not the same as post-treatment level; requires a counterfactual.
Drift Diagnostic Distribution shift over time in inputs (data drift) or in the input→output relation (concept drift). Triggers retraining/recalibration.

Delivery & MLOps

Term Type Definition / Usage
Canary release Deployment pattern Route a small traffic slice to the new version, monitor, then expand. Limits blast radius; needs an automatic rollback trigger.
Blue-Green deployment Deployment pattern Two full environments; switch traffic atomically. Fast rollback, double the infrastructure cost.
Feature Store Artifact / infrastructure Central registry of computed features with lineage and versioning. Guarantees train/serve consistency — the top cause of silent production degradation.
Model Card Artifact Documentation of intended use, training data, metrics by segment, limitations, fairness checks. Required for handover and audit.
Runbook Artifact Operational procedure: alerts, thresholds, on-call actions, rollback steps.
RACI Governance Responsible / Accountable / Consulted / Informed. Attach to every KPI and every deliverable; unassigned KPIs do not move.
Data lineage Governance Traceable path from source system to metric. Prerequisite for trusting any number in a board deck.
Data contract Governance Agreement between producer and consumer on schema, semantics, and SLA. Versioned; breaking changes require backfill plans.

Evidence hierarchy

Term Type Definition / Usage
Observation Evidence level What the data shows. Safe to state. “Mobile CR is 1.2pp below desktop.”
Hypothesis Evidence level Proposed explanation, falsifiable. “Mobile CR is lower because the checkout form requires re-entry.” Must name the test that would kill it.
Causation Evidence level Established only by randomization or a credible quasi-experiment (DID, IV, RDD). Never claim it from correlation alone.
A/B test Method Randomized controlled experiment. Fix randomization unit, primary metric, guardrails, and duration before launch. Do not peek.
Holdout Method Untreated group kept aside long-term to measure cumulative program effect, not a single change.
Quasi-experiment Method DID (difference-in-differences), IV (instrumental variables), RDD (regression discontinuity). Use when randomization is impossible; assumptions are strong and must be stated.

Definition of Done for this series

Next step: start with CRISP-DM, KPI Tree & Problem Framing — it is the spine the other files hang from.


🔗 Full series

Bài 1: Business Analytics Foundations & Types of Analytics
Bài 2: CRISP-DM, KPI Tree & Problem Framing
Bài 3: Data Quality & Feature Engineering
Bài 4: EDA & Visualization for Decision-Making
Bài 5: Linear Regression (Simple & Multiple)
Bài 6: Classification Evaluation: ROC/PR, Thresholding, Cost-Sensitive, Calibration
Bài 7: Segmentation & Dimensionality Reduction


Nếu bài này hữu ích, hãy chia sẻ cho người đang học Data Science / Business Analytics. Mọi góp ý về lỗi kỹ thuật rất được hoan nghênh.