Template Overview

This is a basic tutorial for utilizing model templates and adding additional customization. After reading this tutorial, you will be able to

• Choose the desired template(s) to use.

• Customize the parameters in the templates.

• Perform grid search via multiple templates/parameters.

Model templates include parameters and their default values that are ready to use for forecast models. These templates can be run with the run_forecast_config function by defining the ForecastConfig class. The intention of using these templates is to easily combine with our pipeline forecast_pipeline and do grid search, cross-validation and evaluation. In this tutorial, we will go over available templates, their default values and how to customize the parameters.

A General View of the Templates and the ForecastConfig Class

Greykite supports a few forecast models, including SILVERKITE, PROPHET, and ARIMA. The SILVERKITE model has a high-level estimator named SimpleSilverkiteEstimator and a low-level estimator named SilverkiteEstimator. The PROPHET model has an estimator named ProphetEstimator. The ARIMA model has an estimator named AutoArimaEstimator.

The estimators only implement fit and predict methods. They are used in the forecast_pipeline function, for grid search, cross-validation, backtest, and forecast. To easily configure and run the forecast pipeline, pass ForecastConfig to the run_forecast_config function.

The ForecastConfig class takes the following parameters

• computation_param: the ComputationParam class, defines the computation parameters of the pipeline.

• coverage: the forecast interval coverage.

• evaluation_metric_param: the EvaluationMetricParam class, defines the metrics used to evaluate performance and choose the best model.

• evaluation_period_param: the EvaluationPeriodParam class, defines the cross-validation and train-test-split rules.

• forecast_horizon: the forecast horizon.

• metadata_param: the MetadataParam class, defines the metadata of the training data.

• model_components_param: the ModelComponentsParam class, defines the model parameters.

• model_template: the name or dataclass of the base model template, corresponding to one or some pre-defined ModelComponentsParam class(es).

Among these parameters, model_components_param and model_template define the parameters used in the estimators.

1. The full estimator parameters can be specified through the ModelComponentsParam class, as described in Greykite models and components.

2. We have pre-defined ModelComponentsParam classes that serve as default estimator parameters for different use cases. These pre-defined ModelComponentsParam classes have names.

3. You can specify in the model_template parameter a valid model template name. The function will automatically map the model_template input to the corresponding estimator and its default parameters.

4. To override the default values, you can create a ModelComponentsParam class with only the parameters you want to override, and pass it to the model_components_param parameter.

Note that you don’t have to specify all values in the ModelComponentsParam to override the defaults. If a parameter is not specified, the default value for the parameter specified by the model template name will be used. In the later sections we will go over the valid model_template and ModelComponentsParam for each of the three estimators. For details about how to configure the other parameters and how to use the run_forecast_config function, see Tune your first forecast model.

The three estimators accept different input for model_template and ModelComponentsParam. Below are the valid input types for the model_template parameter.

• High-level Silverkite template: for the high-level SimpleSilverkiteEstimator, we have model templates named "AUTO", "SILVERKITE", "SILVERKITE_HOURLY_1", "SILVERKITE_HOURLY_24", "SILVERKITE_HOURLY_168", "SILVERKITE_HOURLY_336", "SILVERKITE_DAILY_1", "SILVERKITE_DAILY_90", "SILVERKITE_WEEKLY", "SILVERKITE_MONTHLY", "SILVERKITE_EMPTY" and a set of generic naming following some rules. This type of model templates support list input for both model_template and model_components_param parameters. This type of model templates are most recommended for ease of use.

• Low-level Silverkite template: for the low-level SilverkiteEstimator, we have a model template named "SK". This template allows you to configure lower-level parameters in the Silverkite model. This template does not support list input.

• Prophet template: for the ProphetEstimator, we have a model template named "PROPHET". This template does not support list input.

To customize the default parameters in the templates, the ModelComponentsParam dataclass takes the following parameters

• growth: defines how the trend of the time series grows.

• seasonality: defines the seasonality components and orders.

• changepoints: defines when trend and/or seasonality should change, including automatic options.

• events: defines short term events and holidays.

• autoregression: defines the lags and aggregations for the past values.

• regressors: defines extra regressors.

• uncertainty: defines the forecast interval parameters.

• custom: defines parameters that do not belong to the other sections.

• hyperparameter_override: used to create overrides for the parameters specified above; useful in grid search.

The model’s tuning parameters are set according to the categories above. However, different estimators take different types of values for these categories. We will go over each of the three types of templates, their default values, and how to customize the ModelComponentsParam for them. For more general details, see Greykite models and components.

# Imports related libraries.
import pandas as pd

from greykite.common.constants import GrowthColEnum
from greykite.framework.templates.autogen.forecast_config import ForecastConfig
from greykite.framework.templates.autogen.forecast_config import ModelComponentsParam
from greykite.framework.templates.model_templates import ModelTemplateEnum
from greykite.framework.templates.simple_silverkite_template import SimpleSilverkiteTemplate

The High-level Templates in SILVERKITE

The high-level templates in SILVERKITE provides many good defaults that work under different scenarios. All templates in this section use SimpleSilverkiteEstimator. The two most basic templates are "SILVERKITE" and "SILVERKITE_EMPTY".

"SILVERKITE" is a template with automatic growth, seasonality, holidays, and interactions. It works best for hourly and daily frequencies. If you specify "SILVERKITE" as model_template, the following ModelComponentsParam class is used as default template values.

model_components_param_silverkite = ModelComponentsParam(
    seasonality={
        "auto_seasonality": False,
        "yearly_seasonality": "auto",
        "quarterly_seasonality": "auto",
        "monthly_seasonality": "auto",
        "weekly_seasonality": "auto",
        "daily_seasonality": "auto",
    },
    growth={
        "growth_term": GrowthColEnum.linear.name
    },
    events={
        "auto_holiday": False,
        "holidays_to_model_separately": "auto",
        "holiday_lookup_countries": "auto",
        "holiday_pre_num_days": 2,
        "holiday_post_num_days": 2,
        "holiday_pre_post_num_dict": None,
        "daily_event_df_dict": None,
    },
    changepoints={
        "auto_growth": False,
        "changepoints_dict": {
            "method": "auto",
            "yearly_seasonality_order": 15,
            "resample_freq": "3D",
            "regularization_strength": 0.6,
            "actual_changepoint_min_distance": "30D",
            "potential_changepoint_distance": "15D",
            "no_changepoint_distance_from_end": "90D"
        },
        "seasonality_changepoints_dict": None
    },
    autoregression={
        "autoreg_dict": "auto",
        "simulation_num": 10  # simulation is not triggered with ``autoreg_dict="auto"``
    },
    regressors={
        "regressor_cols": []
    },
    lagged_regressors={
        "lagged_regressor_dict": None
    },
    uncertainty={
        "uncertainty_dict": None
    },
    custom={
        "fit_algorithm_dict": {
            "fit_algorithm": "ridge",
            "fit_algorithm_params": None,
        },
        "feature_sets_enabled": "auto",  # "auto" based on data freq and size
        "max_daily_seas_interaction_order": 5,
        "max_weekly_seas_interaction_order": 2,
        "extra_pred_cols": [],
        "drop_pred_cols": None,
        "explicit_pred_cols": None,
        "min_admissible_value": None,
        "max_admissible_value": None,
        "normalize_method": "zero_to_one"
    }
)

To customize this template, create a ModelComponentsParam class like above with the parameters you would like to use to override the defaults, and feed it to the model_components_param parameter in ForecastConfig. For example

custom_model_components = ModelComponentsParam(
    seasonality={
        "yearly_seasonality": 15
    },
    custom={
        "fit_algorithm_dict": {
            "fit_algorithm": "ridge",
            "fit_algorithm_params": None
        }
    }
)

These two parameters can be put in the ForecastConfig class. The parameters used by the model will be those in the model_components_param_silverkite with "yearly_seasonality" and "fit_algorithm_dict" overridden by the custom parameters.

forecast_config = ForecastConfig(
    model_template=ModelTemplateEnum.SILVERKITE.name,
    model_components_param=custom_model_components
)

Detailed explanations for these parameters are in Greykite models and components. The following paragraphs briefly summarized what each parameter does.

The seasonality parameter recognizes the keys "yearly_seasonality", "quarterly_seasonality", "monthly_seasonality", "weekly_seasonality" and "daily_seasonality". Their values are the corresponding Fourier series values. For "SILVERKITE" template, the values are "auto" and their defaults are defined in SilverkiteSeasonalityEnum.

The growth parameter recognizes the key "growth_term", which describes the growth rate of the time series model. For "SILVERKITE" template, the value is "linear" and indicates linear growth.

The events parameter recognizes the keys "holidays_to_model_separately", "holiday_lookup_countries", "holiday_pre_num_days", "holiday_post_num_days", "holiday_pre_post_num_dict" and "daily_event_df_dict". More details can be found at Holidays and Events. For "SILVERKITE" template, it automatically looks up holidays in a holiday dictionary and model major holidays plus minus 2 days with separate indicators.

The changepoints parameter recognizes the keys "changepoints_dict" and "seasonality_changepoints_dict", which correspond to trend changepoints and seasonality changepoints. For more details of configuring these two parameters, see Changepoint Detection. For "SILVERKITE" template, both parameters are None, indicating that neither trend changepoints nor seasonality changepoints is included.

The autoregression parameter recognizes the key "autoreg_dict". You can specify lags and aggregated lags through the dictionary to trigger autoregressive terms. Specify the value as "auto" to automatically include recommended autoregressive terms for the data frequency and forecast horizon. Note that “auto” ensures that all the lag orders in the autoregressive terms are at least forecast_horizon. Therefore, simulation is not used in the prediction phase, hence simulation_num does not matter. More details can be found at Auto-regression. For "SILVERKITE" template, autoregression is not included.

The regressors parameter recognizes the key "regressor_cols", which takes a list of regressor column names. These regressor columns have to be included in the training df for both training and forecast periods. For more details about regressors, see Regressors. For "SILVERKITE" template, no regressors are included.

The lagged_regressors parameter recognizes the key "lagged_regressor_dict". It is a dictionary with keys being the regressor’s name, value being a dictionary that specifies the lag structure in a similar manner as "autoreg_dict". For more details about lagged regressors, see Regressors. For "SILVERKITE" template, no lagged regressors are included.

The uncertainty parameter recognizes the key "uncertainty_dict", which takes a dictionary to specify how forecast intervals are calculated. For more details about uncertainty, see Uncertainty. For "SILVERKITE" template, the default value is None. If coverage in ForecastConfig is not None, the template uses a default setting based on data frequency. We will see how to set coverage later.

The custom parameter recognizes specific keys for SILVERKITE type of templates that correspond to SimpleSilverkiteEstimator. These keys include

• "fit_algorithm_dict" takes a dictionary to specify what regression method is used to fit the time series. The default is the ridge regression in sklearn. For a detailed list of algorithms, see Algorithms.

• "feature_sets_enabled" defines the interaction terms to be included in the model. A list of pre-defined interaction terms can be found at Feature sets. The default is None, which automatically finds the proper interaction terms that fit the data frequency.

• "max_daily_seas_interaction_order" is the maximum order of Fourier series components in daily seasonality to be used in interactions. The default is 5.

• "max_weekly_seas_interaction_order" is the maximum order of Fourier series components in daily seasonality to be used in interactions. The default is 2.

• "extra_pred_cols" defines extra predictor column names. For details, see Extra predictors. The default is no extra predictors.

• "min_admissible_value" is the minimum admissible value in forecast. All values below this will be clipped at this value. The default is None.

• "max_admissible_value" is the maximum admissible value in forecast. All values above this will be clipped at this value. The default is None.

• "normalize_method" is the way of normalizing all columns in the feature matrix before fitting the model. Supported methods are “statistical”, “zero_to_one”, “minus_half_to_half” and “zero_at_origin”. The default is “zero_to_one”.

All default high-level SILVERKITE templates are defined through this framework. The "SILVERKITE_EMPTY" template is an empty template that does not include any component. If you provide ModelComponentsParam via model_components_param with "SILVERKITE_EMPTY", the final model parameter to be used will be exactly what you provided through ModelComponentsParam. It’s not like "SILVERKITE", where the values you do not provide within model_components_param will be filled with the defaults in "SILVERKITE". If you choose to use the "SILVERKITE_EMPTY" template but do not provide any ModelComponentsParam via model_components_param, the model will only fit the intercept term.

Auto model template

The Greykite library supports auto model template selection. If you are not sure which model template to use, you can try out the auto model template selection. Simply specify “AUTO” for model_template, the model will automatically select the most appropriate model template based on the data frequency, forecast horizon and evaluation configs.

forecast_config_auto = ForecastConfig(
    model_template=ModelTemplateEnum.AUTO.name
)

The auto model template selection will always map to a high-level Silverkite model template. That means you can specify model_components_param to override the model configurations in the same way as if you specified a high-level Silverkite model template.

The default model template is SILVERKITE, with the following exceptions:

• Hourly data

  • enough cv (cv splits >= 5 and cv evaluation >= 30 points)

    • Forecast horizon == 1: uses SILVERKITE_HOURLY_1.

    • Forecast horizon <= 48: uses SILVERKITE_HOURLY_24.

    • Forecast horizon <= 192: uses SILVERKITE_HOURLY_168.

    • Forecast horizon <= 504: uses SILVERKITE_HOURLY_336.

  • not enough cv

    • Forecast horizon == 1: uses SILVERKITE.

    • Forecast horizon <= 48: uses first template in SILVERKITE_HOURLY_24.

    • Forecast horizon <= 192: uses first template in SILVERKITE_HOURLY_168.

    • Forecast horizon <= 504: uses first template in SILVERKITE_HOURLY_336.

• Daily data

  • enough cv (cv splits >= 5 and cv evaluation >= 30 points)

    • Forecast horizon <= 7: uses SILVERKITE_DAILY_1.

    • Forecast horizon >= 90: uses SILVERKITE_DAILY_90.

  • not enough cv

    • Forecast horizon <= 7: uses SILVERKITE_DAILY_1_CONFIG_1.

    • Forecast horizon >= 90: uses first template in SILVERKITE_DAILY_90.

• Weekly data

  • enough cv (cv splits >= 5 and cv evaluation >= 30 points)

    • Uses SILVERKITE_WEEKLY.

  • not enough cv

    • Uses first template in SILVERKITE_WEEKLY.

• Monthly data

  • Uses SILVERKITE_MONTHLY.

Pre-defined Generic High-level SILVERKITE Templates

It can happen that you would like to customize the ModelComponentsParam but are not sure which values to set for each parameter. The high-level SILVERKITE template pre-defines sets of values for different components, indexed by human-readable language. This allows you to try sensible options for the components using a directive language. For example, “setting seasonality to normal and changepoints to light” is specified by sk.SEAS.value.NM and sk.CP.value.LT. This option provides rough tuning knobs before fine tuning the exact parameter values. This type of template name must be initialized through the SimpleSilverkiteTemplateOptions dataclass. You can choose a value for each component and assemble them as a template.

from greykite.framework.templates.simple_silverkite_template_config \
    import SimpleSilverkiteTemplateOptions as st
from greykite.framework.templates.simple_silverkite_template_config \
    import SILVERKITE_COMPONENT_KEYWORDS as sk
# The model template specifies
# hourly frequency, normal seasonality (no quarterly or monthly), linear growth, light trend changepoints,
# separate holidays with plus/minus 2 days, automatic feature sets, ridge regression, automatic autoregression,
# automatic max daily seasonality interaction order and automatic max weekly seasonality interaction order.
model_template = st(
    freq=sk.FREQ.value.HOURLY,
    seas=sk.SEAS.value.NM,
    gr=sk.GR.value.LINEAR,
    cp=sk.CP.value.LT,
    hol=sk.HOL.value.SP2,
    feaset=sk.FEASET.value.AUTO,
    algo=sk.ALGO.value.RIDGE,
    ar=sk.AR.value.AUTO,
    dsi=sk.DSI.value.AUTO,
    wsi=sk.WSI.value.AUTO
)

This option provides rough tuning knobs to intuitively try out different model component parameters. You can then fine tune the model using ModelComponentsParams directly. A complete list of the key-values are

• FREQ: the data frequency, can be “HOURLY”, “DAILY” or “WEEKLY”, default “DAILY”.

• SEAS: the seasonality, can be “LT”, “NM”, “HV”, “NONE”, “LTQM”, “NMQM” or “HVQM”, default “LT”. The “QM” versions include quarterly and monthly seasonality while the others do not.

• GR: the growth term, can be “LINEAR” or “NONE”, default “LINEAR”, corresponding to linear growth or constant growth.

• CP: the automatically detected trend change points, can be “NONE”, “LT”, “NM”, “HV”, default “NONE”.

• HOL: the holidays, can be “NONE”, “SP1”, “SP2”, “SP4” or “TG”, default “NONE”. The default configuration looks up popular holidays in a list of popular countries. The “SP{n}” values models major holidays with plus/minus n days around them separately, while “TG” models all holidays along with plus/minus 2 days together as one indicator.

• FEASET: the feature sets that defines the interaction terms, can be “AUTO”, “ON” or “OFF”, default “OFF”. “AUTO” choose the pre-defined interaction terms automatically, while “ON” and “OFF” includes or excludes all pre-defined interaction terms, respectively.

• ALGO: the algorithm used to fit the model, can be “LINEAR”, “RIDGE”, “SGD” or “LASSO”, default “LINEAR”. Ridge and Lasso use cross-validation to identify the tuning parameter, while “SGD” (stochastic gradient descent) implements L2 norm regularization with tuning parameter 0.001.

• AR: the autoregressive terms, can be “AUTO” or “OFF”, default “OFF”.

• DSI: the maximum daily seasonality order used for interaction in feature sets, can be “AUTO” or “OFF”, default “AUTO”.

• WSI: the maximum weekly seasonality order used for interaction in feature sets, can be “AUTO” or “OFF”, default “AUTO”.

Note that if you do not specify any parameter, the default value will be used: FREQ=DAILY, SEAS=LT, GR=LINEAR, CP=NONE, HOL=NONE, FEASET=OFF, ALGO=LINEAR, AR=OFF, DSI=AUTO, WSI=AUTO. To see how these keywords are converted to these model component params, see COMMON_MODELCOMPONENTPARAM_PARAMETERS. However, you can print the ModelComponentsParam class for a model template with the util function get_model_components_from_model_template.

sst = SimpleSilverkiteTemplate()
model_components = sst.get_model_components_from_model_template("SILVERKITE_EMPTY")
print(model_components[0])  # `model_components` is a list of length 1.

Out:

ModelComponentsParam(autoregression={'autoreg_dict': None, 'simulation_num': 10, 'fast_simulation': False}, changepoints={'auto_growth': False, 'changepoints_dict': None, 'seasonality_changepoints_dict': None}, custom={'feature_sets_enabled': False, 'fit_algorithm_dict': {'fit_algorithm': 'linear', 'fit_algorithm_params': None}, 'max_daily_seas_interaction_order': 0, 'max_weekly_seas_interaction_order': 0, 'extra_pred_cols': [], 'drop_pred_cols': None, 'explicit_pred_cols': None, 'min_admissible_value': None, 'max_admissible_value': None, 'regression_weight_col': None, 'normalize_method': 'zero_to_one', 'remove_intercept': False}, events={'auto_holiday': False, 'holidays_to_model_separately': [], 'holiday_lookup_countries': [], 'holiday_pre_num_days': 0, 'holiday_post_num_days': 0, 'holiday_pre_post_num_dict': None, 'daily_event_df_dict': None, 'auto_holiday_params': None, 'daily_event_neighbor_impact': None, 'daily_event_shifted_effect': None}, growth={'growth_term': None}, hyperparameter_override=None, regressors={'regressor_cols': []}, lagged_regressors={'lagged_regressor_dict': None}, seasonality={'auto_seasonality': False, 'yearly_seasonality': 0, 'quarterly_seasonality': 0, 'monthly_seasonality': 0, 'weekly_seasonality': 0, 'daily_seasonality': 0}, uncertainty={'uncertainty_dict': None})

You can also pass a dataclass.

model_components = sst.get_model_components_from_model_template(model_template)
print(model_components[0])  # `model_components` is a list of length 1.

Out:

ModelComponentsParam(autoregression={'autoreg_dict': 'auto', 'simulation_num': 10, 'fast_simulation': False}, changepoints={'auto_growth': False, 'changepoints_dict': {'method': 'auto', 'resample_freq': 'D', 'regularization_strength': 0.6, 'potential_changepoint_distance': '7D', 'no_changepoint_distance_from_end': '30D', 'yearly_seasonality_order': 15, 'yearly_seasonality_change_freq': None}, 'seasonality_changepoints_dict': None}, custom={'feature_sets_enabled': 'auto', 'fit_algorithm_dict': {'fit_algorithm': 'ridge', 'fit_algorithm_params': None}, 'max_daily_seas_interaction_order': 5, 'max_weekly_seas_interaction_order': 2, 'extra_pred_cols': [], 'drop_pred_cols': None, 'explicit_pred_cols': None, 'min_admissible_value': None, 'max_admissible_value': None, 'regression_weight_col': None, 'normalize_method': 'zero_to_one', 'remove_intercept': False}, events={'auto_holiday': False, 'holidays_to_model_separately': 'auto', 'holiday_lookup_countries': 'auto', 'holiday_pre_num_days': 2, 'holiday_post_num_days': 2, 'holiday_pre_post_num_dict': None, 'daily_event_df_dict': None, 'auto_holiday_params': None, 'daily_event_neighbor_impact': None, 'daily_event_shifted_effect': None}, growth={'growth_term': 'linear'}, hyperparameter_override=None, regressors={'regressor_cols': []}, lagged_regressors={'lagged_regressor_dict': None}, seasonality={'auto_seasonality': False, 'yearly_seasonality': 15, 'quarterly_seasonality': 0, 'monthly_seasonality': 0, 'weekly_seasonality': 4, 'daily_seasonality': 8}, uncertainty={'uncertainty_dict': None})

Provide a List of Templates

For the high-level "SILVERKITE" templates through the SimpleSilverkiteEstimator estimator, you are allowed to provide a list of model_template or/and a list of model_components_param. This option allows you to do grid search and compare over different templates/model component overrides at the same time.

For model_template, you can provide a list of any templates defined above. For example, you can do

model_templates_list = ["SILVERKITE", "SILVERKITE_EMPTY", model_template]

The get_model_components_from_model_template also takes a list as input.

model_components = sst.get_model_components_from_model_template(model_templates_list)
print(model_components)  # There are 3 elements.

Out:

[ModelComponentsParam(autoregression={'autoreg_dict': 'auto', 'simulation_num': 10, 'fast_simulation': False}, changepoints={'auto_growth': False, 'changepoints_dict': {'method': 'auto', 'yearly_seasonality_order': 15, 'resample_freq': '3D', 'regularization_strength': 0.6, 'actual_changepoint_min_distance': '30D', 'potential_changepoint_distance': '15D', 'no_changepoint_distance_from_end': '90D'}, 'seasonality_changepoints_dict': None}, custom={'fit_algorithm_dict': {'fit_algorithm': 'ridge', 'fit_algorithm_params': None}, 'feature_sets_enabled': 'auto', 'max_daily_seas_interaction_order': 5, 'max_weekly_seas_interaction_order': 2, 'extra_pred_cols': [], 'drop_pred_cols': None, 'explicit_pred_cols': None, 'min_admissible_value': None, 'max_admissible_value': None, 'regression_weight_col': None, 'normalize_method': 'zero_to_one', 'remove_intercept': False}, events={'auto_holiday': False, 'holidays_to_model_separately': 'auto', 'holiday_lookup_countries': 'auto', 'holiday_pre_num_days': 2, 'holiday_post_num_days': 2, 'holiday_pre_post_num_dict': None, 'daily_event_df_dict': None, 'auto_holiday_params': None, 'daily_event_neighbor_impact': None, 'daily_event_shifted_effect': None}, growth={'growth_term': 'linear'}, hyperparameter_override=None, regressors={'regressor_cols': []}, lagged_regressors={'lagged_regressor_dict': None}, seasonality={'auto_seasonality': False, 'yearly_seasonality': 'auto', 'quarterly_seasonality': 'auto', 'monthly_seasonality': 'auto', 'weekly_seasonality': 'auto', 'daily_seasonality': 'auto'}, uncertainty={'uncertainty_dict': None}), ModelComponentsParam(autoregression={'autoreg_dict': None, 'simulation_num': 10, 'fast_simulation': False}, changepoints={'auto_growth': False, 'changepoints_dict': None, 'seasonality_changepoints_dict': None}, custom={'feature_sets_enabled': False, 'fit_algorithm_dict': {'fit_algorithm': 'linear', 'fit_algorithm_params': None}, 'max_daily_seas_interaction_order': 0, 'max_weekly_seas_interaction_order': 0, 'extra_pred_cols': [], 'drop_pred_cols': None, 'explicit_pred_cols': None, 'min_admissible_value': None, 'max_admissible_value': None, 'regression_weight_col': None, 'normalize_method': 'zero_to_one', 'remove_intercept': False}, events={'auto_holiday': False, 'holidays_to_model_separately': [], 'holiday_lookup_countries': [], 'holiday_pre_num_days': 0, 'holiday_post_num_days': 0, 'holiday_pre_post_num_dict': None, 'daily_event_df_dict': None, 'auto_holiday_params': None, 'daily_event_neighbor_impact': None, 'daily_event_shifted_effect': None}, growth={'growth_term': None}, hyperparameter_override=None, regressors={'regressor_cols': []}, lagged_regressors={'lagged_regressor_dict': None}, seasonality={'auto_seasonality': False, 'yearly_seasonality': 0, 'quarterly_seasonality': 0, 'monthly_seasonality': 0, 'weekly_seasonality': 0, 'daily_seasonality': 0}, uncertainty={'uncertainty_dict': None}), ModelComponentsParam(autoregression={'autoreg_dict': 'auto', 'simulation_num': 10, 'fast_simulation': False}, changepoints={'auto_growth': False, 'changepoints_dict': {'method': 'auto', 'resample_freq': 'D', 'regularization_strength': 0.6, 'potential_changepoint_distance': '7D', 'no_changepoint_distance_from_end': '30D', 'yearly_seasonality_order': 15, 'yearly_seasonality_change_freq': None}, 'seasonality_changepoints_dict': None}, custom={'feature_sets_enabled': 'auto', 'fit_algorithm_dict': {'fit_algorithm': 'ridge', 'fit_algorithm_params': None}, 'max_daily_seas_interaction_order': 5, 'max_weekly_seas_interaction_order': 2, 'extra_pred_cols': [], 'drop_pred_cols': None, 'explicit_pred_cols': None, 'min_admissible_value': None, 'max_admissible_value': None, 'regression_weight_col': None, 'normalize_method': 'zero_to_one', 'remove_intercept': False}, events={'auto_holiday': False, 'holidays_to_model_separately': 'auto', 'holiday_lookup_countries': 'auto', 'holiday_pre_num_days': 2, 'holiday_post_num_days': 2, 'holiday_pre_post_num_dict': None, 'daily_event_df_dict': None, 'auto_holiday_params': None, 'daily_event_neighbor_impact': None, 'daily_event_shifted_effect': None}, growth={'growth_term': 'linear'}, hyperparameter_override=None, regressors={'regressor_cols': []}, lagged_regressors={'lagged_regressor_dict': None}, seasonality={'auto_seasonality': False, 'yearly_seasonality': 15, 'quarterly_seasonality': 0, 'monthly_seasonality': 0, 'weekly_seasonality': 4, 'daily_seasonality': 8}, uncertainty={'uncertainty_dict': None})]

For model_components_param, you can also create a list of ModelComponentsParam classes to override the base templates. Each single ModelComponentsParam is used to override each single base template. Therefore, if you provide a list of 4 ModelComponentsParam via model_components_param and the list of 3 base templates above via model_template, a total of 12 different sets of model parameters is expected. However, only unique sets of parameters will be kept.

There are also pre-defined model templates that are defined through lists. The "SILVERKITE_DAILY_1" is a pre-tuned model template on daily data with 1 day forecast horizon. It is defined as a list of 3 model configs. The "SILVERKITE_DAILY_90" is a pre-tuned model template on daily data with 90 day’s forecast horizon. It is defined as a list of 4 model configs. The "SILVERKITE_WEEKLY" is a pre-tuned model template on weekly data. It is defined as a list of 4 model configs. The "SILVERKITE_HOURLY_1", "SILVERKITE_HOURLY_24", "SILVERKITE_HOURLY_168", "SILVERKITE_HOURLY_336" are pre-tuned model templates on hourly data with horizons 1 hour, 1 day, 1 week and 2 weeks, respectively. They are defined as a list of 3 (for "SILVERKITE_HOURLY_1") or 4 (for others) model configs.

You are also allowed to put these names in the model_template list, for example

model_templates_list2 = ["SILVERKITE_DAILY_90", model_template]

This corresponds to 5 single base templates. Whenever you specify multiple sets of parameters (list of templates, list of model components, etc.), it’s best to have a sufficient number of cross validation folds so that the model does not pick a biased set of parameters.

The Low-level Templates in SILVERKITE

There is a pre-defined low-level template named "SK" that takes low-level parameters and uses SilverkiteEstimator.

The attributes in ModelComponentsParam are the same as in "SILVERKITE" but they take different types of inputs.

model_components_param_sk = ModelComponentsParam(
    growth={
    },  # growth does not accept any parameters, pass growth term via `extra_pred_cols` instead.
    seasonality={
        "fs_components_df": [pd.DataFrame({
            "name": ["tod", "tow", "tom", "toq", "toy"],
            "period": [24.0, 7.0, 1.0, 1.0, 1.0],
            "order": [3, 3, 1, 1, 5],
            "seas_names": ["daily", "weekly", "monthly", "quarterly", "yearly"]})],
    },
    changepoints={
        "changepoints_dict": [None],
        "seasonality_changepoints_dict": [None]
    },
    events={
        "daily_event_df_dict": [None]
    },
    autoregression={
        "autoreg_dict": [None]
    },
    # `regressors` does not take any input for SK template,
    # please specify regressor columns in ``custom.extra_pred_cols`` below.
    regressors={},
    uncertainty={
        "uncertainty_dict": [None]
    },
    custom={
        "fit_algorithm_dict": {
            "fit_algorithm": "ridge",
            "fit_algorithm_params": None,
        },
        "extra_pred_cols": ["ct1"],  # extra predictor columns and regressors
        "min_admissible_value": [None],
        "max_admissible_value": [None],
    }
)

The growth parameter, the dictionary should be empty. The growth term’s name is specified via extra_pred_cols in custom. The default growth term is "ct1", which corresponds to linear growth.

The seasonality parameter, it recognizes the key "fs_components_df", which is a pandas dataframe that specifies the fourier series generation information. For more information, see SilverkiteEstimator. For "SK" template, the default includes daily, weekly, monthly, quarterly and yearly seasonality with orders 3, 3, 1, 1, 5, respectively.

The changepoints parameter recognizes the keys "changepoints_dict" and "seasonality_changepoints_dict". Each of the two keys takes a parameter dictionary that corresponds to trend changepoints and seasonality changepoints. For more details of configuring these two parameters, see Changepoint Detection. For "SK" template, both parameters are None, indicating that neither trend changepoints nor seasonality changepoints is included.

The events parameter recognizes the key "daily_event_df_dict". Specify any events or holidays through the “daily_event_df_dict”. The usage is the same as this parameter in SILVERKITE. For "SK" template, the default is no daily events (holidays).

The autoregression parameter recognizes the key "autoreg_dict". You can specify lags and aggregated lags through the dictionary to trigger autoregressive terms. Specify the value as "auto" to automatically include the proper order of lags. For "SK" template, autoregression is not included.

The regressors parameter does not recognize any key. If you have external regressors you would like to include, you need to include them in custom.extra_pred_cols instead of regressors.regressor_cols. For more details about regressors, see Regressors. For "SK" template, no regressors are included.

The uncertainty parameter recognizes the key "uncertainty_dict", which takes a dictionary to specify how forecast intervals are calculated. For more details about uncertainty, see Uncertainty. For "SK" template, the default value is None. If coverage in ForecastConfig is not None, it will automatically finds the most proper conditional residual to compute forecast intervals. We will see how to set coverage later.

The custom parameter recognizes specific keys for "SK" type of template that correspond to SilverkiteEstimator. These keys include

• "fit_algorithm_dict" takes a dictionary to specify what regression method is used to fit the time series. The default is the linear regression in sklearn. For a detailed list of algorithms, see Algorithms.

• "extra_pred_cols" defines extra predictor column names. It accepts any valid patsy model formula term. Every column name needs to be either generated by build_silverkite_features or included in the data df. For details, see Extra predictors. The default is ["ct1"], which is the linear growth term. If you have external regressors, this is the place to include the regressor names.

• "min_admissible_value" is the minimum admissible value in forecast. All values below this will be clipped at this value. The default is None.

• "max_admissible_value" is the maximum admissible value in forecast. All values above this will be clipped at this value. The default is None.

A major difference between the high-level and low-level interfaces is that the lower-level interface does not have pre-defined holidays or feature sets (interaction terms), and takes more customizable seasonality information. Note that "SK" is the only low-level template in SILVERKITE estimators, and does not support a list of model_template or model_components_param.

The "PROPHET" Template

The "PROPHET" template uses ProphetEstimator, which is a wrapper for the Prophet model.

The attributes in ModelComponentsParam are the same as in "SILVERKITE" but they take different types of inputs.

model_components_param_prophet = ModelComponentsParam(
    growth={
        "growth_term": ["linear"]
    },
    seasonality={
        "seasonality_mode": ["additive"],
        "seasonality_prior_scale": [10.0],
        "yearly_seasonality": ['auto'],
        "weekly_seasonality": ['auto'],
        "daily_seasonality": ['auto'],
        "add_seasonality_dict": [None]
    },
    changepoints={
        "changepoint_prior_scale": [0.05],
        "changepoints": [None],
        "n_changepoints": [25],
        "changepoint_range": [0.8]
    },
    events={
        "holiday_lookup_countries": "auto",
        "holiday_pre_num_days": [2],
        "holiday_post_num_days": [2],
        "start_year": 2015,
        "end_year": 2030,
        "holidays_prior_scale": [10.0]
    },
    regressors={
        "add_regressor_dict": [None]
    },
    uncertainty={
        "mcmc_samples": [0],
        "uncertainty_samples": [1000]
    }
)

The growth parameter recognizes the key "growth_term", which describes the growth rate of the time series model. For "PROPHET" template, the value indicates linear growth.

The seasonality parameter recognizes the keys "seasonality_mode", "seasonality_prior_scale", "yearly_seasonality", "weekly_seasonality", "daily_seasonality" and "add_seasonality_dict". For "PROPHET" template, the seasonality model is “additive” with prior scale 10 and automatic components.

The changepoints parameter recognizes the keys "changepoint_prior_scale", "changepoints", "n_changepoints" and "changepoint_range". The Prophet model supports trend changepoints only. For "PROPHET" template, it puts 25 potential trend changepoints uniformly over the first 80% data and use regularization with prior scale 0.05.

The events parameter recognizes the keys "holiday_lookup_countries", "holiday_pre_num_days", "holiday_post_num_days", "start_year", "end_year" and "holidays_prior_scale". The algorithm automatically looks up holidays in "holiday_lookup_countries". For "PROPHET" template, it automatically looks up holidays between 2015 and 2030 with their plus/minus 2 days. The holiday prior scale is 10.

The Prophet model does not support autoregression, so the autoregression value should be empty.

The regressors parameter recognizes the key "add_regressor_dict". For more details about regressors, see Regressors. For "PROPHET" template, no regressors are included.

The uncertainty parameter recognizes the key "mcmc_samples" and "uncertainty_samples". For more details about uncertainty, see Uncertainty. For "PROPHET" template, the default value is to sample 1000 uncertainty samples.

The Prophet model does not have any specific value in the custom parameter.

Extra Notes

• All templates take the hyperparameter_override key in their ModelComponentsParam class, which is used to define extra grid search options. For details, see Grid Search.

• To specify a string as a template name, it is recommended to use the ModelTemplateEnum to avoid typos. For example,

silverkite_template = ModelTemplateEnum.SILVERKITE.name
silverkite_templates = [
    ModelTemplateEnum.SILVERKITE_EMPTY.name,
    ModelTemplateEnum.SILVERKITE_DAILY_90.name
]
prophet_template = ModelTemplateEnum.PROPHET.name