Predictive Analytics for Business

• A. To describe what happened in the past using historical reports
• B. To forecast likely future outcomes based on historical data and statistical models ✓
• C. To store large volumes of data in a centralized database
• D. To visually display current business performance on dashboards

Why: Predictive analytics uses historical data, statistical techniques, and machine learning to estimate what is likely to happen in the future. Describing past events is descriptive analytics, storage is a data management task, and dashboards are a reporting tool.

• A. What was our total revenue last quarter?
• B. How many employees are currently on payroll?
• C. Which customers are most likely to cancel their subscription next month? ✓
• D. What is the average delivery time for orders placed in March?

Why: Predicting which customers will churn requires forecasting a future behavior, which is the core of predictive analytics. The other options ask about past or current facts that can be answered with descriptive reporting.

• A. Test data
• B. Training data ✓
• C. Live data
• D. Metadata

Why: Training data is the labeled or historical dataset a model learns from to identify relationships between input variables and outcomes. Test data is used afterward to evaluate performance, live data is streaming input used after deployment, and metadata describes data rather than serving as learning input.

• A. Descriptive analytics
• B. Diagnostic analytics
• C. Predictive analytics ✓
• D. Prescriptive analytics

Why: Predictive analytics focuses on estimating future outcomes. Descriptive analytics summarizes what happened, diagnostic analytics explains why it happened, and prescriptive analytics recommends actions to take.

• A. To reduce the amount of data needed for future projects
• B. To check whether the model generalizes well to new, unseen cases rather than simply memorizing the training data ✓
• C. To make the model run faster in production
• D. To satisfy legal requirements for data storage

Why: A model can overfit, performing very well on training data while failing on new data. Testing on held-out data reveals how well it will generalize to real-world situations, which is what matters for business decisions.

• A. Building the final predictive model and tuning its hyperparameters
• B. Cleaning, transforming, and organizing raw data so it is suitable for modeling ✓
• C. Presenting model results to business stakeholders through dashboards
• D. Deploying the trained model into a production environment

Why: Data preparation covers the work of cleaning, transforming, and organizing raw data into a usable form before any modeling takes place. Model building, presentation, and deployment all come later in the workflow.

• A. Delete the entire dataset because any missing values make it unusable
• B. Ignore the column entirely in all future analysis
• C. Impute the missing values, for example with the median age, after checking why the data is missing ✓
• D. Replace every missing value with zero regardless of context

Why: Imputation with a suitable statistic such as the median is a common way to handle missing values, especially after investigating the reason for the gaps. Replacing with zero would distort the distribution, and discarding the whole dataset or column throws away useful information.

• A. Creating new input variables from existing data that may help a model make better predictions ✓
• B. Writing the software code that runs the predictive model
• C. Selecting the business metric used to evaluate model success
• D. Collecting new data from external sources without transforming it

Why: Feature engineering means deriving new predictive variables, such as ratios, date components, or aggregated totals, from the raw data. These features often improve model performance by expressing patterns more clearly.

• A. Most predictive models require numerical inputs, so categorical text must be encoded, for example with one-hot encoding ✓
• B. Text variables always reduce model accuracy and should be removed
• C. Text labels must be converted into dates before they can be used
• D. Categorical variables can only be used in unsupervised learning

Why: Many algorithms require numeric inputs, so categorical variables need an encoding step such as one-hot encoding or label encoding. Categorical features are not inherently harmful, and they are widely used in supervised learning.

• A. It makes the dataset smaller so that models train faster in all cases
• B. It ensures that the model is evaluated on data it has not seen, giving a more honest estimate of performance on new data ✓
• C. It removes outliers automatically from both sets
• D. It is required so that feature engineering can be skipped

Why: Holding out a test set allows evaluation on unseen data, which reveals whether the model generalizes rather than simply memorizing the training data. Splitting does not automatically remove outliers or eliminate the need for feature engineering.

• A. Median
• B. Mode
• C. Mean ✓
• D. Range

Why: The mean is calculated using every value in the dataset, so a few extremely large or small values pull it away from the typical value. The median and mode are far less sensitive to outliers.

• A. The middle value of the data
• B. The most frequently occurring value
• C. The average distance of data points from the mean ✓
• D. The difference between the largest and smallest values

Why: Standard deviation quantifies how spread out the values are around the mean, indicating the typical distance of observations from the average. The range, not the standard deviation, is the difference between maximum and minimum.

• A. The mean, because it is always the most accurate measure
• B. The median, because it is resistant to the influence of high values ✓
• C. The range, because it shows the full spread
• D. The variance, because it accounts for all squared deviations

Why: In right-skewed data, the high values inflate the mean, making it unrepresentative of a typical month. The median is not affected by extreme values and better reflects the central location of the data.

• A. Pie chart
• B. Scatter plot
• C. Histogram ✓
• D. Line chart over time

Why: A histogram groups numerical values into bins and displays their frequencies, revealing shape, center, spread, and skewness of a single variable. Scatter plots compare two variables, and pie charts suit categorical proportions.

• A. Calculating the exact average of the dataset
• B. Measuring the spread of the middle 50% of the data and identifying potential outliers ✓
• C. Counting the number of unique values
• D. Determining the most common category

Why: The IQR is the difference between the 75th and 25th percentiles, capturing the spread of the central half of the data. It is commonly used with fences at 1.5 times the IQR to flag potential outliers.

• A. The value of the dependent variable when the independent variable is zero
• B. The average change in the dependent variable for a one-unit increase in the independent variable ✓
• C. The correlation between the two variables
• D. The total variance explained by the model

Why: The slope measures how much the predicted value of the dependent variable changes, on average, for each one-unit increase in the independent variable. The intercept is the value when the predictor is zero.

• A. Mean absolute error
• B. Root mean squared error
• C. R-squared ✓
• D. Standard deviation of residuals

Why: R-squared (the coefficient of determination) ranges from 0 to 1 and expresses the share of variation in the outcome that the model's predictors account for. MAE and RMSE measure prediction error magnitude, not explained variance.

• A. Cross-sectional regression
• B. Time series regression that includes a time trend or lagged variables ✓
• C. Logistic regression
• D. Principal component analysis

Why: Because the data are ordered over time and the goal is to forecast future values, a time series regression that accounts for trend and lagged effects is appropriate. Logistic regression predicts categories, and PCA is a dimensionality reduction technique.

• A. To make the dataset larger
• B. To evaluate how well the model predicts data it has not seen during fitting ✓
• C. To remove outliers automatically
• D. To ensure all variables are normally distributed

Why: A model can fit the training data very closely without generalizing well. Evaluating on a held-out test set estimates real forecasting performance and helps detect overfitting.

• A. The model is perfectly specified
• B. The model may be missing a trend, seasonality, or other time-related structure ✓
• C. The sample size is too large
• D. The dependent variable is categorical

Why: Random residuals indicate the model has captured the systematic signal. A pattern over time suggests autocorrelation or an omitted time-related component, meaning the forecasts can likely be improved.

• A. A continuous numerical value such as total spend
• B. A categorical label such as 'churned' or 'stayed' ✓
• C. The date of each customer's last purchase
• D. The number of support calls made by each customer

Why: Classification predicts discrete categories. For churn, the target is a categorical label indicating whether a customer left or remained, while the other options are either continuous values, dates, or input features.

• A. Precision
• B. Accuracy
• C. Recall ✓
• D. Mean squared error

Why: Recall (also called sensitivity or true positive rate) is the share of actual positives the model correctly identifies. Precision measures how many predicted positives are correct, and mean squared error is a regression metric.

• A. It is a node that partitions the data to improve class separation ✓
• B. It is the final predicted class for all customers
• C. It is a regularization penalty added to the loss function
• D. It is a measure of model accuracy on the test set

Why: Each internal node of a decision tree tests a feature condition and divides the data into subsets, aiming to make the resulting groups more homogeneous in their class labels. Leaf nodes hold final predictions, not splits.

• A. It shows the exact weights learned by a logistic regression model
• B. It displays counts of true positives, true negatives, false positives, and false negatives ✓
• C. It automatically tunes the model's hyperparameters
• D. It compresses high-dimensional data into two principal components

Why: A confusion matrix tabulates correct and incorrect predictions by class, revealing the types of errors a model makes. This gives more insight than accuracy alone, especially when classes are imbalanced.

• A. False positives, where legitimate transactions are flagged
• B. False negatives, where fraudulent transactions go undetected ✓
• C. True negatives, where legitimate transactions are correctly passed
• D. Training error, measured on the data used to fit the model

Why: A false negative means fraud was missed, which is the costly outcome described. Businesses in this situation often prioritize recall, accepting more false positives to catch more actual fraud.

• A. Because training data is always corrupted and unusable
• B. Because evaluating on training data overstates performance, since the model may have memorized noise rather than learned general patterns
• C. Because the law requires a separate test dataset for all business models
• D. Because unseen data always contains more predictive signal than training data

• A. The proportion of all actual positives that were correctly identified
• B. The proportion of all predictions that were correct, regardless of class
• C. The proportion of predicted positives that are actually positive
• D. The proportion of actual negatives that were correctly identified

• A. To increase the size of the training dataset by duplicating rows
• B. To reduce the number of features in the model automatically
• C. To estimate model performance more reliably by training and testing on different data partitions
• D. To speed up model training by splitting the computation across machines

• A. The model is perfectly calibrated and ready for deployment
• B. The accuracy metric is misleading because a model that always predicts 'no churn' would also score about 98%
• C. The model has overfit to the minority class
• D. The dataset must contain duplicate records

• A. A model performs well on training data but poorly on new, unseen data
• B. A model performs poorly on both training and test data
• C. A model uses too few input variables to capture the business problem
• D. A model's predictions are consistently higher than actual values across all records

• A. To make the model's code more complex and accurate on training data
• B. To integrate the model into business processes so its predictions can support real decisions ✓
• C. To store historical data in a centralized database
• D. To replace all human decision-makers with automated systems

Why: Deployment means putting the model into production so that its predictions are actually used to inform business decisions, such as scoring customers or forecasting demand.

• A. Overfitting
• B. Feature engineering
• C. Model serving ✓
• D. Cross-validation

Why: Model serving refers to hosting a trained model behind an endpoint or API so that applications can send inputs and receive predictions in real time or in batches.

• A. Because deployed models never need to be updated once released
• B. Because model performance can degrade as the data patterns in the real world change, a problem known as drift ✓
• C. Because monitoring reduces the need to collect any new data
• D. Because monitoring is only required for regulatory paperwork

Why: Data drift and concept drift cause a model's accuracy to decline as customer behavior, markets, or inputs shift, so ongoing monitoring is needed to detect when retraining is necessary.

• A. Real-time online inference for every single customer click
• B. Batch scoring run on a weekly schedule ✓
• C. Manual calculation by analysts in spreadsheets each day
• D. Deploying the model only on a developer's laptop

Why: When predictions are needed periodically rather than instantly, batch scoring on a schedule is simpler and more cost-effective than maintaining a real-time endpoint.

• A. Hiding the model's logic so competitors cannot copy it
• B. Documenting the model's assumptions, data sources, and limitations, and making its outputs explainable to stakeholders ✓
• C. Deploying the model without testing it on data it has not seen before
• D. Using the most complex algorithm available regardless of interpretability

Why: Clear documentation and explainability help stakeholders understand what the model does, where it may fail, and how to use its predictions responsibly, which builds trust and supports governance.