Scenario parameters
|
Use the Parameters tab to set basic forecasting parameters. |
Parameters
| Parameter | Description |
|---|---|
| Data Parameters | |
| Algorithms |
Algorithms to be used in time series forecasting. This is considered when Forecast_Strategy parameter is not AutoTuning. Available algorithms are: ArimaX, Stochastic Gradient Boosting (Global), SVM Linear, Prophet, Auto Arima, Exponential Smoothing, Croston (Intermittent), Simple Moving Average, Naive, and Seasonal Naive. The default algorithm is ArimaX. |
| Auto Detect Time Horizon | Automatically detects the start timestamp for forecasting. The entire history is used for training time series models. This setting is enabled by default. |
| Backtest Window Size | Determines the number of periods in each prediction set. This is recommended to be the same as the 'Forecast Horizon' parameter for short-term forecasting. The default value is 7. |
| Causal Elimination Strategy |
The causal selection strategy to be used. Available strategies are:
|
| Custom End Date | Data after this timestamp will not be considered for training time series models. Specify only if the Auto Detect Time Horizon parameter is disabled. |
| Custom Start Date |
Data before this timestamp will not be considered for training time series models. Specify only if the Auto Detect Time Horizon parameter is disabled. |
| Demand Correlation Threshold | The minimum correlation required for the causal to be included with an LOD's demand required for a causal to be selected. Default value is 0.15. |
| Effort Level |
Available effort levels are:
For higher effort levels, more models are trained with extensive hyper parameter tuning. This setting is considered only if the Forecast Strategy parameter is set to AutoTuning. |
| Error Metric |
The metric used to evaluate forecasting models. Available metrics are:
|
| Forecast Horizon | The number of periods into the future to include in the forecast. Default value is 20. |
| Forecast Strategy |
Determines whether trained models are down-selected (AutoTuning, Tournament) or combined (Ensemble). Also determines whether algorithms are automatically chosen for training (AutoTuning) or manually chosen by the user (Ensemble, Tournament). Available forecast strategies are:
AutoTuning mode automatically tunes models and parameters based on Effort_Level. With Ensemble or Tournament mode, you should select the Algorithms. |
| Include Causal Lags | Enable this setting to include lags for causal factors. This setting is enabled by default. |
| Lead Time | The number of periods between the end of the training set and the period to forecast. The default is 0. |
| Level Of Detail |
The level of detail for forecast generation. Select a dimension to include its detail data on analysis pages. The default is all three dimensions as defined in the "Dimension definitions" page. When you select a dimension, the analysis pages will allow you to filter that dimension by its members. If you do not select a dimension for Level of Detail, it is aggregated and the analysis pages will only display the aggregated data (the filter for that dimension will only present All as a choice). For example, if you have Location, Sub-Location, and Product for your three dimensions, and you select the Sub-Location (Dimension 2) in Level of Detail, the analysis pages will allow you to filter by Sub-Location, and users will only see "All" as a choice in the Location and Product filters. |
| Multicollinearity Threshold | The maximum correlation allowed between any pair of causals selected for the same level of detail. Among causals with a correlation higher than this threshold, only one of the causals is considered for time series forecasting. The default is 0.95. |
| Time Aggregation |
This is the time parameter for forecast generation. Available periods are:
|
Last modified: Friday May 12, 2023