Feature engineering and preprocessing · insight-1723
Re-express money features in a stable unit when the currency's purchasing power moves between training and serving
- Evidence type
- practitioner heuristic
- Difficulty
- medium
- Roles
- Data scientist, ML engineer
- Industries
- aviation, retail, finance
Summary: situation, action, reason, boundary
Situation
A model uses nominal fares or spend in a currency with sustained inflation or a drifting exchange rate, and is trained on past months and scored on later ones.
Action
Convert money features to constant prices with a dated index, or to a price relative to a reference such as the route's median fare that week. Version the index with the model and apply the same rule at serving.
Reason
Reasoning, not a cited finding: a model learns from the number it sees, so if the unit's value changes the feature drifts while customer behaviour need not.
Boundary
Multi-month training on prices or spend in high-inflation or multi-currency settings.
Symptom or risk
The same real price lands on a new nominal level: linear models extrapolate it and trees treat unseen high values as the top training bin, so the learned price response goes stale silently.
Exceptions
Prices customers react to in nominal terms, or very short windows.
Trade-offs
The index is another input to version and may lag.
Pitfalls
Using today's index for history; deflating target and feature with different series.
Example
Hypothetical arithmetic at 4 percent monthly drift: an unchanged product costs 1.04^11 = 1.54 times its month-1 price in month 12, while a months 1-6 training window only reaches 1.04^5 = 1.22 times.