7-Day ML Leaderboard Booster Kit

By Mathurin Aché, Kaggle Grand Master (public Kaggle profile: kaggle.com/mathurinache).

This kit is designed for data scientists who want a sharper optimization workflow without chasing noise.

Day 1: Freeze the Validation Contract

• Write the target, split logic, metric, and leakage risks before training.
• Keep a fold_id column in every experiment file.
• Create one baseline that you are willing to keep as the reference for a full week.

Day 2: Build a Clean Baseline

• Start with simple preprocessing.
• Save out-of-fold predictions.
• Log seeds, folds, metric, and notes.
• Do not tune before the baseline is reproducible.

Day 3: Leakage Sweep

• Check time leakage, grouped entities, duplicates, target-derived features, and preprocessing fitted on full data.
• If the validation score is suspiciously high, treat it as a bug until proven otherwise.

Day 4: Search Space Discipline

• Start broad, then shrink ranges from observed failures.
• Avoid adding ten knobs before learning from three.
• Repeat promising trials with another seed.

Day 5: Error Slicing

• Slice errors by feature bins, categorical levels, prediction confidence, and time.
• Promote changes that improve the painful slices without destroying global performance.

Day 6: Ensemble Only After OOF Hygiene

• Blend out-of-fold predictions, not leaderboard guesses.
• Keep a simple weighted average baseline before stacking.
• Reject ensembles that only work on one split.

Day 7: Decision Review

• Keep the changes that survived repeated evidence.
• Document what failed.
• Turn the best workflow into a reusable notebook.

Next step: package these habits into repeatable assets with the Kaggle GM ML Optimization Pack.