Financial Data, Measurement and Preparation

Introduction

Financial analytics begins long before a model is estimated. Analysts must determine what an observation represents, when it became available, how it was measured, and whether it can be compared with other observations. An inaccurate timestamp, duplicated customer, unadjusted share price, or improperly joined accounting record can invalidate an otherwise sophisticated analysis.

Financial data are especially demanding because they are heterogeneous, time-dependent and regulated. A bank may combine account transactions, loan contracts, credit-bureau reports and macroeconomic indicators. An insurer may integrate policies, claims, telematics and weather records. Investment analysts combine prices, corporate disclosures, analyst forecasts and news. These sources differ in frequency, granularity, units, reliability and permissible use.

Data preparation therefore involves more than correcting errors. It is the controlled process through which raw observations become analytically meaningful evidence. This chapter develops the concepts and practical methods required to collect, measure, clean, transform, integrate, sample and explore financial data while controlling quality, bias and regulatory risk.

Learning Objectives

After studying this chapter, students should be able to:

1. Distinguish market, accounting, banking, insurance, macroeconomic and alternative data.
2. Explain measurement scales, units, timing conventions and levels of granularity.
3. Design a controlled financial-data pipeline from source to decision output.
4. Diagnose and treat missing values, duplicates, outliers and inconsistent records.
5. Transform financial variables for comparison and statistical analysis.
6. Integrate datasets without creating invalid matches or look-ahead bias.
7. Apply probability sampling and calculate a design-consistent stratified estimate.
8. Conduct exploratory data analysis and identify data-quality problems.
9. Evaluate sampling, measurement, survivorship and algorithmic bias.
10. apply governance, ethical and regulatory principles throughout the data lifecycle.

1. The Financial Data Landscape

A dataset is a structured collection of observations about entities, events or periods. Its unit of observation may be a customer, transaction, security, policy, claim, firm or country-month. Correctly identifying that unit is essential. A file containing one row per transaction cannot be joined directly to a file containing one row per customer without first specifying an aggregation rule.

1.1 Major categories of financial data

Market data describe trading activity and market valuation. A price alone is not always an economically complete measure. Equity prices may require adjustment for dividends, share splits and rights issues. Bond analysis requires clarity about clean price, accrued interest, maturity and yield convention.

Accounting data arise from financial statements and related disclosures. They are prepared for reporting periods rather than continuous time. Reported profit for a financial year must not be treated as known before its publication date. Restatements can also create multiple versions of the same historical figure.

Banking data include customer characteristics, applications, account balances, credit limits, repayments and delinquency events. They are frequently hierarchical: one customer may hold several accounts, and each account may contain many transactions.

Insurance data connect policies, insured exposures, premiums, claims and settlements. A claim reported today may relate to an earlier accident and may remain unsettled for years. Analysts must distinguish occurrence date, reporting date, payment date and valuation date.

Macroeconomic data provide the economic environment in which institutions operate. Many official series are revised after initial release. A forecasting exercise should use the vintage actually available at the historical forecast date, not a subsequently revised value.

Alternative data are generated outside conventional reporting systems. Examples include payment-terminal activity, mobile-app usage, text sentiment and vehicle telematics. Such information can improve timeliness but may have uncertain provenance, weak population coverage or serious privacy implications.

1.2 Structured, semi-structured and unstructured data

Structured data occupy predefined rows and columns. Semi-structured data, such as JSON messages and regulatory filings marked with XBRL tags, contain labels but may vary in structure. Unstructured data include reports, emails, images and audio. Natural-language processing can convert text into variables, but the resulting sentiment or topic score is a model-generated measurement rather than a directly observed fact.

1.3 Internal and external data

Internal sources include core banking systems, policy administration platforms, customer relationship systems and general ledgers. External sources include exchanges, regulators, statistical agencies, credit bureaus and commercial vendors. External acquisition does not transfer responsibility for quality: the receiving institution must still validate licensing, provenance, definitions and delivery controls.

2. Measurement: What Does a Financial Number Mean?

Measurement assigns values to attributes according to explicit rules. A number without its definition, unit, entity and reference time is incomplete.

2.1 Measurement scales

A credit rating is ordinal: grade A is better than grade B, but the numerical distance between grades is not necessarily constant. Treating such categories as equally spaced numerical values requires justification.

2.2 Stocks, flows and exposure

A stock is measured at a point in time, such as a loan balance on 31 December. A flow is measured over an interval, such as annual interest income. Combining a flow with a stock requires a suitable denominator. For example, return on average assets generally uses average assets because income accumulates throughout the period.

Insurance claim frequency must also be related to exposure. Ten claims among 100 one-year policies differ from ten claims among 100 one-month policies. If exposure is measured in policy-years, a basic frequency measure is:

The numerator and exposure period must refer to compatible populations and time intervals.

2.3 Event time, effective time and system time

Financial records commonly contain several timestamps:

• Event time: when the economic event occurred.
• Effective time: when a value legally or operationally became applicable.
• Publication time: when information became publicly available.
• Ingestion time: when the analytical system received it.
• Processing time: when it was transformed.

These distinctions prevent look-ahead bias, which occurs when an analysis uses information that would not have been available at the decision date. Point-in-time datasets preserve what was known at each historical date.

2.4 Nominal and real values

Nominal monetary values include changes in the price level. If R_n is a nominal growth rate and π is inflation over the same period, the exact real growth rate is:

Here, R_r is the real growth rate, and all rates are decimal rates over the same interval. The approximation R_r ≈ R_n − π is suitable only when both rates are relatively small.

Example. If deposits grow by 8% while inflation is 5%, real growth is 1.08 ÷ 1.05 − 1 = 0.02857, or approximately 2.86%, not exactly 3%.

2.5 Returns and corporate actions

For an asset with initial price P_t-1, final price P_t and cash distribution D_t, the simple holding-period return is:

Prices and distributions must use the same currency and refer to the same holding period. If prices are already adjusted for distributions, adding dividends again would double count them.

3. The Governed Data Pipeline

A financial-data pipeline converts source records into controlled analytical and decision outputs. Governance is not a final pipeline stage. It is a cross-cutting system of ownership, metadata, access control, validation, security, retention and auditability that applies to every stage.

Figure: Governance spans the complete financial-data pipeline, which extends from source systems to reports and decisions and includes feedback controls.

3.1 Source contracts and ingestion

A data contract states the expected schema, field definitions, units, permissible values, refresh schedule and responsibilities of provider and consumer. Ingestion controls should verify file identity, row counts, schema, timestamps, checksums and delivery completeness. Raw data should normally be retained unchanged in a restricted layer so that transformations can be reproduced.

3.2 Metadata and lineage

Metadata describe data. Business metadata define meaning; technical metadata define formats and locations; operational metadata record processing events. Data lineage documents where a variable originated and which transformations produced it. A model input such as “90-day delinquency rate” should be traceable to the relevant accounts, dates, status codes and aggregation rule.

4. Data Quality Assessment

Data quality is fitness for a stated purpose. A dataset can be accurate enough for aggregate reporting yet inadequate for individual credit decisions.

For a field containing n records, of which m are missing, the missing rate is:

A low overall rate can conceal concentration in a protected or high-risk group. Quality metrics should therefore be examined by source, period, product and relevant customer segment.

4.1 Rule-based profiling

Useful profiling tests include:

• uniqueness of keys;
• minimum, maximum and quantile checks;
• allowed-value and data-type validation;
• date-order checks, such as settlement date not preceding claim date;
• reconciliation identities;
• abrupt distribution changes between periods;
• referential-integrity checks between related tables.

Accounting provides powerful validation identities. Subject to the reporting framework and consolidation scope:

A failed identity may indicate a missing record, sign error, inconsistent reporting scope or inappropriate currency conversion. It should trigger investigation rather than automatic alteration.

5. Cleaning Financial Data

Cleaning identifies and resolves defects while preserving an audit trail. Original values should not be overwritten silently; corrections should be implemented through reproducible code or controlled adjustment records.

5.1 Missing data

Missingness may arise because a value was not collected, was inapplicable, failed validation or arrived late. These causes are analytically different. A missing income value should not automatically be encoded as zero.

Common treatments include:

1. Deletion: reasonable when missingness is rare and unlikely to distort the sample.
2. Simple imputation: replacement by a median, mode or specified constant; easy but understates uncertainty.
3. Conditional imputation: estimates values using related variables or groups.
4. Multiple imputation: creates several plausible completed datasets and combines results, reflecting imputation uncertainty under stated assumptions.
5. Missingness indicators: represent whether a value was absent, where the absence itself may be informative.

The conventional missing-data mechanisms are MCAR—missing completely at random; MAR—missing at random conditional on observed information; and MNAR—missingness related to the unobserved value even after conditioning. These are assumptions about the data-generating process, not labels established merely by inspecting a file.

5.2 Duplicates

Exact duplicate rows are not the only duplication problem. The same customer may appear under different identifiers or spellings. Entity resolution uses identifiers, names, addresses and other attributes to infer whether records refer to the same entity. False merges can be as harmful as missed matches, especially where decisions concern individuals.

Deduplication requires a declared key and retention rule. For example, retaining the latest customer record is valid only if timestamps are reliable and older versions are not needed for point-in-time analysis.

5.3 Outliers

An outlier is an observation distant from much of the distribution. It may be an error, fraud signal, crisis event or legitimate large exposure. Automatic deletion can remove precisely the risk information that matters most.

For a univariate distribution, the interquartile range is:

A common screening rule flags values below Q₁ − 1.5IQR or above Q₃ + 1.5IQR. This is a diagnostic convention, not proof that a record is erroneous. Treatments include verification, correction, robust modelling, transformation, capping with justification, or separate analysis of extreme cases.

5.4 Inconsistent formats and units

Dates, currencies, decimal conventions and signs must be standardized. Currency conversion requires an exchange rate with an appropriate rate type and date. Stock values at a period end may use a closing rate, whereas flows over a period may use transaction-date or average rates, depending on the analytical and accounting purpose.

6. Transformation and Feature Construction

Transformations alter representation while retaining or enhancing analytical meaning.

6.1 Standardization

For observations x_i with sample mean \bar{x} and sample standard deviation s, a z-score is:

This produces a transformed sample with mean 0 and sample standard deviation 1, provided s > 0. Standardization changes scale but does not remove skewness or outliers. In predictive work, parameters must be estimated from the training data and then applied to validation and test data to avoid leakage.

6.2 Logarithmic transformation

For strictly positive x, ln(x) can reduce right skew and convert multiplicative relationships into additive ones. It is undefined for zero and negative values. Adding an arbitrary constant changes interpretation and must be justified.

Continuously compounded return is:

This requires positive prices and consistent price adjustments. Log returns are additive across consecutive periods, whereas simple returns compound multiplicatively.

6.3 Ratios and rates

Ratios improve comparability but may become unstable when the denominator is small or negative. A debt-to-income ratio also requires aligned periods and consistent definitions of debt and income. Analysts should retain numerator and denominator because identical ratios may represent very different exposure sizes.

6.4 Aggregation and temporal features

Transaction data may be aggregated into monthly inflows, balance volatility or missed-payment counts. Every feature needs:

• an entity key;
• an observation window;
• an ending or cutoff date;
• treatment of reversals and late transactions;
• a rule for partial periods.

A feature for a decision dated 30 June must not include a transaction posted on 2 July, even if the transaction later receives an effective date in June, unless the decision system genuinely had that information.

7. Data Integration

Integration combines sources into a coherent analytical dataset. Joins must reflect the business relationship and time structure.

7.1 Join types and granularity

An inner join retains matched records only and may silently exclude customers absent from one source. A left join retains all records from the primary table and indicates unmatched records. Many-to-many joins can multiply rows and distort totals.

Before joining, analysts should verify the cardinality:

• one customer to one demographic record;
• one customer to many accounts;
• one policy to many claims;
• one firm-period to one accounting statement version.

If customer-level data are joined to transactions, transactions should generally be aggregated first or the resulting repeated customer attributes must be intentionally handled.

7.2 Temporal and as-of joins

An as-of join matches each event with the most recent eligible record available no later than the event time. This is important for interest rates, credit ratings and macroeconomic releases. Eligibility should be based on publication or availability time when modelling historical decisions.

7.3 Reconciliation after integration

After a join, compare pre- and post-join row counts, unique keys, unmatched rates and financial totals. If a loan portfolio balance doubles after integration, a many-to-many join is a likely cause. Reconciliation converts integration from a coding task into a controlled accounting process.

8. Sampling Financial Populations

Sampling is used when full-population analysis is costly, slow or operationally impossible. A valid sample design defines the target population, sampling frame, selection method and inclusion probabilities.

8.1 Probability and non-probability sampling

In simple random sampling without replacement (SRSWOR), every sample of a specified size has the same probability of selection. Systematic sampling selects every kth item after a random start but can be affected by periodic ordering. Stratified sampling divides the population into internally defined strata and samples within each. Cluster sampling selects groups such as branches and then studies units within those groups.

Convenience and judgment samples are non-probability samples. They may support operational investigation but do not support design-based population inference unless strong additional assumptions are made.

8.2 Stratified estimation

Suppose stratum h contains N_h population units and n_h sampled units. Let N = ∑_h N_h, W_h = N_h ÷ N, and \bar{y}_h be the sample mean in stratum h. The stratified population-mean estimator is:

If independent SRSWOR is conducted within each stratum, an estimated design variance is:

where f_h = n_h ÷ N_h is the stratum sampling fraction and s_h² is the sample variance within stratum h. The factor 1-f_h is the finite-population correction. This formula does not apply unchanged to cluster, unequal-probability or non-probability designs.

8.3 Worked example: stratified default estimation

A lender has 8,000 retail loans and 2,000 small-business loans. It independently selects 200 loans by SRSWOR from each stratum. Sixteen sampled retail loans and 30 sampled small-business loans defaulted.

The stratum estimates are:

• Retail: \bar{y}₁ = 16 ÷ 200 = 0.08, with W₁ = 0.80.
• Small business: \bar{y}₂ = 30 ÷ 200 = 0.15, with W₂ = 0.20.

Therefore:

The estimated portfolio default proportion is 9.4%. The unweighted sample proportion is 46 ÷ 400 = 11.5%, which is inappropriate because the two strata were sampled at different rates.

For binary data, the unbiased within-stratum sample variance is s_h² = n_h\bar{y}_h(1-\bar{y}_h) ÷ (n_h-1). Thus:

The sampling fractions are f₁=200÷8000=0.025 and f₂=200÷2000=0.10. Hence:

The estimated standard error is:

Using a normal large-sample approximation, a 95% confidence interval is:

The approximate interval is 0.0628 to 0.1252, or 6.28% to 12.52%. This interval is design-consistent with independent stratified SRSWOR and includes stratum-specific variances and finite-population corrections. The normal approximation should be assessed cautiously when stratum event counts are very small; replication methods or design-aware alternatives may then be preferable.

9. Exploratory Data Analysis

Exploratory data analysis (EDA) examines structure before formal modelling. It should reveal both economic patterns and preparation defects.

9.1 Univariate analysis

For numerical variables, inspect counts, missingness, mean, median, standard deviation, quantiles and histograms. For categorical variables, inspect frequencies and rare levels. The mean is sensitive to extreme values; the median is more robust but does not replace examination of the full distribution.

9.2 Bivariate and multivariate analysis

Cross-tabulations can compare default rates by product or rating. Scatterplots can reveal nonlinearity and heteroscedasticity. Correlation measures linear association but does not establish causation. Highly correlated variables may represent duplicated information, common economic drivers or mechanically related ratios.

Time plots are indispensable. A sudden shift may reflect an economic shock, product redesign, system migration or altered definition. Segment-level analysis can expose patterns hidden in aggregate statistics.

9.3 Leakage and unstable relationships

EDA must respect the eventual prediction date. A variable such as “final recovery amount” may strongly predict default severity but is unavailable when credit is approved. Analysts should also test whether distributions and relationships change over time. Random train–test splits can conceal temporal instability; chronological validation is often more realistic for financial applications.

10. Bias, Ethics and Governance

10.1 Sources of bias

• Selection bias: included observations differ systematically from the target population.
• Survivorship bias: failed firms, closed funds or terminated accounts are omitted.
• Measurement bias: variables are measured differently across groups or periods.
• Historical bias: data reproduce past inequalities or institutional practices.
• Label bias: outcomes such as fraud or default are observed imperfectly.
• Publication bias: available information disproportionately reflects reported or successful outcomes.
• Automation bias: users place excessive confidence in system outputs.

A credit model trained only on previously approved applicants observes repayment outcomes for accepted borrowers but not rejected applicants. This reject-inference problem limits what the historical labels reveal about the full applicant population.

10.2 Privacy, security and lawful use

Financial and insurance records are highly sensitive. Institutions should apply purpose limitation, data minimization, role-based access, encryption, retention schedules and secure deletion. Pseudonymization reduces direct identifiability but does not guarantee anonymity, particularly when datasets can be linked.

Alternative data require special scrutiny. Legal availability does not automatically imply ethical appropriateness. Geolocation, social-network or behavioural information may act as a proxy for protected characteristics. Institutions must consider applicable data-protection, consumer-credit, insurance, banking-secrecy and anti-discrimination requirements.

10.3 Accountability and model-ready data

A governed analytical dataset should have a named owner, approved purpose, version, lineage record, quality report and access classification. Human review remains necessary when automated preparation systems infer entities, impute sensitive variables or classify text. Generative AI can assist with schema mapping or documentation, but generated transformations must be tested because plausible output may still be incorrect.

11. Illustrative Case: Building an Early-Warning Dataset

Background

This is an illustrative case. Meridian Bank wants to identify small-business borrowers at risk of entering 30-day arrears within the next three months. Relevant data are distributed across loan, transaction, customer, credit-bureau and macroeconomic systems.

Problem

The first prototype produced unusually high accuracy. Investigation found several defects: current arrears status was included among predictors, bureau reports were joined using their latest available version, customer records were duplicated after a many-to-many account join, and missing transaction histories were treated as zero activity.

Analysis

Current arrears status directly revealed the future label for some observations and therefore caused leakage. Latest-version bureau records introduced information published after historical decision dates. Row multiplication gave customers with multiple accounts excessive weight. Zero imputation incorrectly equated absent feeds with genuine inactivity. Furthermore, closed failed businesses were missing from one customer extract, creating survivorship bias.

Approach and solution

The bank established one row per borrower and monthly observation date. Transactions were aggregated by borrower using windows ending at each observation date. Bureau records were attached through an as-of join based on receipt time. Account balances were reconciled to controlled totals before and after integration.

Missing transaction data were assigned reason codes. Genuine zero activity remained zero, while unavailable feeds were represented by missing values and availability indicators. Closed borrowers were recovered from the historical archive. All transformations were fitted on training periods, and later months were reserved for out-of-time validation.

Governance controls documented variable definitions, source ownership, access rights and lineage. Segment-level performance and missingness were reviewed by business size, region and customer tenure. High-risk flags supported human review rather than automatic withdrawal of credit.

Outcome

Measured predictive performance fell from the implausibly high prototype result, but out-of-time performance became more stable and operationally credible. Reconciliations showed complete borrower coverage, and decision reviewers could trace each feature to source records available at the historical observation date.

Learning points

1. Exceptional model performance can indicate leakage rather than insight.
2. Observation units and temporal availability must be defined before integration.
3. Missingness reasons carry more meaning than a universal replacement value.
4. Reconciliation, lineage and fairness review are integral to preparation.
5. A less impressive but reproducible model is more valuable than a contaminated one.

12. Contemporary Practice and Best-Practice Checklist

Open-banking interfaces, real-time payments, cloud lakehouses and machine-readable disclosures are increasing data volume and speed. These technologies do not remove the need for controls. Streaming systems require event-time handling, late-arrival policies and continuously monitored schemas. Data vendors and AI-derived features require provenance and validation just as internal fields do.

Professional practice should follow these principles:

• define the decision, population, entity and time horizon first;
• preserve immutable raw data and versioned transformation code;
• use explicit units, currencies, calendars and availability timestamps;
• profile quality by time, source and relevant segment;
• investigate outliers rather than deleting them mechanically;
• prevent leakage by using point-in-time information;
• verify join cardinality and reconcile totals after integration;
• use sampling estimators and variances appropriate to the design;
• separate training, validation and test periods before fitting transformations;
• document lineage, permissions, limitations and unresolved bias.

Chapter Summary

Financial analytics depends on accurate interpretation and disciplined preparation of data. Market, accounting, banking, insurance, macroeconomic and alternative sources differ in timing, granularity, structure and risk. Measurement requires explicit units, scales, entity definitions and availability dates. Cleaning addresses missingness, duplication, inconsistent formats and outliers without erasing economically meaningful events. Transformation supports comparability, while integration requires controlled keys, cardinality checks, as-of joins and reconciliation.

Sampling inference must correspond to the actual design. In stratified sampling, population weights, stratum-specific variances and finite-population corrections determine the estimator and its uncertainty. Exploratory analysis examines distributions, relationships, time variation and possible leakage. Throughout the pipeline—from source systems to reports and decisions—governance must ensure quality, lineage, privacy, security, lawful use and accountability. Reliable preparation is therefore not a preliminary clerical step; it is the evidential foundation of financial analysis.

Key Terms

Alternative data: Information generated outside conventional financial reporting and transaction systems.

As-of join: A temporal join using the most recent eligible record available by a specified time.

Data contract: An agreement defining a dataset’s schema, meaning, delivery and quality expectations.

Data leakage: Use of information unavailable at the intended prediction or decision time.

Data lineage: Documentation of data origins and transformations.

Data quality: The fitness of data for a specified purpose.

Entity resolution: The process of determining whether different records refer to the same real-world entity.

Exposure: The amount of time or risk subject to a possible event.

Finite-population correction: A variance adjustment used when sampling without replacement from a finite population.

Granularity: The level of detail represented by each observation.

Imputation: Replacement of missing values using a defined method.

Look-ahead bias: Bias caused by using information not yet available at the historical decision time.

Metadata: Information describing the meaning, structure, origin or processing of data.

Outlier: An observation unusually distant from much of a distribution.

Point-in-time data: Data preserving what was known or available at each historical date.

Sampling frame: The operational list or mechanism from which a sample is selected.

Stratified sampling: Probability sampling conducted separately within defined population subgroups.

Survivorship bias: Distortion caused by excluding entities that failed or left the population.

Unit of observation: The entity, event or period represented by one dataset row.

Review Questions

Short Answer Questions

1. Distinguish a stock variable from a flow variable.
2. Why should an analyst retain both event time and ingestion time?
3. What is the difference between a legitimate extreme value and a data error?
4. What is the purpose of a data contract?
5. Why can an inner join create selection bias?

Long Answer Questions

1. Compare market, accounting, banking, insurance, macroeconomic and alternative data in terms of timing, granularity and quality risks.
2. Explain the main missing-data mechanisms and evaluate alternative treatment methods.
3. Describe a governed financial-data pipeline and explain why governance must span every stage.
4. Discuss how point-in-time integration prevents look-ahead bias in investment or credit analysis.
5. Evaluate the ethical and regulatory risks associated with alternative data in financial decisions.

Analytical / Application Questions

1. Design a preparation plan for combining policy, claim, weather and customer data to predict insurance claims. Specify the unit of observation, timestamps, joins and quality controls.
2. A bank discovers that model performance declines sharply in later periods. Develop an EDA strategy to distinguish economic change from a data-pipeline defect.
3. A loan portfolio contains retail, small-business and corporate strata with different sampling fractions. Explain how you would estimate the overall default proportion and its sampling uncertainty.

Multiple Choice Questions with Answers

Question 1

Which timestamp is most directly relevant when preventing look-ahead bias in a historical forecasting study?

A. The date on which the analyst downloaded the final dataset

B. The date on which the information became available to the historical decision-maker

C. The current date of model estimation

D. The earliest date stored in any source table

Correct Answer: B. The date on which the information became available to the historical decision-maker

Question 2

A dataset contains one row per transaction and another contains one row per customer. What should normally be done before constructing one-row-per-customer features?

A. Join both tables without checking cardinality

B. Delete customers with multiple transactions

C. Aggregate transactions by customer using a defined time window

D. Convert customer identifiers into continuous numbers

Correct Answer: C. Aggregate transactions by customer using a defined time window

Question 3

Nominal deposits increase by 8% while inflation over the same period is 5%. Using the exact relationship, what is the real growth rate?

A. 2.86%

B. 3.00%

C. 3.81%

D. 13.40%

Correct Answer: A. 2.86%

Question 4

Which statement about an outlier is correct?

A. Every outlier must be deleted before modelling

B. An outlier always proves that fraud occurred

C. An outlier may be an error or an economically meaningful extreme event

D. An outlier cannot affect a sample mean

Correct Answer: C. An outlier may be an error or an economically meaningful extreme event

Question 5

In the worked stratified sample, retail loans have weight 0.80 and estimated default rate 0.08, while small-business loans have weight 0.20 and estimated rate 0.15. What is the stratified estimate?

A. 8.0%

B. 9.4%

C. 11.5%

D. 15.0%

Correct Answer: B. 9.4%

Question 6

Why is the unweighted sample default rate inappropriate in the worked stratified example?

A. Default is a categorical outcome

B. The sample contains more than one hundred loans

C. The strata were sampled at different rates relative to their population sizes

D. Finite-population corrections can never be used for proportions

Correct Answer: C. The strata were sampled at different rates relative to their population sizes

Question 7

Which control best detects row multiplication caused by a many-to-many join?

A. Standardizing all numerical fields

B. Comparing row counts, unique keys and financial totals before and after the join

C. Replacing missing values with zero

D. Sorting records alphabetically

Correct Answer: B. Comparing row counts, unique keys and financial totals before and after the join

Question 8

Which transformation requires strictly positive original values?

A. A natural-log transformation ln(x)

B. A missingness indicator

C. A category-frequency table

D. A rank ordering

Correct Answer: A. A natural-log transformation ln(x)

Question 9

Which action most directly supports reproducibility and auditability?

A. Overwriting raw values after each correction

B. Deleting old dataset versions

C. Preserving immutable raw data and versioned transformation code

D. Allowing every employee unrestricted data access

Correct Answer: C. Preserving immutable raw data and versioned transformation code

Question 10

A credit model uses the final recovery amount to predict risk at loan approval. What is the primary problem?

A. Sampling variance

B. Data leakage

C. Currency conversion

D. Stratification

Correct Answer: B. Data leakage