WGU D502 BHN1 Task 3: Project Report Data Analytics
Capstone, Random Forest Purchase Prediction, Model Results
& Panopto Presentation 2026/2027 Updated Complete
Submission
This capstone assessment evaluates mastery of an end-to-end data analytics project: problem framing, data
preparation, random forest model development, rigorous model evaluation, and professional presentation of
results via Panopto. It emphasizes defensible methodology, correct interpretation of performance metrics, and
clear communication to stakeholders. It contains 69 multiple-choice questions, each with four distractors and a
fully worked rationale that explains why the keyed answer is correct. Questions are organized into clearly labelled
sections that mirror the major content areas of the course. Targeted learning outcomes include: Formulate a
business-relevant purchase prediction problem and select appropriate data, features, and evaluation metrics.;
Build, tune, and validate a random forest classifier while avoiding data leakage and overfitting.; Interpret model
results (confusion matrix, ROC-AUC, feature importance) and communicate actionable insights in a recorded
Panopto presentation.. Every item has been reviewed for clinical accuracy, current guidelines, and clarity so that
students can study with confidence and self-correct as they work through the bank. Use it as a high-yield review
immediately before the exam, or as a structured practice tool during the unit - the rationales double as concise
teaching notes. The recommended writing time is 3 hours, with a passing score of Competent (equivalent to B or
higher; A+ target). Aligned with Aligned with WGU D502 BHN1 Task 3 (2026/2027) and US graduate-level
assessment standards for data analytics capstones. standards and reflects the question style commonly seen on
Section 1: General (Questions 1-69)
1 In a random forest, which mechanism most directly reduces variance without
materially increasing bias?
A) Increasing tree depth until training error reaches zero
B) Averaging decorrelated trees grown on bootstrap samples
C) Using a single decision tree with cost-complexity pruning
D) Standardizing all features before splitting
Answer: B
Rationale: Averaging many decorrelated trees reduces variance while keeping bias
low; deeper trees increase variance, a single pruned tree is not a forest, and
standardization is not the core variance-reduction mechanism.
2 For an imbalanced purchase dataset (2% buyers), which metric best reflects
minority-class performance?
A) Overall accuracy
B) ROC-AUC only
C) Precision-Recall AUC
D) Mean squared error
Answer: C
,Rationale: PR-AUC focuses on the positive (buyer) class and is more informative
than accuracy or ROC-AUC under severe imbalance; MSE is for regression.
3 Which practice most likely introduces data leakage into a random forest purchase
model?
A) Fitting the scaler on the full dataset before splitting
B) Using out-of-bag samples for validation
C) Setting a random seed for reproducibility
D) Tuning max_depth via cross-validation
Answer: A
Rationale: Fitting preprocessing on the full dataset leaks test information into
training; OOB validation, seeds, and CV tuning are legitimate practices.
4 Which hyperparameter primarily controls the number of features considered at
each split?
A) n_estimators
B) max_features
C) min_samples_leaf
D) bootstrap
Answer: B
Rationale: max_features sets the number of features evaluated per split, directly
controlling tree decorrelation; n_estimators is tree count, min_samples_leaf affects
leaf size, and bootstrap controls sampling.
5 A model yields 0.95 training accuracy and 0.62 test accuracy. Which adjustment
is most appropriate?
A) Increase max_depth
B) Decrease n_estimators to 10
C) Increase min_samples_leaf and reduce max_depth
D) Remove cross-validation
Answer: C
Rationale: The large gap indicates overfitting; increasing min_samples_leaf and
reducing depth regularize the model. Increasing depth or reducing trees worsens
overfitting, and removing CV reduces validation rigor.
6 In a Panopto presentation of model results, which element best demonstrates
analytical rigor to stakeholders?
A) Listing every hyperparameter tried without context
, B) Showing baseline vs. model metrics with a confusion matrix
C) Reading the code aloud line by line
D) Omitting limitations to appear confident
Answer: B
Rationale: Comparing to a baseline and showing a confusion matrix communicates
performance and error trade-offs clearly; exhaustive lists, code reading, and omitting
limitations reduce clarity and credibility.
7 Which statement about random forest feature importance is most accurate?
A) Impurity-based importance is biased toward high-cardinality features
B) Permutation importance cannot be computed on test data
C) Importance values sum to the model's accuracy
D) Importance proves causal effect on purchase
Answer: A
Rationale: Mean decrease in impurity is biased toward high-cardinality features, so
permutation importance is often preferred; importance does not sum to accuracy and
does not establish causality.
8 Which sampling strategy is most defensible for evaluating a purchase model on
temporal data?
A) Random k-fold cross-validation
B) Time-series split with forward chaining
C) Stratified k-fold with shuffling
D) Leave-one-out cross-validation
Answer: B
Rationale: Temporal data require time-aware splits to avoid lookahead bias; random,
shuffled, or leave-one-out CV can leak future information into training.
9 A stakeholder asks why recall was prioritized over precision. Which justification
is strongest?
A) Recall is always higher than precision
B) False negatives (missed buyers) are costlier than false positives
C) Precision cannot be computed for imbalanced data
D) Recall equals ROC-AUC in random forests
Answer: B
Rationale: Metric choice should reflect business costs; if missing a buyer is more
costly than a wasted outreach, recall is prioritized. Recall is not always higher,
precision is computable, and recall is not ROC-AUC.
Capstone, Random Forest Purchase Prediction, Model Results
& Panopto Presentation 2026/2027 Updated Complete
Submission
This capstone assessment evaluates mastery of an end-to-end data analytics project: problem framing, data
preparation, random forest model development, rigorous model evaluation, and professional presentation of
results via Panopto. It emphasizes defensible methodology, correct interpretation of performance metrics, and
clear communication to stakeholders. It contains 69 multiple-choice questions, each with four distractors and a
fully worked rationale that explains why the keyed answer is correct. Questions are organized into clearly labelled
sections that mirror the major content areas of the course. Targeted learning outcomes include: Formulate a
business-relevant purchase prediction problem and select appropriate data, features, and evaluation metrics.;
Build, tune, and validate a random forest classifier while avoiding data leakage and overfitting.; Interpret model
results (confusion matrix, ROC-AUC, feature importance) and communicate actionable insights in a recorded
Panopto presentation.. Every item has been reviewed for clinical accuracy, current guidelines, and clarity so that
students can study with confidence and self-correct as they work through the bank. Use it as a high-yield review
immediately before the exam, or as a structured practice tool during the unit - the rationales double as concise
teaching notes. The recommended writing time is 3 hours, with a passing score of Competent (equivalent to B or
higher; A+ target). Aligned with Aligned with WGU D502 BHN1 Task 3 (2026/2027) and US graduate-level
assessment standards for data analytics capstones. standards and reflects the question style commonly seen on
Section 1: General (Questions 1-69)
1 In a random forest, which mechanism most directly reduces variance without
materially increasing bias?
A) Increasing tree depth until training error reaches zero
B) Averaging decorrelated trees grown on bootstrap samples
C) Using a single decision tree with cost-complexity pruning
D) Standardizing all features before splitting
Answer: B
Rationale: Averaging many decorrelated trees reduces variance while keeping bias
low; deeper trees increase variance, a single pruned tree is not a forest, and
standardization is not the core variance-reduction mechanism.
2 For an imbalanced purchase dataset (2% buyers), which metric best reflects
minority-class performance?
A) Overall accuracy
B) ROC-AUC only
C) Precision-Recall AUC
D) Mean squared error
Answer: C
,Rationale: PR-AUC focuses on the positive (buyer) class and is more informative
than accuracy or ROC-AUC under severe imbalance; MSE is for regression.
3 Which practice most likely introduces data leakage into a random forest purchase
model?
A) Fitting the scaler on the full dataset before splitting
B) Using out-of-bag samples for validation
C) Setting a random seed for reproducibility
D) Tuning max_depth via cross-validation
Answer: A
Rationale: Fitting preprocessing on the full dataset leaks test information into
training; OOB validation, seeds, and CV tuning are legitimate practices.
4 Which hyperparameter primarily controls the number of features considered at
each split?
A) n_estimators
B) max_features
C) min_samples_leaf
D) bootstrap
Answer: B
Rationale: max_features sets the number of features evaluated per split, directly
controlling tree decorrelation; n_estimators is tree count, min_samples_leaf affects
leaf size, and bootstrap controls sampling.
5 A model yields 0.95 training accuracy and 0.62 test accuracy. Which adjustment
is most appropriate?
A) Increase max_depth
B) Decrease n_estimators to 10
C) Increase min_samples_leaf and reduce max_depth
D) Remove cross-validation
Answer: C
Rationale: The large gap indicates overfitting; increasing min_samples_leaf and
reducing depth regularize the model. Increasing depth or reducing trees worsens
overfitting, and removing CV reduces validation rigor.
6 In a Panopto presentation of model results, which element best demonstrates
analytical rigor to stakeholders?
A) Listing every hyperparameter tried without context
, B) Showing baseline vs. model metrics with a confusion matrix
C) Reading the code aloud line by line
D) Omitting limitations to appear confident
Answer: B
Rationale: Comparing to a baseline and showing a confusion matrix communicates
performance and error trade-offs clearly; exhaustive lists, code reading, and omitting
limitations reduce clarity and credibility.
7 Which statement about random forest feature importance is most accurate?
A) Impurity-based importance is biased toward high-cardinality features
B) Permutation importance cannot be computed on test data
C) Importance values sum to the model's accuracy
D) Importance proves causal effect on purchase
Answer: A
Rationale: Mean decrease in impurity is biased toward high-cardinality features, so
permutation importance is often preferred; importance does not sum to accuracy and
does not establish causality.
8 Which sampling strategy is most defensible for evaluating a purchase model on
temporal data?
A) Random k-fold cross-validation
B) Time-series split with forward chaining
C) Stratified k-fold with shuffling
D) Leave-one-out cross-validation
Answer: B
Rationale: Temporal data require time-aware splits to avoid lookahead bias; random,
shuffled, or leave-one-out CV can leak future information into training.
9 A stakeholder asks why recall was prioritized over precision. Which justification
is strongest?
A) Recall is always higher than precision
B) False negatives (missed buyers) are costlier than false positives
C) Precision cannot be computed for imbalanced data
D) Recall equals ROC-AUC in random forests
Answer: B
Rationale: Metric choice should reflect business costs; if missing a buyer is more
costly than a wasted outreach, recall is prioritized. Recall is not always higher,
precision is computable, and recall is not ROC-AUC.
