Predictive Modeling Flashcards
7 cards from real MAC practice questions. Tap to flip, then mark Knew It or Still Learning — missed cards come back until you master them.
Read the first 7 Predictive Modeling flashcards as text
A predictive model has high precision but low recall for identifying high-value customers. What does this imply?
Answer: The model is conservative, identifying few high-value customers but those it flags are usually correct
High precision means most flagged customers are truly high-value, but low recall means the model is missing many actual high-value customers.
In marketing mix modeling (MMM), what does the 'adstock' transformation account for?
Answer: The lagged and decaying effect of advertising on consumer behavior over time
Adstock models the carryover effect of advertising, where exposure in one period continues to influence consumer behavior in subsequent periods with diminishing impact.
Which statistical assumption does multicollinearity violate in linear regression models?
Answer: Independence of predictor variables
Multicollinearity means predictor variables are highly correlated with each other, violating the assumption of predictor independence and making coefficient estimates unstable.
What is the purpose of a 'holdout set' in predictive modeling?
Answer: To provide an unbiased final evaluation of the model after all tuning is complete
A holdout set is kept completely separate during model development and used only for final evaluation, ensuring an unbiased estimate of real-world performance.
When using logistic regression for churn prediction, the model output before applying the sigmoid function is called the:
Answer: Log-odds (logit)
Logistic regression computes a log-odds (logit) value, which is then transformed by the sigmoid function into a probability between 0 and 1.
A data scientist notices that a model trained on Q4 data performs poorly in Q2. This is an example of:
Answer: Model drift caused by concept or data distribution shift
Model drift occurs when the statistical relationship between inputs and outputs changes over time, causing a model trained in one period to degrade in another.
Which sampling technique is most appropriate when a churn dataset has only 2% positive (churned) cases?
Answer: Oversampling the minority class (e.g., SMOTE)
With severe class imbalance, techniques like SMOTE oversample the minority class to give the model enough positive examples to learn meaningful patterns.