QM003 Hands-on Lab
Move the feature across the historical information boundary
Goal
Use the same synthetic series as the article to compare an admissible feature with two future-aware alternatives.
- Clean: \(x_t\)
- Centered: \((x_{t-1}+x_t+x_{t+1})/3\)
- Future: \(x_{t+1}\)
At forecast origin \(t\), the last two features use information that is not yet available.
Python
python labs/python/qm003_hands_on.pyYou can isolate one feature with --feature clean, --feature centered, or --feature future.
R
Rscript labs/r/qm003_hands_on.R allUse clean, centered, or future instead of all to run one feature.
Interpretation
The exercise is intentionally synthetic. Do not treat the size of the RMSE difference as a general estimate of leakage bias. The point is that a held-out sample is not valid when the feature itself contains future information.