ISYE 6501 Final Exam Practice: 25
Questions with Answers & Explanations
(Georgia Tech)
Questions 1–5: Model Selection
1. A healthcare researcher has daily counts of emergency room visits over five years and notices a strong
weekly pattern (higher on weekends) and increasing overall demand. Which model is most appropriate
for forecasting future daily visits?
A) Principal Components Analysis (PCA)
B) K-Means Clustering
C) Holt-Winters Exponential Smoothing ✅
D) Logistic Regression
Explanation: Holt-Winters exponential smoothing explicitly handles trend (increasing demand)
and seasonality (weekly pattern). PCA and K-means are unsupervised methods not designed for
forecasting. Logistic regression is for classification, not time series.
2. An e-commerce company wants to test five different website layouts. They cannot test all 32
combinations of features simultaneously, so they test a carefully chosen subset to identify which factors
have the largest effect on click-through rate. This approach is called:
A) Multi-armed bandit
B) Fractional factorial design ✅
C) K-fold cross-validation
D) LASSO regression
Explanation: Fractional factorial design tests only a fraction of all possible combinations to identify the
most important factors efficiently. Multi-armed bandit balances exploration/exploitation in real time.
Cross-validation evaluates models. LASSO performs variable selection.
3. A financial analyst models stock returns and finds that the variance of prediction errors changes over
time – large errors tend to cluster together. Which model is specifically designed to handle this?
A) GARCH ✅
, B) ARIMA
C) k-Nearest Neighbors
D) CART
Explanation: GARCH (Generalized Autoregressive Conditional Heteroskedasticity) models changing
volatility (variance clustering) over time. ARIMA models the mean, not variance. K-NN and CART are
general-purpose prediction methods.
4. Which of the following is an unsupervised learning method that reduces dimensionality by creating
new uncorrelated variables that capture maximum variance?
A) Linear regression
B) Logistic regression
C) Principal Components Analysis (PCA) ✅
D) Support vector machine (SVM)
Explanation: PCA transforms original correlated variables into new uncorrelated principal components
ordered by the amount of variance they explain. It is unsupervised (no outcome variable needed).
5. A credit card company wants to predict which transactions are fraudulent. The dataset has 99.9%
legitimate transactions and 0.1% fraudulent ones. Which metric is most appropriate to evaluate model
performance?
A) Accuracy
B) Precision
C) F1 score ✅
D) R-squared
Explanation: With severe class imbalance, accuracy is misleading (predicting all as legitimate gives
99.9% accuracy but catches no fraud). F1 score, the harmonic mean of precision and recall, balances
both and is robust for imbalanced classification. R-squared is for regression.
Questions 6–10: Statistical Concepts
6. In linear regression, adding an irrelevant predictor variable will typically cause:
A) R-squared to increase and Adjusted R-squared to decrease ✅
B) Both R-squared and Adjusted R-squared to increase
Questions with Answers & Explanations
(Georgia Tech)
Questions 1–5: Model Selection
1. A healthcare researcher has daily counts of emergency room visits over five years and notices a strong
weekly pattern (higher on weekends) and increasing overall demand. Which model is most appropriate
for forecasting future daily visits?
A) Principal Components Analysis (PCA)
B) K-Means Clustering
C) Holt-Winters Exponential Smoothing ✅
D) Logistic Regression
Explanation: Holt-Winters exponential smoothing explicitly handles trend (increasing demand)
and seasonality (weekly pattern). PCA and K-means are unsupervised methods not designed for
forecasting. Logistic regression is for classification, not time series.
2. An e-commerce company wants to test five different website layouts. They cannot test all 32
combinations of features simultaneously, so they test a carefully chosen subset to identify which factors
have the largest effect on click-through rate. This approach is called:
A) Multi-armed bandit
B) Fractional factorial design ✅
C) K-fold cross-validation
D) LASSO regression
Explanation: Fractional factorial design tests only a fraction of all possible combinations to identify the
most important factors efficiently. Multi-armed bandit balances exploration/exploitation in real time.
Cross-validation evaluates models. LASSO performs variable selection.
3. A financial analyst models stock returns and finds that the variance of prediction errors changes over
time – large errors tend to cluster together. Which model is specifically designed to handle this?
A) GARCH ✅
, B) ARIMA
C) k-Nearest Neighbors
D) CART
Explanation: GARCH (Generalized Autoregressive Conditional Heteroskedasticity) models changing
volatility (variance clustering) over time. ARIMA models the mean, not variance. K-NN and CART are
general-purpose prediction methods.
4. Which of the following is an unsupervised learning method that reduces dimensionality by creating
new uncorrelated variables that capture maximum variance?
A) Linear regression
B) Logistic regression
C) Principal Components Analysis (PCA) ✅
D) Support vector machine (SVM)
Explanation: PCA transforms original correlated variables into new uncorrelated principal components
ordered by the amount of variance they explain. It is unsupervised (no outcome variable needed).
5. A credit card company wants to predict which transactions are fraudulent. The dataset has 99.9%
legitimate transactions and 0.1% fraudulent ones. Which metric is most appropriate to evaluate model
performance?
A) Accuracy
B) Precision
C) F1 score ✅
D) R-squared
Explanation: With severe class imbalance, accuracy is misleading (predicting all as legitimate gives
99.9% accuracy but catches no fraud). F1 score, the harmonic mean of precision and recall, balances
both and is robust for imbalanced classification. R-squared is for regression.
Questions 6–10: Statistical Concepts
6. In linear regression, adding an irrelevant predictor variable will typically cause:
A) R-squared to increase and Adjusted R-squared to decrease ✅
B) Both R-squared and Adjusted R-squared to increase
