Data Analysis Process
The Data Analysis Process domain tests your knowledge of the lifecycle from business question to deployed result. Expect questions on CRISP-DM, preparation, sampling, model evaluation and iteration.
CRISP-DM phases
CRISP-DM has six phases: business understanding, data understanding, data preparation, modeling, evaluation and deployment. The phases are iterative, not a strict waterfall; evaluation can send you back to data preparation or business understanding. Business understanding defines objectives, success criteria and constraints before data work begins.
Preparation and modeling
Data preparation commonly consumes most project effort: cleaning, integrating, transforming, handling missing values and engineering features. ETL extracts, transforms and loads data into a warehouse; ELT loads raw data first and transforms it in the target system. Sampling must match the question: simple random, stratified, systematic and cluster samples differ, while convenience samples invite bias. Split data into training, validation and test sets. Training fits the model; validation tunes it; the test set gives a final unbiased estimate. Cross-validation reuses data for tuning. Overfitting means fitting noise and failing on new data; underfitting means missing real structure. Data leakage occurs when information unavailable at prediction time enters training. A data dictionary and versioned datasets make preparation reproducible.
Worked example
A bank wants to reduce churn. Business understanding sets a goal to cut voluntary closures. Data understanding reviews account and service-call data. Preparation removes duplicate records and creates tenure features. A logistic regression model is trained, tuned on validation data, and judged on a held-out test set using recall and business cost, not accuracy alone. Deployment scores customers monthly; monitoring detects drift. Reasoning: the metric follows the business cost of missing a churner.
Traps
Skipping business understanding produces a technically valid answer to the wrong question. Reusing the test set during tuning inflates results. Accuracy misleads on imbalanced churn data. Confusing ETL with ELT, or treating CRISP-DM as one-pass, also costs marks.