Theta Method
Decomposition-based forecasting method, strong performer on many benchmarks
The Theta Method decomposes the time series into two "theta lines" — one capturing long-term trend and one capturing short-term dynamics — then combines their forecasts. It won the M3 forecasting competition and is competitive with far more complex models.
When to use:
- General univariate forecasting where simplicity and reliability are valued
- When you want a strong baseline without complex parameter tuning
- Short-to-medium forecast horizons
Input:
- Trained model checkpoint — exported Theta model
- Preprocessing config — scaling settings
- Training tail — last N observations
- Steps — forecast horizon
Output: Forecasted values for the specified number of steps
Model Settings (set during training, used at inference)
Theta (default: 2.0)
Theta parameter controlling the decomposition. theta=2 is the classic optimal setting derived from the M3 competition.
Seasonality (default: auto-detected) Whether and what seasonal period to use.
Inference Settings
No dedicated inference-time settings.