ForecastingTrainingConfig

Training config for the FORECASTING problem type

KEY TYPE Description
ADDITIONAL_FORECAST_KEYS None List[str]: List of categoricals in timeseries that can act as multi-identifier.
PROBABILITY_QUANTILES List[float] Prediction quantiles.
USE_ITEM_ID bool Include a feature to indicate the item being forecast.
HISTORY_LENGTH int While training, how much history to consider.
USE_TIMESERIES_WEIGHTS_IN_OBJECTIVE bool If True, we include weights from column set as "TIMESERIES WEIGHT COLUMN" in objective functions.
CUSTOM_METRICS List[str] Registered custom metrics available for selection.
FILTER_ITEMS bool Filter items with small history and volume.
ZERO_PREDICTOR bool Include subnetwork to classify points where target equals zero.
EXPERIMENTATION_MODE ExperimentationMode Selecting Thorough Experimentation will take longer to train.
BATCH_RENORMALIZATION bool Enable batch renormalization between layers.
USE_ALL_ITEM_TOTALS bool Include as input total target across items.
ENABLE_COLD_START bool Enable cold start forecasting by training/predicting for zero history items.
LOCAL_SCALE_TARGET bool Using per training/prediction window target scaling.
PREDICTION_LENGTH int How many timesteps in the future to predict.
ENABLE_MULTIPLE_BACKTESTS bool Whether to enable multiple backtesting or not.
NUMBER_OF_SAMPLES int Number of samples for ancestral simulation
SKIP_TIMESERIES_WEIGHT_SCALING bool If True, we will avoid normalizing the weights.
RETURN_FRACTIONAL_FORECASTS None Use this to return fractional forecast values while prediction
LOSS_FUNCTION ForecastingLossFunction Loss function for training neural network.
INITIAL_LEARNING_RATE float Initial learning rate.
HANDLE_ZEROS_AS_MISSING_VALUES bool If True, handle zero values in demand as missing data.
FULL_DATA_RETRAINING bool Train models separately with all the data.
TEST_START str Limit training data to dates before the given test start.
ENABLE_CLUSTERING bool Enable clustering in forecasting.
ALLOW_TRAINING_WITH_SMALL_HISTORY None Allows training with fewer than 100 rows in the dataset
FORECAST_FREQUENCY ForecastingFrequency Forecast frequency.
SKIP_MISSING bool Make the RNN ignore missing entries rather instead of processing them.
NUM_BACKTESTING_WINDOWS int Total backtesting windows to use for the training.
ENABLE_FEATURE_SELECTION bool Enable feature selection.
BACKTESTING_WINDOW_STEP_SIZE int Use this step size to shift backtesting windows for model training.
SORT_OBJECTIVE ForecastingObjective Ranking scheme used to sort models on the metrics page.
TEST_SPLIT int Percent of dataset to use for test data. We support using a range between 5% to 20% of your dataset to use as test data.
UNDERPREDICTION_WEIGHT float Weight for underpredictions
TIMESERIES_WEIGHT_COLUMN str If set, we use the values in this column from timeseries data to assign time dependent item weights during training and evaluation.
TYPE_OF_SPLIT ForecastingDataSplitType Type of data splitting into train/test.
TEST_BY_ITEM bool Partition train/test data by item rather than time if true.
SMOOTH_HISTORY float Smooth (low pass filter) the timeseries.
L2_REGULARIZATION_FACTOR float L2 regularization factor.
CUSTOM_LOSS_FUNCTIONS List[str] Registered custom losses available for selection.
BATCH_SIZE ForecastingBatchSize Batch size.
USE_ITEM_WEIGHTS_IN_OBJECTIVE bool If True, we include weights from column set as "ITEM ATTRIBUTES WEIGHT COLUMN" in objective functions.
ITEM_ATTRIBUTES_WEIGHT_COLUMN str If set, we use the values in this column from item attributes data to assign weights to items during training and evaluation.
RECURRENT_UNITS int Number of units in each recurrent layer.
CONVOLUTION_FILTERS int Number of filters in each convolution.
CONVOLUTIONAL_LAYERS int Number of convolutional layers to stack on top of recurrent layers in network.
LOCAL_SCALING_MODE ForecastingLocalScaling Options to make NN inputs stationary in high dynamic range datasets.
FILL_MISSING_VALUES List[List[dict]] Strategy for filling in missing values.
RECURRENT_LAYERS int Number of recurrent layers to stack in network.
USE_CLIPPING bool Apply clipping to input data to stabilize the training.
PREDICTION_STEP_SIZE int Number of future periods to include in objective for each training sample.
DISABLE_NETWORKS_WITHOUT_ANALYTIC_QUANTILES bool Disable neural networks, which quantile functions do not have analytic expressions (e.g, mixture models)
TIMESERIES_LOSS_WEIGHT_COLUMN str Use value in this column to weight the loss while training.
MAX_SCALE_CONTEXT int Maximum context to use for local scaling.
SYMMETRIZE_QUANTILES bool Force symmetric quantiles (like in Gaussian distribution)
QUANTILES_EXTENSION_METHOD ForecastingQuanitlesExtensionMethod Quantile extension method
DATA_SPLIT_FEATURE_GROUP_TABLE_NAME str Specify the table name of the feature group to export training data with the fold column.
ENABLE_PADDING bool Pad series to the max_date of the dataset
TRAINING_POINT_OVERLAP float Amount of overlap to allow between training samples.
DATETIME_HOLIDAY_CALENDARS List[HolidayCalendars] Holiday calendars to augment training with.
OBJECTIVE ForecastingObjective Ranking scheme used to select final best model.
DROPOUT_RATE int Dropout percentage rate.
FORCE_PREDICTION_LENGTH int Force length of test window to be the same as prediction length.
USE_LOG_TRANSFORMS bool Apply logarithmic transformations to input data.
© Copyright