Improve low R² without p‑hacking

A linear regression predicting contribution per order has R²=0.07. a) List concrete steps to raise predictive performance without invalidating inference: feature transformations (e.g., splines for basket size), interactions (treatment×daypart), appropriate error distribution (Gamma/log link), and leakage checks. b) Will simply adding another covariate reliably increase R² out‑of‑sample? Demonstrate with CV why/why not and propose alternatives (GAMs, quantile regression, gradient boosting) while keeping effect‑estimation goals in mind. c) Show how you’d use nested cross‑validation and target‑leakage tests to guard against p‑hacking while iterating. d) Explain when a low R² is acceptable for unbiased ATE but unacceptable for accurate individual predictions.