Time series and forecasting · insight-0234
Detrend or difference the target before using tree models on trending series
- Evidence type
- documented fact
- Difficulty
- medium
- Roles
- Data scientist, ML engineer
- Industries
- saas, retail, telecom
Summary: situation, action, reason, boundary
Situation
A gradient-boosted forecast of a growing subscription base flattens at the highest level seen in training.
Action
Predict differences, growth rates or ratios to a recent level, or remove the trend with a simple model and let the trees learn the residual, then backtest across origins with new highs.
Reason
Regression trees partition the input space and predict a constant within each region, so a tree ensemble cannot continue a trend beyond the feature and target ranges it saw.
Boundary
Growing or declining series modelled with random forests, XGBoost, LightGBM or CatBoost.
Symptom or risk
Tree predictions are piecewise constant over training targets, so they cannot extrapolate a trend beyond the observed range and under-forecast growth.
Exceptions
Stationary series, or series normalised by a recent level, rarely need this.
Trade-offs
Differenced targets accumulate errors when integrated back over long horizons.
Pitfalls
Adding a time index feature and assuming the trees will extrapolate it.
Lessons and concepts
Claims and their sources
Tree-based methods partition the feature space into regions and fit a simple model, typically a constant, in each one.