Train test split concept

A train/test split reserves held-out data to estimate how models generalize. Training only on part of the data reduces overfitting to quirks you would not see in production.

Typical ratios

• 80/20 or 70/30 — common starting points for medium datasets
• Small data — cross-validation instead of single split (next modeling lessons)

Stratified split

For classification, keep class proportions similar in train and test—important when labels are imbalanced.

Time-based splits

For forecasting, random shuffles leak future into past. Split by time: train on older dates, test on recent.

Random seed

import random
random.seed(42)
# shuffle indices, cut at split point

Same seed → same split for debugging. Document seed in experiment logs.

Important interview questions and answers

1. Q: Why hold out test data?
   A: Unbiased estimate of performance on unseen rows—tune on train/validation only.
2. Q: Data leakage in split?
   A: Duplicates across train/test (same user twice) inflate metrics—dedupe first.

Self-check

1. What is the purpose of a test set?
2. When should you split by time instead of randomly?
3. Why set random.seed before splitting?

Pitfall: Random split on time-series data causes leakage—use time-based split.