Week 1: Foundation
- EDA (2 days): Distribution analysis, missing values, target leakage check
- Baseline (1 day): Simple model, establish CV score
- Reproducibility (1 day): Fix seeds, set up CV pipeline
Week 2-3: Feature Engineering
Work in sprints of 2-3 features, commit only if CV improves:
- Start with domain-obvious features
- Add aggregations by key entity columns
- Extract date/time parts
- Compute pairwise interactions
Week 4: Ensemble
- Train 5-10 diverse base models
- Hold out 20% as ensemble validation set
- Use simple Optuna-tuned weighted average
- Stacking only if you have time for proper CV
Last 3 Days: Final Push
- Submit 3-4 ensemble variants
- Select final based on CV+LB balance
- Don't overfit to leaderboard (keep 2 submissions in reserve)