Contents x
    Forecast Strategy
    • 13 Jan 2023
    • 3 Minutes to read
    • Print
    • Dark
      Light
    Contents

    Forecast Strategy

    • Updated on 13 Jan 2023
    • 3 Minutes to read
    • Print
    • Dark
      Light
    Article summary

    Description

    The Forecast Strategy node applies several forecast models to a data set.


    Configuration Options

    Basic Configuration Options

    SettingDescription\Parameters
    Split ColumnsThe dataset is split into groups according to the selected column(s). The resulting dataset is not transformed but treated as a unique grouping of data when processed within a node.
    Time GranularityLength of time between records in the input data set. I.e., if each input record represents a week of data, the time granularity is Week.

    Options
    • Day
    • Week
    • Month
    • Quarter
    • Year
    Role ColumnSplits the data set into scoring and training data.
    Training ValueValue in the Role Column for training data.
    Scoring ValueValue is the Role Column for scoring data.
    PredictorsThe columns being used to predict with.
    Model Type FilterPreset selections of models/ensemble models based on the data quality of the input data set.

    Options
    • All
    • Good Signal
    • Poor Signal
    • Heavy Zeros
    ModelsSee Model Details
    Ensemble ModelsSee Model Details
    Forecast HorizonNumber of periods to forecast forward from the end of the input data set.
    Time ColumnOptional column to seed time values for future forecast periods.
    TimeoutThe specified timeout for the action. If this time is exceeded, the node action will fail, regardless of the success of the action.

    Model Details

    Except where specified, each predictive model comes from the R Fable package. Training data sets are used to build the predictive models, which can then forecast future values.

    Some models have been optimized for performance; these are denoted by a Rabbit icon in the model selection dropdown.

    Basic Models


    ModelDescription
    MeanThe forecasts of all future values are equal to the average (or mean) of the historical data.
    ARIMAAutoRegressive Integrated Moving Average. Combination of differencing with autoregression and a moving average model. ARIMA models describe the autocorrelations in the data.
    Exponential Smoothing State SpaceForecasts produced using exponential smoothing methods are weighted averages of past observations, with the weights decaying exponentially as the observations get older; the more recent the observation the higher the associated weight.
    ThetaEquivalent to simple exponential smoothing with drift.
    Linear RegressionTime series linear models.
    ProphetProphet is a procedure for forecasting time series data based on an additive model where non-linear trends are fit with yearly, weekly, and daily seasonality, plus holiday effects. It works best with time series that have strong seasonal effects and several seasons of historical data. Prophet is robust to missing data and shifts in the trend, and typically handles outliers well.
    Seasonal DecompositionDecomposition combination model using an additive/multiplicative decomposition. Decompose a time series into seasonal, trend, and remainder components. Seasonal components are estimated iteratively using STL. Secondary models are used to fit each of the components from the resulting decomposition. Seasonality type (additive/multiplicative) is automatically determined by minimizing MAPE.
    Croston'sBased on Croston's method for intermittent demand forecasting; involves using simple exponential smoothing on the non-zero elements of the time series and a separate application of SES to the times between non-zero elements of the time series. Does not produce point forecasts, i.e., a unique forecast for each time point. It produces the average demand per future period, instead.
    Improved Croston'sBased on the above method. The single average demand per future period generated by Croston’s method may be insufficient if predictions for each month for a product that has intermittent demand are desired. Creates a one-step-ahead forecast using Croston’s method, then incorporates that prediction back into the training set. Uses this expanded training set to produce another one-step-ahead forecast. This process is repeated until you have produced as many predictions as there are time periods in the test set.
    Neural NetworkFeed-forward neural networks with a single hidden layer and lagged inputs for forecasting univariate time series. Does not rely on autoregression to make predictions; can be utilized for poorer quality/intermittent demand data.

    Ensemble Models


    ModelDescription
    Good Signal EnsembleAverage of Mean, Theta, Exponential Smoothing, ARIMA, Seasonal Decomposition, Prophet, Linear Regression. Output column name is shown as GOOD_SIG_ENS.
    Poor Signal EnsembleAverage of Mean, Neural Network. Output column name is shown as POOR_SIG_ENS.
    Intermittent Data EnsembleAverage of Mean, Improved Croston's, Croston's, Neural Network. Output column name is shown as INTER_DATA_ENS.

    Actions

    ActionDescription
    PreviewOnce any forecast models have been built, the combined result set can be previewed at any time.
    Rebuild ModelRebuild any selected models. Produces an .Rdata file that is used to forecast values.
    Was this article helpful?
    What's Next