Internal and External Data Sources

Internal and External Data Sources

Quick Answer: After you know what data the research question needs, Domain 2 asks you to identify candidate sources—internal systems and stores plus external or public datasets—and evaluate them for fitness, trust, cost, latency, and rights. Platform names matter less than whether a source can supply the right elements at the right grain for the decision.

Source identification is still planning work. You are building options for acquisition, not yet claiming that every source is clean enough for final analysis (quality depth follows in later Domain 2 sections). On the CBDA exam, strong answers match source type to analytical use case and call out governance implications early.

A practical map of source types

Operational systems (systems of record / engagement)

These are day-to-day applications: CRM, ERP, billing, claims, EHR modules, LMS, POS, ticketing, HRIS. They generate transactions and states the business runs on.

CBDA use cases: current process truth, event-level history, status flags used in operations.
Watch-outs: schemas optimize transactions, not analytics; definitions may be local; extracts can stress production; historical depth may be limited by archival rules.

Data warehouses

Warehouses store integrated, often structured data optimized for analytical query—commonly dimensional models or governed enterprise tables.

CBDA use cases: cross-functional metrics, historical snapshots, consistent reporting grain for descriptive and diagnostic work.
Watch-outs: latency (batch loads), incomplete subject areas, “official” metrics that still disagree with operational reality if ETL transforms are opaque.

Data marts

Marts are subject- or department-oriented analytical stores (finance mart, marketing mart) derived from warehouse or operational feeds.

CBDA use cases: faster access for a domain team; pre-joined conformed dimensions for common questions.
Watch-outs: local marts can diverge; marketing “customer” may not match finance “customer.”

Data lakes (and lakehouses)

Lakes hold large volumes of raw or lightly processed data in varied formats; lakehouses add warehouse-like table management on lake storage.

CBDA use cases: exploratory access to multi-format history, machine learning feature raw zones, cheap retention of detailed events.
Watch-outs: “dumping ground” risk without catalog and quality; schema-on-read shifts burden to the analyst; access control and PII sprawl.

Data vaults (awareness level)

Data vault modeling emphasizes hubs, links, and satellites for agile, auditable integration of historical sources.

CBDA use cases: enterprise integration with strong historization and lineage-friendly design.
Watch-outs: business users rarely query vault structures directly; marts or warehouse presentation layers usually sit on top. Exam awareness: vault ≠ automatic business definitions.

Third-party and public data

External commercial feeds (credit, firmographics, weather, mobility, market prices) and public/open data (census, macro indicators, government releases, academic sets).

CBDA use cases: enrichment, benchmarking, market context, features not generated inside the firm.
Watch-outs: licensing, redistribution limits, opaque collection methods, refresh surprises, join-key weakness, bias relative to your population.

Structured vs semi-structured vs unstructured (awareness)

CBDA does not require you to be a deep learning engineer. It does require you to recognize that form changes sourcing and preparation cost. Combining structured transactions with unstructured feedback is common; plan time and skills for both. On scenarios, choosing only structured data when the research question is about why customers complain in their own words is a fitness miss.

Internal vs external: trust, cost, latency, rights

When internal is preferred

• The decision hinges on how your process actually runs.
• You need lineage and remediation with people you can call.
• External purchase cannot clear legal review inside the decision window.
• Metrics must reconcile to financial or operational systems of record.

When external is justified

• Required attributes do not exist internally (market share, competitor price index, weather at store location).
• Internal coverage is structurally incomplete (new market entry with no history).
• Benchmarking against industry baselines is part of the research question.
• Primary collection would be slower or costlier than a governed purchase.

Exam pattern: Internal incomplete + external complementary → evaluate both, integrate deliberately, govern rights—not “external only because it is trendy,” and not “internal only because external feels scary” when the gap is real.

Reference data and master data (awareness)

Reference data is relatively stable codes and lookups: country codes, product category lists, reason codes, status enumerations, currency codes. Bad reference data breaks filters and group-bys (“CA” meaning California vs Canada; retired cancel codes still appearing).

Master data is the authoritative business entities the enterprise tries to keep consistent: customer, product, supplier, location, chart of accounts. Master data management (MDM) programs create golden records and survivorship rules.

Why CBDA candidates care during sourcing:

• Joining CRM to billing fails when customer master duplicates or mismatched keys exist.
• Product hierarchy in merchandising disagrees with finance SKU rolls—marts disagree.
• Reference lists differ by region; global analysis silently drops rows.
• “One customer” research questions assume master resolution you may not have—log it as a gap.

You are not expected to implement MDM in Domain 2, but you must notice when analysis depends on master/reference consistency and plan keys, survivorship assumptions, and limitations accordingly.

Source-type table with CBDA use cases

Scenario patterns

Pattern A — Warehouse first bias: Sponsor says “only use the warehouse.” Inventory shows claims denials detail never landed. Best move: document gap; seek operational claims extract or lake raw zone; do not pretend warehouse completeness.

Pattern B — External shortcut: Team wants to buy a lookalike audience file instead of defining internal conversion events. Best move: confirm whether the research question is about your funnel (internal first) or market universe (external may fit); check rights before purchase.

Pattern C — Unstructured gold: NPS scores are flat but ticket text explodes with a new defect theme. Best move: plan semi/unstructured sourcing (tickets, notes) with privacy review—not only structured score tables.

Pattern D — Master data pain: Two “customer counts” differ by 12%. Best move: pause cross-system claims; involve stewards; define survivorship for this analysis; state limitation if unresolved.

Pattern E — Mart convenience trap: Marketing mart makes charts easy but excludes returns. Best move: validate mart business rules against the research question; supplement from warehouse/ops if returns matter.

Selecting sources: a short decision path

1. Start from the question-to-data matrix (prior section).
2. List internal candidates that hold must-have elements.
3. Rate fit (coverage, grain, window, definition match).
4. If gaps remain, evaluate external or primary options with rights and cost.
5. Prefer sources with owners, dictionaries, and access paths inside the decision window.
6. Record multi-source join keys and master/reference dependencies.
7. Carry forward quality unknowns—do not equate “source identified” with “source trusted.”

CBDA exam habits for source types

• Prefer fitness to the research question over newest architecture buzzwords.
• Prefer internal + external combination when each closes a real gap.
• Prefer answers that mention rights, licensing, and privacy for third-party data.
• Prefer recognition of structured vs unstructured preparation cost.
• Prefer flags for master/reference data issues when joins or golden IDs are assumed.
• Reject “analyze all data everywhere” and “never use external data” as absolute rules.