XGBoost Time Series
XGBoost trained on lag features for time series forecasting
XGBoost Time Series transforms the forecasting problem into supervised learning by engineering lag features, rolling statistics, and calendar features from the time index. The trained XGBoost model then predicts future values from these derived features.
When to use:
- Non-linear patterns in the time series that statistical models miss
- Multiple related time series (XGBoost can use cross-series features)
- When rich feature engineering (external variables, calendar features) is available
Input:
- Trained model checkpoint — exported XGBoost model
- Preprocessing config — lag feature engineering settings
- Training tail — last N observations for lag feature computation
- Steps — forecast horizon
Output: Forecasted values for the specified horizon
Model Settings (set during training, used at inference)
The lag features used (e.g., lag-1, lag-7, rolling-mean-14) are defined during training. Key XGBoost parameters:
N Estimators (default: 100) Number of boosting rounds.
Max Depth (default: 6) Tree depth.
Learning Rate (default: 0.1) Shrinkage per step.
Lags (set during training) Which historical lag steps to include as features.
Inference Settings
No dedicated inference-time settings. The same lag features computed during training are generated from the provided training tail.