Time forecasting models - Automatic algorithm
This topic presents implementation details of the Automatic forecasting algorithm. This procedure selects the best forecasting algorithm (among the available ones) in BMC Helix Continuous Optimization and computes the prediction of the time series using it.
To provide the best algorithm, this procedure works in two steps. These are:
Models vector pruning
Initially, the vector of feasible algorithms contains all available algorithms in BMC Helix Continuous Optimization. Then, the time series is inspected to acquire all its features. According to these features, the vectors of the model can be reduced by removing forecasting methods which are not suited for available data.
The following rules apply to reducing the vector:
- If time series monitors less than 1 year of observations > Yearly Time Shifts is removed
- If no significant regime changes are detected > all Last Ramp algorithms are removed
- If data regression is not strongly polynomial or exponential > Quadratic, Cubic and Exponential are removed
- If time series variation is not sufficiently high > Robust Exponential Damping and Robust Exponential Damping - Last Ramp are removed
- If seasonality is not relevant in the data > Holt-Winters Exponential Smoothing is removed.
Best model selection
As a comparison metric, a stable version of the Root Mean Squared Error (RMSE) is computed.
The algorithm achieving the lowest Robust Root Mean Squared Error (RRMSE) is considered the best for data and is used for the prediction. For more information, see Computation of the RRMSE index.
Automatic forecasting algorithm
You also have the option to select the Automatic selection mode from the Scenario editor in the BMC Capacity Optimization console. Possible values for this option are:
- Conservative (default): All available forecasting algorithms are selectable from Automatic procedure, except for Quadratic, Cubic, Exponential, and Yearly time shift.
- Aggressive: All available forecasting algorithms are selectable from Automatic procedure
The Automatic forecasting algorithm is the default prediction method for time series. In this method, every series in the model is forecasted using the best algorithm that fits the shape of the series.
If you want to know which algorithm has been chosen for a specific series, you can refer to the points table and refer to the selected algorithm in the corresponding row. The following image shows an illustration.
