Problem Framing and Data Preparation

A business stakeholder describes a goal in plain English (e.g., reduce policy lapse) and you must translate it into a modeling problem.

Identify the response (lapse/no lapse, claim amount, etc.) and decide whether it is regression or classification. Define the unit of analysis (policy, claim, customer) and the prediction horizon. State assumptions about data availability at scoring time — do not use post-event variables as predictors.

Translate "what business problem" → "what response, unit, and prediction horizon".

A raw dataset is loaded into R and you must perform an EDA before any modeling.

Inspect each variable: type, distribution, missingness, outliers. Plot the response against key candidate predictors. Check for class imbalance in classification problems. Identify potential leakage variables (those known only after the prediction time). Note categorical variables with rare levels that may need consolidation.

EDA = data audit: types, distributions, missingness, leakage, response relationships.

EDA reveals a skewed variable, a categorical with many rare levels, and a non-linear relationship between two predictors and the response.

Apply log or power transforms to skewed continuous variables. Bin or collapse categorical variables with few observations per level. Add interaction terms when the relationship between a predictor and the response varies across groups. Document each transformation with a business rationale so it can be defended in the report.

Common transforms: log for skew, bins for sparse categories, interactions for differential effects.