Defining algorithm options

This topic describes how to define algorithm options for a time forecasting model scenario, which is part of Adding-a-time-forecasting-model-scenario.

To define algorithm options for a time forecasting scenario

From the Add a new scenario page, click Advanced.
Click the Algorithm Options tab.

Define the algorithm options as listed in the following table.

Note

The options available in the Algorithm Options tab are based on the algorithm you select in the Forecasting Parameters tab.

Algorithm name	Options available	Value
Automatic	Prediction bands confidence	Percentage
	Automatic selection mode	Conservative or Aggressive
	Reliability evaluation mode	Loose - Accepts wider range of relative error or Tight - Requires smaller relative error
By time shifts	Degree of the trend fit	Integer
By time shifts	Reliability evaluation mode	Loose - Accepts wider range of relative error or Tight - Requires smaller relative error
Exponential - multiplicative trend	Degree of the trend fit	Integer
Exponential - multiplicative trend	Reliability evaluation mode	Loose - Accepts wider range of relative error or Tight - Requires smaller relative error
Quadratic	Reliability evaluation mode	Loose - Accepts wider range of relative error or Tight - Requires smaller relative error
Cubic	Reliability evaluation mode	Loose - Accepts wider range of relative error or Tight - Requires smaller relative error
Linear	Prediction bands confidence	Percentage
	Peak detection	None or Weekly peak detection
	Peak selectivity (displayed when Weekly peak detection is selected)	High, Medium, or Low
	Peak neighborhood amplitude (displayed when Weekly peak detection is selected)	Percentage
	Reliability evaluation mode	Loose - Accepts wider range of relative error or Tight - Requires smaller relative error
Robust Linear	Prediction bands confidence	Percentage
	Peak detection	None or Weekly peak detection
	Peak selectivity (displayed when Weekly peak detection is selected)	High, Medium, or Low
	Peak neighborhood amplitude (displayed when Weekly peak detection is selected)	Percentage
	Reliability evaluation mode	Loose - Accepts wider range of relative error or Tight - Requires smaller relative error
Robust Linear - Last Ramp	Prediction bands confidence	Percentage
	Last ramp selection policy	Latest or Latest or most significant (if any)
	Last ramp detection confidence	Percentage
	Peak detection	None or Weekly peak detection
	Peak selectivity (displayed when Weekly peak detection is selected)	High, Medium, or Low
	Peak neighborhood amplitude(displayed when Weekly peak detection is selected)	Percentage
	Reliability evaluation mode	Loose - Accepts wider range of relative error or Tight - Requires smaller relative error
Robust Linear - Smoothed Last Ramp	Prediction bands confidence	Percentage
	Last ramp selection policy	Latest or Latest or most significant (if any)
	Last ramp detection confidence	Percentage
	Reliability evaluation mode	Loose - Accepts wider range of relative error or Tight - Requires smaller relative error
Holt-Winters exponential smoothing	Prediction bands confidence	Percentage
Holt-Winters exponential smoothing	Reliability evaluation mode	Loose - Accepts wider range of relative error or Tight - Requires smaller relative error
Yearly time shift	Prediction bands confidence	Percentage
	Type of the trend fit	Best fitting, Linear, Quadratic, or Cubic
	Use a robust linear trend	Integer
	Past years decay policy	Uniform, Linear, Uniform, or Exponential
	Seasonal Period	Number of days
	Reliability evaluation mode	Loose - Accepts wider range of relative error or Tight - Requires smaller relative error
Robust Exponential Damping	Prediction bands confidence	Percentage
	Sample interval	Integer
	Normalized factor of the sample interval	Integer
	Smoothing factor	Integer
	Reliability evaluation mode	Loose - Accepts wider range of relative error or Tight - Requires smaller relative error
Robust Exponential Damping - Last Ramp	Prediction bands confidence	Percentage
	Sample interval	Integer
	Normalized factor of the sample interval	Integer
	Smoothing factor	Integer
	Last ramp selection policy	Latest or Latest or most significant (if any)
	Last ramp detection confidence	Percentage
	Reliability evaluation mode	Loose - Accepts wider range of relative error or Tight - Requires smaller relative error
Automated Box and Jenkins	Prediction bands confidence	Percentage
	Seasonal period	Based on time resolution, Dominant, or Custom. If you select *Custom, type the number of samples.
	Bounding feature	Active or Not active
	Reliability evaluation mode	Loose - Accepts wider range of relative error or Tight - Requires smaller relative error
Automated Box and Jenkins - Last Ramp	Prediction bands confidence	Percentage
	Seasonal period	Based on time resolution, Dominant, or Custom. If you select *Custom, type the number of samples.
	Bounding feature	Active or Not active
	Last ramp selection policy	Latest or Latest or most significant (if any)
	Last ramp detection confidence	Percentage
	Reliability evaluation mode	Loose - Accepts wider range of relative error or Tight - Requires smaller relative error

Tip

If you are not sure which algorithm is best suited to your data set, BMC recommends selecting Automatic, the default algorithm. When you select Automatic, BMC Helix Capacity Optimization applies the most appropriate mathematical algorithm based on the actual data in the time series.

Prediction bands confidence

Determines the level of confidence interval of the prediction bands.

Reliability evaluation mode

Determines the level of internal thresholds that are used for classifying the prediction accuracy. It drives the classification of the forecast as reliable, warning or unreliable. These three levels are represented with a traffic-light icon that has three different symbols depending on the classification. Tight means small prediction error computed by cross validation is needed in order for the prediction to be classified as reliable. For details, see Reliability Evaluator.

Peak detection

Peak detection allows to isolate peaks in a data series, so that the forecast can be computed based on these peaks only, that means the regression is performed on the peak data only. Hence, only the samples around the peaks are used for training the regressor.

The available options are:

None: No peaks are detected. The regression is done over the provided samples.
Daily peak detection: Peaks are detected on a daily basis. Consider using this option if you have hourly-sampled data.
Weekly peak detection: Peaks are detected on a weekly basis. Consider using this option if you have a several months of data.

The peak detection algorithm is highly heuristics-based and hence BMC recommends to use it for experiment purpose only. If you want a pessimistic future prediction, BMC recommends to perform a standard regression and consider the upper bound of the confidence interval as your pessimistic prediction, instead of performing the regression over the peaks. In this way, more samples are provided for training the regressor. Also, the computation of prediction bands has a strong mathematical foundation.

Peak selectivity

Determines the selectivity of the significance filter applied during the peak detection. High means more peak candidates will be considered as peaks.

Peak neighborhood amplitude

After peaks are detected, for each peak, a subset of neighbor samples is included in the final data series. This subset contains samples whose values are over a defined threshold, computed as a percentage of the peak value.

Last ramp selection policy

Determines the selection criteria of the last ramp. The available options are:

Latest: The latest detected regime change (jumps) will be considered as the starting point for the last ramp. The regime change is an abrupt change into the statistical behavior of a time series.
Latest or most significant: The detected regime changes are considered in the order of their magnitude. If the second largest change is significantly smaller than the largest one, then the largest change is considered as the most significant and its position will be be considered as the starting point for the last ramp. Otherwise, latest detected regime change will be considered.

Last ramp detection confidence

Determines the confidence level for the detection of a regime change. The regime change is an abrupt change into the statistical behavior of a time series. The highest confidence level means the largest confidence interval.

High last ramp detection confidence: Small jumps will not be detected.
Low last ramp detection confidence: Potentially every small jump will be detected.

Where to go from here

After you have defined algorithms for a time forecasting model scenario, you can continue with Output Options.