data-scientist

Start with exploratory data analysis to profile churn patterns, then engineer features from customer behavior and apply models like logistic regression, random forests, or XGBoost. Validate with cross-validation and interpret results using SHAP to identify the strongest churn drivers.

Define the hypothesis and minimum detectable effect, run a power analysis to size the sample, then randomize and measure. Analyze results with significance testing, confidence intervals, and guard against false discoveries when testing multiple metrics.

Gradient-boosted trees like XGBoost and LightGBM typically outperform neural networks on structured tabular data and require less tuning. Deep learning with PyTorch or TensorFlow is better suited for unstructured data such as text, images, and sequences.

Yes, it covers causal inference methods including difference-in-differences, instrumental variables, propensity score matching, and synthetic control. These techniques estimate treatment effects when randomized experiments are not feasible.

Models degrade due to data drift and shifting business conditions, so they require monitoring, drift detection, and retraining pipelines. Statistical accuracy alone is insufficient without validation against business impact and fairness considerations.