Custom Parameters¶
Parameters unique to the Silverkite algorithm are included under model_components.custom
.
Tip
Refer to this section for customizing Silverkite model.
model_components.custom
is currently not supported for Prophet.
custom = dict(
fit_algorithm_dict=fit_algorithm_dict,
feature_sets_enabled=feature_sets_enabled,
max_daily_seas_interaction_order=max_daily_seas_interaction_order,
max_weekly_seas_interaction_order=max_weekly_seas_interaction_order,
extra_pred_cols=extra_pred_cols,
drop_pred_cols=drop_pred_cols,
explicit_pred_cols=explicit_pred_cols,
min_admissible_value=min_admissible_value,
max_admissible_value=max_admissible_value,
)
Fit algorithm¶
Silverkite abstracts feature generation from model fitting. First, Silverkite generates features for forecasting using time-series domain knowledge. Then, any standard algorithm can be applied for fitting.
Silverkite supports many fit algorithms from scikit-learn
and statsmodels
, as
well as our own fast quantile regression implementation.
Here are a few important ones:
fit_algorithm |
implementation |
notes |
---|---|---|
ridge |
Default alpha = np.logspace(-5, 5, 30) |
|
lasso |
||
elastic_net |
||
linear |
||
quantile_regression |
Default quantile=0.5, alpha=0 |
Note
For mean forecasts, "ridge"
or "linear"
are recommended.
"ridge"
is preferred when there are many features.
For quantile prediction, "quantile_regression"
is recommended.
"sgd"
is often unstable. Tree-based methods ("rf"
, "gradient_boosting"
)
are generally not good at projecting growth into the future.
Here is the full list of supported fit algorithms. You can select the algorithm via
fit_algorithm
and set its initialization parameters via fit_algorithm_params
.
fit_algorithm_dict : `dict` or a list of such values for grid search
How to fit the model. A dictionary with the following optional keys.
``"fit_algorithm"`` : `str`, default "ridge"
The type of model used in fitting
(implemented by sklearn and statsmodels).
Available models are:
- "statsmodels_ols" : `statsmodels.regression.linear_model.OLS`
- "statsmodels_wls" : `statsmodels.regression.linear_model.WLS`
- "statsmodels_gls" : `statsmodels.regression.linear_model.GLS`
- "statsmodels_glm" : `statsmodels.genmod.generalized_linear_model.GLM`
- "linear" : `statsmodels.regression.linear_model.OLS`
- "elastic_net" : `sklearn.linear_model.ElasticNetCV`
- "ridge" : `sklearn.linear_model.RidgeCV`
- "lasso" : `sklearn.linear_model.LassoCV`
- "sgd" : `sklearn.linear_model.SGDRegressor`
- "lars" : `sklearn.linear_model.LarsCV`
- "lasso_lars" : `sklearn.linear_model.LassoLarsCV`
- "rf" : `sklearn.ensemble.RandomForestRegressor`
- "gradient_boosting" : `sklearn.ensemble.GradientBoostingRegressor`
- "quantile_regression" : `greykite.algo.common.l1_quantile_regression.QuantileRegression`
See `~greykite.algo.common.ml_models.fit_model_via_design_matrix`
for the sklearn and statsmodels classes that implement these methods, and their default parameters.
"linear" is the same as "statsmodels_ols", because `statsmodels.regression.linear_model.OLS`
is more stable than `sklearn.linear_model.LinearRegression`.
``"fit_algorithm_params"`` : `dict` or None, default None
Parameters passed to the requested fit algorithm.
If None, uses the defaults in `~greykite.algo.common.ml_models.fit_model_via_design_matrix`.
Examples:
custom = dict(
fit_algorithm_dict=dict(
fit_algorithm="linear"
)
)
custom = dict(
fit_algorithm_dict=dict(
fit_algorithm="ridge",
fit_algorithm_params={
"alphas": np.logspace(-5, 5, 30)
}
)
)
custom = dict(
fit_algorithm_dict=dict(
fit_algorithm="quantile_regression",
fit_algorithm_params={
"quantile": 0.5,
"alpha": 1,
}
)
)
# Grid search is possible
custom = dict(
fit_algorithm_dict=[
dict(
fit_algorithm="linear"
),
dict(
fit_algorithm="ridge",
fit_algorithm_params={
"alphas": np.logspace(-5, 5, 30)
}
),
]
)
Interactions¶
You can include interactions via the feature_sets_enabled
parameter. Setting feature_sets_enabled="auto"
adds interactions appropriate to the data frequency and amount of training history.
The fourier order of seasonality interaction terms can be capped by
max_daily_seas_interaction_order
and max_weekly_seas_interaction_order
for daily and weekly seasonality, respectively.
Options (defaults shown for SILVERKITE
template):
feature_sets_enabled : `dict` [`str`, `bool` or "auto" or None] or `bool` or "auto" or None; or a list of such values for grid search, default "auto"
Whether to include interaction terms and categorical variables to increase model flexibility.
If a `dict`, boolean values indicate whether include various sets of features in the model.
The following keys are recognized
(from `~greykite.algo.forecast.silverkite.constants.silverkite_column.SilverkiteColumn`):
``"COLS_HOUR_OF_WEEK"`` : `str`
Constant hour of week effect
``"COLS_WEEKEND_SEAS"`` : `str`
Daily seasonality interaction with is_weekend
``"COLS_DAY_OF_WEEK_SEAS"`` : `str`
Daily seasonality interaction with day of week
``"COLS_TREND_DAILY_SEAS"`` : `str`
Allow daily seasonality to change over time by is_weekend
``"COLS_EVENT_SEAS"`` : `str`
Allow sub-daily event effects
``"COLS_EVENT_WEEKEND_SEAS"`` : `str`
Allow sub-daily event effect to interact with is_weekend
``"COLS_DAY_OF_WEEK"`` : `str`
Constant day of week effect
``"COLS_TREND_WEEKEND"`` : `str`
Allow trend (growth, changepoints) to interact with is_weekend
``"COLS_TREND_DAY_OF_WEEK"`` : `str`
Allow trend to interact with day of week
``"COLS_TREND_WEEKLY_SEAS"`` : `str`
Allow weekly seasonality to change over time
The following dictionary values are recognized:
- True: include the feature set in the model
- False: do not include the feature set in the model
- None: do not include the feature set in the model
- "auto" or not provided: use the default setting based on data frequency and size
If not a `dict`:
- if a boolean, equivalent to a dictionary with all values set to the boolean.
- if None, equivalent to a dictionary with all values set to False.
- if "auto", equivalent to a dictionary with all values set to "auto".
max_daily_seas_interaction_order : `int` or None or a list of such values for grid search, default 5
Max fourier order to use for interactions with daily seasonality
(COLS_EVENT_SEAS, COLS_EVENT_WEEKEND_SEAS, COLS_WEEKEND_SEAS, COLS_DAY_OF_WEEK_SEAS, COLS_TREND_DAILY_SEAS).
Model includes interactions terms specified by ``feature_sets_enabled``
up to the order limited by this value and the available order from ``seasonality``.
max_weekly_seas_interaction_order : `int` or None or a list of such values for grid search, default 2
Max fourier order to use for interactions with weekly seasonality (COLS_TREND_WEEKLY_SEAS).
Model includes interactions terms specified by ``feature_sets_enabled``
up to the order limited by this value and the available order from ``seasonality``.
feature set |
max freq |
when to use |
(human-readable) formula |
---|---|---|---|
COLS_HOUR_OF_WEEK |
hourly |
hour of week effect to help daily seasonality model |
hour_of_week |
COLS_WEEKEND_SEAS |
hourly |
weekend has a different daily seasonality pattern |
is_weekend:daily_seas |
COLS_DAY_OF_WEEK_SEAS |
hourly |
each day has a different daily seasonality pattern |
day_of_week:daily_seas |
COLS_TREND_DAILY_SEAS |
hourly |
daily seasonality pattern changes over time, by is_weekend |
trend:is_weekend:daily_seas |
COLS_EVENT_SEAS |
hourly |
events have a different daily seasonality pattern |
event:daily_seas |
COLS_EVENT_WEEKEND_SEAS |
hourly |
events have a different daily event seasonality pattern, by is_weekend |
event:is_weekend:daily_seas |
COLS_DAY_OF_WEEK |
daily |
day of week effect to help weekly seasonality model |
day_of_week |
COLS_TREND_WEEKEND |
daily |
growth rate differs for weekend/weekday |
trend:is_weekend |
COLS_TREND_DAY_OF_WEEK |
daily |
growth rate differs by day of week |
trend:day_of_week |
COLS_TREND_WEEKLY_SEAS |
daily |
weekly seasonality pattern changes over time |
trend:weekly_seas |
Examples:
from greykite.algo.forecast.silverkite.constants.silverkite_column import SilverkiteColumn
# Uses the default for all feature sets based on data frequency and size (training data)
custom = dict(
feature_sets_enabled="auto"
)
# Turns off all feature sets
custom = dict(
feature_sets_enabled=False
)
custom = dict(
feature_sets_enabled=None # same as False (prefer False to be explicit)
)
# Turns on all feature sets
# (Not recommended. Use "auto" to enable all relevant feature sets, or
# enable specific feature sets as shown below.)
custom = dict(
feature_sets_enabled=True
)
# Turns on specific feature sets
custom = dict(
feature_sets_enabled={
# Not included in the model.
SilverkiteColumn.COLS_HOUR_OF_WEEK: False,
SilverkiteColumn.COLS_WEEKEND_SEAS: False,
SilverkiteColumn.COLS_DAY_OF_WEEK_SEAS: False,
SilverkiteColumn.COLS_TREND_DAILY_SEAS: False,
SilverkiteColumn.COLS_EVENT_SEAS: False,
# None is the same as False (prefer False to be explicit)
SilverkiteColumn.COLS_EVENT_WEEKEND_SEAS: None,
# Included in the model.
SilverkiteColumn.COLS_DAY_OF_WEEK: True,
SilverkiteColumn.COLS_TREND_WEEKEND: True,
# Auto uses the default based on data frequency and size.
SilverkiteColumn.COLS_TREND_DAY_OF_WEEK: "auto",
# Omitted key is treated the same as "auto".
# SilverkiteColumn.COLS_TREND_WEEKLY_SEAS: "auto"
},
# Allows up to fourier order 2 for weekly seasonality interactions
max_weekly_seas_interaction_order=2
)
# Turns on a few feature sets relevant for hourly data
custom = dict(
feature_sets_enabled={
SilverkiteColumn.COLS_HOUR_OF_WEEK: False,
SilverkiteColumn.COLS_WEEKEND_SEAS: True,
SilverkiteColumn.COLS_DAY_OF_WEEK_SEAS: True,
SilverkiteColumn.COLS_TREND_DAILY_SEAS: True,
SilverkiteColumn.COLS_EVENT_SEAS: False, # unnecessary when COLS_EVENT_WEEKEND_SEAS is used
SilverkiteColumn.COLS_EVENT_WEEKEND_SEAS: True,
SilverkiteColumn.COLS_DAY_OF_WEEK: False,
SilverkiteColumn.COLS_TREND_WEEKEND: True,
SilverkiteColumn.COLS_TREND_DAY_OF_WEEK: True,
SilverkiteColumn.COLS_TREND_WEEKLY_SEAS: True
},
# Allows up to fourier order 2 for daily/weekly seasonality interactions
max_daily_seas_interaction_order=2,
max_weekly_seas_interaction_order=2
)
To check which features sets are enabled by default for your dataset, call
get_feature_sets_enabled
.
The parameter
num_days
is the number of days in your input timeseries (historical data for training, without future dates for regressors). It does not need to be exact.
from greykite.algo.forecast.silverkite.forecast_simple_silverkite import SimpleSilverkiteForecast
from greykite.common.enums import SimpleTimeFrequencyEnum
silverkite = SimpleSilverkiteForecast()
# 60 days of hourly data
silverkite._SimpleSilverkiteForecast__get_feature_sets_enabled(
simple_freq=SimpleTimeFrequencyEnum.HOUR.name,
num_days=60,
feature_sets_enabled="auto")
# 60 days of daily data
silverkite._SimpleSilverkiteForecast__get_feature_sets_enabled(
simple_freq=SimpleTimeFrequencyEnum.DAY.name,
num_days=60,
feature_sets_enabled="auto")
# 30 weeks of weekly data
silverkite._SimpleSilverkiteForecast__get_feature_sets_enabled(
simple_freq=SimpleTimeFrequencyEnum.WEEK.name,
num_days=7*30,
feature_sets_enabled="auto")
# 30 months of monthly data
silverkite._SimpleSilverkiteForecast__get_feature_sets_enabled(
simple_freq=SimpleTimeFrequencyEnum.MONTH.name,
num_days=30*30,
feature_sets_enabled="auto")
# 20 quarters of quarterly data
silverkite._SimpleSilverkiteForecast__get_feature_sets_enabled(
simple_freq=SimpleTimeFrequencyEnum.QUARTER.name,
num_days=90*20,
feature_sets_enabled="auto")
# 12 years of yearly data
silverkite._SimpleSilverkiteForecast__get_feature_sets_enabled(
simple_freq=SimpleTimeFrequencyEnum.YEAR.name,
num_days=365*12,
feature_sets_enabled="auto")
Specify model terms¶
For even finer control than feature_sets_enabled
, you can specify additional model terms
via extra_pred_cols
. Any valid patsy model formula term is accepted. You need to know how
columns are internally coded to use this function. Most users will not need this option.
Note
While it’s possible to add many terms in the model, a high degree of complexity may not generalize well into the forecast period and should not be necessary for most forecasts.
extra_pred_cols : `list` [`str`] or None or a list of such values for grid search, default None
Names of extra predictor columns to pass to ``forecast_silverkite``.
The standard interactions can be controlled via ``feature_sets_enabled`` parameter.
Accepts any valid patsy model formula term. Can be used to model complex interactions
of time features, events, seasonality, changepoints, regressors. Columns should be
generated by ``build_silverkite_features`` or included with input data.
These are added to any features already included by ``feature_sets_enabled`` and
terms specified by ``model``.
Example:
import greykite.common.constants as cst
from greykite.framework.utils.result_summary import patsy_categorical_term
# Provides all holidays of interest
holiday_names = ["Christmas Day", "Thanksgiving", "Labor Day"]
# Adds day-of-holiday interaction with is_weekend
# (does not add the interaction on days offset from the holiday)
extra_pred_cols = []
event_levels = [cst.EVENT_DEFAULT, cst.EVENT_INDICATOR]
for holiday_name in holiday_names:
holiday_term = patsy_categorical_term(
term=f"{cst.EVENT_PREFIX}_{holiday_name}", # holiday column name
levels=event_levels) # levels for holiday categorical variable
extra_pred_cols += [f"is_weekend:{holiday_term}"]
# Tells the model to include these parameters
custom = dict(extra_pred_cols=extra_pred_cols)
Note
Contact us if you’re using extra_pred_cols
. If your model terms are useful for others, we can add
them to feature_sets_enabled
.
Similarly, you may specify terms to exclude via "drop_pred_cols"
:
drop_pred_cols : `list` [`str`] or None, default None
Names of predictor columns to be dropped from the final model.
Ignored if None.
To directly specify all the terms used in the final model, use "explicit_pred_cols"
:
explicit_pred_cols : `list` [`str`] or None, default None
Names of the explicit predictor columns which will be
the only variables in the final model. Note that this overwrites
the generated predictors in the model and may include new
terms not appearing in the predictors (e.g. interaction terms).
Ignored if None.
Forecast limits¶
You may prevent the forecast from going above or below pre-set values
via min_admissible_value
and max_admissible_value
.
This can be useful, for example, if you are forecasting a non-negative metric.
min_admissible_value : `float` or `double` or `int` or None, default None
The lowest admissible value for the forecasts and prediction
intervals. Any value below this will be mapped back to this value.
If None, there is no lower bound.
max_admissible_value : `float` or `double` or `int`, default None
The highest admissible value for the forecasts and prediction
intervals. Any value above this will be mapped back to this value.
If None, there is no upper bound.
Examples:
# enforce non-negative forecast
custom = dict(
min_admissible_value=0
)
# specifies an acceptable range
custom = dict(
min_admissible_value=1e3
max_admissible_value=1e6
)
Normalization¶
It can be helpful to normalize features, especially when features have different magnitudes and regularization is used.
normalize_method : `str` or None, default None
The normalization method for the feature matrix.
Available values are "statistical", "zero_to_one" and "minus_half_to_half".
Examples:
custom = dict(
normalize_method="statistical"
)
custom = dict(
normalize_method="zero_to_one"
)
The statistical
method removes the “mean” and divides by “std” for each column.
The zero_to_one
method removes the “min” and divides by the “max - min”
The ``minus_half_to_half``
method removes the “(min + max)/2” and divides by the “max - min”
for each column. For details, see normalize_df
.