Exponential Smoothing
Weighted averaging of past observations with exponentially decaying weights
Exponential Smoothing (Holt-Winters) assigns exponentially decreasing weights to past observations. It handles trend and seasonality through additive or multiplicative components and is a fast, robust choice for many business forecasting problems.
When to use:
- Business time series with smooth trend and seasonality
- When a simple, fast, and interpretable model is sufficient
- Short-to-medium forecast horizons without complex dynamics
Input:
- Trained model checkpoint — exported Holt-Winters model
- Preprocessing config — scaling settings
- Training tail — last N observations
- Steps — forecast horizon
Output: Forecasted values for the specified steps
Model Settings (set during training, used at inference)
Trend (default: add)
Trend component type. add for additive trend; mul for multiplicative trend; None for no trend.
Seasonal (default: add)
Seasonal component type. add for additive; mul for multiplicative; None for no seasonality.
Seasonal Periods (default: null — auto) Number of time steps in one seasonal cycle.
Damped Trend (default: false) If true, the trend dampens over the forecast horizon to avoid unrealistic long-term extrapolation.
Smoothing Level (alpha) (default: auto-optimized) Weight for the level component. Lower values smooth more aggressively.
Inference Settings
No dedicated inference-time settings.