def test_train_end_date_without_regressors():
"""Tests make_future_dataframe and train_end_date without regressors"""
ts = UnivariateTimeSeries()
df = pd.DataFrame({
TIME_COL: [
dt(2018, 1, 1, 3, 0, 0),
dt(2018, 1, 1, 4, 0, 0),
dt(2018, 1, 1, 5, 0, 0),
dt(2018, 1, 1, 6, 0, 0),
dt(2018, 1, 1, 7, 0, 0)
],
VALUE_COL: [1, None, 3, None, None],
})
# train_end_date later than last date in df
with pytest.warns(UserWarning) as record:
train_end_date = dt(2018, 1, 1, 8, 0, 0)
ts.load_data(df, TIME_COL, VALUE_COL, train_end_date=train_end_date)
assert f"Input timestamp for the parameter 'train_end_date' " \
f"({train_end_date}) either exceeds the last available timestamp or" \
f"{VALUE_COL} column of the provided TimeSeries contains null " \
f"values at the end. Setting 'train_end_date' to the last timestamp with a " \
f"non-null value ({ts.train_end_date})." in record[0].message.args[0]
assert ts.train_end_date == dt(2018, 1, 1, 5, 0, 0)
result = ts.make_future_dataframe(periods=10, include_history=True)
expected = pd.DataFrame({
TIME_COL:
pd.date_range(start=dt(2018, 1, 1, 3, 0, 0), periods=13, freq="H"),
VALUE_COL:
np.concatenate((ts.fit_y, np.repeat(np.nan, 10)))
})
expected.index = expected[TIME_COL]
expected.index.name = None
assert_frame_equal(result, expected)
def test_plot_grouping_evaluation():
"""Tests plot_grouping_evaluation function"""
df = generate_df_for_tests(freq="D", periods=20)["df"]
df.rename(columns={
TIME_COL: "custom_time_column",
VALUE_COL: "custom_value_column"
},
inplace=True)
ts = UnivariateTimeSeries()
ts.load_data(df,
time_col="custom_time_column",
value_col="custom_value_column")
# groupby_time_feature
fig = ts.plot_grouping_evaluation(aggregation_func=np.mean,
aggregation_func_name="mean",
groupby_time_feature="dow")
assert fig.data[0].name == f"mean of {VALUE_COL}"
assert fig.layout.xaxis.title.text == "dow"
assert fig.layout.yaxis.title.text == f"mean of {VALUE_COL}"
assert fig.layout.title.text == f"mean of {VALUE_COL} vs dow"
assert fig.data[0].x.shape[0] == 7
# groupby_sliding_window_size
fig = ts.plot_grouping_evaluation(
aggregation_func=np.max,
aggregation_func_name="max",
groupby_sliding_window_size=7
) # there are 20 training points, so this creates groups of size (6, 7, 7)
assert fig.data[0].name == f"max of {VALUE_COL}"
assert fig.layout.xaxis.title.text == f"{TIME_COL}_downsample"
assert fig.layout.yaxis.title.text == f"max of {VALUE_COL}"
assert fig.layout.title.text == f"max of {VALUE_COL} vs {TIME_COL}_downsample"
assert fig.data[0].x.shape[0] == 3
# groupby_custom_column
custom_groups = pd.Series(["g1", "g2", "g3", "g4", "g5"],
name="custom_groups").repeat(4)
fig = ts.plot_grouping_evaluation(aggregation_func=np.min,
aggregation_func_name="min",
groupby_custom_column=custom_groups)
assert fig.data[0].name == f"min of {VALUE_COL}"
assert fig.layout.xaxis.title.text == "custom_groups"
assert fig.layout.yaxis.title.text == f"min of {VALUE_COL}"
assert fig.layout.title.text == f"min of {VALUE_COL} vs custom_groups"
assert fig.data[0].x.shape[0] == 5
# custom xlabel, ylabel and title
fig = ts.plot_grouping_evaluation(aggregation_func=np.mean,
aggregation_func_name="mean",
groupby_time_feature="dow",
xlabel="Day of Week",
ylabel="Average of y",
title="Average of y by Day of week")
assert fig.layout.xaxis.title.text == "Day of Week"
assert fig.layout.yaxis.title.text == "Average of y"
assert fig.layout.title.text == "Average of y by Day of week"
def test_plot():
"""Checks plot function"""
# Plots with `color`
ts = UnivariateTimeSeries()
df = pd.DataFrame({
TIME_COL: [
dt(2018, 1, 1, 3, 0, 0),
dt(2018, 1, 1, 4, 0, 0),
dt(2018, 1, 1, 5, 0, 0)
],
VALUE_COL: [1, 2, 3]
})
ts.load_data(df, TIME_COL, VALUE_COL)
fig = ts.plot(color="green")
assert len(fig.data) == 1
assert fig.data[0].line.color == "green"
with pytest.raises(
ValueError,
match=
"There is no `anomaly_info` to show. `show_anomaly_adjustment` must be False."
):
ts.plot(show_anomaly_adjustment=True)
# Plots with `show_anomaly_adjustment`
dl = DataLoaderTS()
df = dl.load_beijing_pm()
value_col = "pm"
# Masks up to 2011-02-04-03, and adds 100.0 to the rest
anomaly_df = pd.DataFrame({
START_DATE_COL: ["2010-01-01-00", "2011-02-04-03"],
END_DATE_COL: ["2011-02-04-03", "2014-12-31-23"],
ADJUSTMENT_DELTA_COL: [np.nan, 100.0],
METRIC_COL: [value_col, value_col]
})
anomaly_info = {
"value_col": value_col,
"anomaly_df": anomaly_df,
"start_date_col": START_DATE_COL,
"end_date_col": END_DATE_COL,
"adjustment_delta_col": ADJUSTMENT_DELTA_COL,
"filter_by_dict": {
METRIC_COL: value_col
},
"adjustment_method": "add"
}
ts = UnivariateTimeSeries()
ts.load_data(df=df, value_col="pm", anomaly_info=anomaly_info)
fig = ts.plot(show_anomaly_adjustment=True)
assert len(fig.data) == 2
assert fig.data[0].name == value_col
assert fig.data[1].name == f"{value_col}_unadjusted"
assert fig.layout.xaxis.title.text == ts.original_time_col
assert fig.layout.yaxis.title.text == ts.original_value_col
assert fig.data[0].y.shape[0] == df.shape[0]
assert fig.data[1].y.shape[0] == df.shape[0]
# adjusted data has more NaNs, since anomalies are replaced with NaN
assert sum(np.isnan(fig.data[0].y)) == 10906
assert sum(np.isnan(fig.data[1].y)) == 2067
def test_make_univariate_time_series(df):
"""Tests make_univariate_time_series function"""
forecast = UnivariateForecast(df, train_end_date=datetime.datetime(2018, 1, 2))
ts = UnivariateTimeSeries()
ts.load_data(pd.DataFrame({
cst.TIME_COL: df[cst.TIME_COL],
cst.VALUE_COL: df[cst.PREDICTED_COL]
}), cst.TIME_COL, cst.VALUE_COL)
assert forecast.make_univariate_time_series().df.equals(ts.df)
def test_get_grouping_evaluation():
"""Tests get_grouping_evaluation function"""
df = pd.DataFrame({
"custom_time_column": [
datetime.datetime(2018, 1, 1),
datetime.datetime(2018, 1, 2),
datetime.datetime(2018, 1, 3),
datetime.datetime(2018, 1, 4),
datetime.datetime(2018, 1, 5)
],
"custom_value_column": [1.0, 2.0, 3.0, 4.0, 5.0],
})
ts = UnivariateTimeSeries()
ts.load_data(df,
time_col="custom_time_column",
value_col="custom_value_column")
# mean, groupby_time_feature
grouped_df = ts.get_grouping_evaluation(aggregation_func=np.mean,
aggregation_func_name="mean",
groupby_time_feature="dow")
expected = pd.DataFrame({
"dow":
[1, 2, 3, 4,
5], # Monday, Tuesday, etc. Time feature is used as column name
f"mean of {VALUE_COL}": [1.0, 2.0, 3.0, 4.0, 5.0]
})
assert_equal(grouped_df, expected)
# max, groupby_sliding_window_size
grouped_df = ts.get_grouping_evaluation(aggregation_func=np.max,
aggregation_func_name="max",
groupby_sliding_window_size=2)
expected = pd.DataFrame({
f"{TIME_COL}_downsample": [
datetime.datetime(2018, 1, 1),
datetime.datetime(2018, 1, 3),
datetime.datetime(2018, 1, 5)
],
f"max of {VALUE_COL}": [1.0, 3.0, 5.0]
})
assert_equal(grouped_df, expected)
# min, groupby_custom_column
grouped_df = ts.get_grouping_evaluation(aggregation_func=np.min,
aggregation_func_name=None,
groupby_custom_column=pd.Series(
["g1", "g2", "g1", "g3", "g2"],
name="custom_groups"))
expected = pd.DataFrame({
"custom_groups": ["g1", "g2", "g3"],
f"aggregation of {VALUE_COL}": [1.0, 2.0, 4.0]
})
assert_equal(grouped_df, expected)
def make_univariate_time_series(self):
"""Converts prediction into a UnivariateTimeSeries
Useful to convert a forecast into the input regressor for a subsequent forecast.
:return: UnivariateTimeSeries
"""
ts = UnivariateTimeSeries()
df = (self.df[[self.time_col, self.predicted_col]]
.rename({self.predicted_col: self.ylabel}, axis=1))
ts.load_data(df, self.time_col, self.ylabel)
return ts
def test_load_data_anomaly():
"""Checks anomaly_info parameter"""
dl = DataLoaderTS()
df = dl.load_beijing_pm()
value_col = "pm"
# no anomaly adjustment
ts = UnivariateTimeSeries()
ts.load_data(df=df, value_col=value_col)
assert ts.df_before_adjustment is None
# adjusts two columns
dim_one = "one"
dim_two = "two"
anomaly_df = pd.DataFrame({
START_DATE_COL: ["2011-04-04-10", "2011-10-10-00", "2012-12-20-10"],
END_DATE_COL: ["2011-04-05-20", "2011-10-11-23", "2012-12-20-13"],
ADJUSTMENT_DELTA_COL: [np.nan, 100.0, -100.0],
METRIC_COL: [dim_one, dim_one, dim_two]
})
anomaly_info = [{
"value_col": value_col,
"anomaly_df": anomaly_df,
"start_date_col": START_DATE_COL,
"end_date_col": END_DATE_COL,
"adjustment_delta_col": ADJUSTMENT_DELTA_COL,
"filter_by_dict": {
METRIC_COL: dim_one
},
"adjustment_method": "add"
}, {
"value_col": "pres",
"anomaly_df": anomaly_df,
"start_date_col": START_DATE_COL,
"end_date_col": END_DATE_COL,
"adjustment_delta_col": ADJUSTMENT_DELTA_COL,
"filter_by_dict": {
METRIC_COL: dim_two
},
"adjustment_method": "subtract"
}]
ts = UnivariateTimeSeries()
ts.load_data(df=df, value_col=value_col, anomaly_info=anomaly_info)
canonical_data_dict = get_canonical_data(df=df,
value_col=value_col,
anomaly_info=anomaly_info)
assert_equal(ts.df, canonical_data_dict["df"])
assert_equal(ts.df_before_adjustment,
canonical_data_dict["df_before_adjustment"])
def test_check_time_series_gaps():
"""Checks gaps filled for non-regular input"""
ts = UnivariateTimeSeries()
df = pd.DataFrame({
TIME_COL: [
dt(2018, 1, 1, 0, 0, 1),
dt(2018, 1, 1, 0, 0, 2),
dt(2018, 1, 1, 0, 0, 10), # intentionally out of order
dt(2018, 1, 1, 0, 0, 4)
],
VALUE_COL: [1, 2, 3, 4]
})
expected = pd.Series([
dt(2018, 1, 1, 0, 0, 1),
dt(2018, 1, 1, 0, 0, 2),
dt(2018, 1, 1, 0, 0, 3),
dt(2018, 1, 1, 0, 0, 4),
dt(2018, 1, 1, 0, 0, 5),
dt(2018, 1, 1, 0, 0, 6),
dt(2018, 1, 1, 0, 0, 7),
dt(2018, 1, 1, 0, 0, 8),
dt(2018, 1, 1, 0, 0, 9),
dt(2018, 1, 1, 0, 0, 10)
])
expected.index = expected
ts.load_data(
df, TIME_COL, VALUE_COL,
freq="S") # the frequency should be provided when there are gaps
assert ts.df[TIME_COL].equals(expected)
assert ts.time_stats["data_points"] == 10 # after filling in gaps
assert ts.value_stats["count"] == 4 # before filling in gaps
assert ts.time_stats["added_timepoints"] == 6
assert ts.time_stats["dropped_timepoints"] == 0
assert ts.df[VALUE_COL].equals(ts.y)
expected_gaps = pd.DataFrame({
"right_before_gap":
pd.Series([dt(2018, 1, 1, 0, 0, 2),
dt(2018, 1, 1, 0, 0, 4)]),
"right_after_gap":
pd.Series([dt(2018, 1, 1, 0, 0, 4),
dt(2018, 1, 1, 0, 0, 10)]),
"gap_size": [1.0, 5.0]
})
assert ts.time_stats["gaps"].equals(expected_gaps)
def test_check_time_series_tz_local():
"""Checks date parsing with localization"""
expected = pd.Series([
dt(2018, 1, 1, 0, 0, 0),
dt(2018, 1, 5, 0, 0, 0),
dt(2018, 1, 9, 0, 0, 0)
])
expected.index = expected
expected = expected.tz_localize("US/Pacific")
ts = UnivariateTimeSeries()
df = pd.DataFrame({
TIME_COL: ["2018-01-01", "2018-01-05", "2018-01-09"],
VALUE_COL: [1, 2, 3]
})
ts.load_data(df, TIME_COL, VALUE_COL, tz="US/Pacific", freq="4D")
assert ts.time_stats["added_timepoints"] == 0
assert ts.time_stats["dropped_timepoints"] == 0
assert ts.df[TIME_COL].equals(expected)
assert ts.df[VALUE_COL].equals(ts.y)
def test_check_time_series2():
"""Checks value column stats"""
ts = UnivariateTimeSeries()
df = pd.DataFrame({
TIME_COL: [
dt(2018, 1, 1, 0, 0, 1),
dt(2018, 1, 1, 0, 0, 2),
dt(2018, 1, 1, 0, 0, 3)
],
VALUE_COL: [1, 2, 3]
})
ts.load_data(df, TIME_COL, VALUE_COL)
assert ts.value_stats["mean"] == 2.0
assert ts.value_stats["std"] == 1.0
assert ts.value_stats["min"] == 1.0
assert ts.value_stats["25%"] == 1.5
assert ts.value_stats["50%"] == 2.0
assert ts.value_stats["75%"] == 2.5
assert ts.value_stats["max"] == 3.0
assert ts.df[VALUE_COL].equals(ts.y)
def test_check_time_series1():
"""Checks if regular data can be properly loaded. Checks time column stats"""
ts = UnivariateTimeSeries()
df = pd.DataFrame({
"time": [
dt(2018, 1, 1, 0, 0, 1),
dt(2018, 1, 1, 0, 0, 2),
dt(2018, 1, 1, 0, 0, 3)
],
"val": [1, 2, 3]
})
ts.load_data(df, "time", "val")
assert ts.original_time_col == "time"
assert ts.original_value_col == "val"
assert ts.time_stats["data_points"] == 3
assert ts.time_stats["mean_increment_secs"] == 1.0
assert ts.time_stats["min_timestamp"] == df.min()[0]
assert ts.time_stats["max_timestamp"] == df.max()[0]
assert ts.time_stats["added_timepoints"] == 0
assert ts.time_stats["dropped_timepoints"] == 0
assert ts.freq == "S"
assert ts.df[VALUE_COL].equals(ts.y)
assert ts.df.index.name is None
def forecast_pipeline(
# input
df: pd.DataFrame,
time_col=TIME_COL,
value_col=VALUE_COL,
date_format=None,
tz=None,
freq=None,
train_end_date=None,
anomaly_info=None,
# model
pipeline=None,
regressor_cols=None,
lagged_regressor_cols=None,
estimator=SimpleSilverkiteEstimator(),
hyperparameter_grid=None,
hyperparameter_budget=None,
n_jobs=COMPUTATION_N_JOBS,
verbose=1,
# forecast
forecast_horizon=None,
coverage=0.95,
test_horizon=None,
periods_between_train_test=None,
agg_periods=None,
agg_func=None,
# evaluation
score_func=EvaluationMetricEnum.MeanAbsolutePercentError.name,
score_func_greater_is_better=False,
cv_report_metrics=CV_REPORT_METRICS_ALL,
null_model_params=None,
relative_error_tolerance=None,
# CV
cv_horizon=None,
cv_min_train_periods=None,
cv_expanding_window=False,
cv_use_most_recent_splits=False,
cv_periods_between_splits=None,
cv_periods_between_train_test=None,
cv_max_splits=3):
"""Computation pipeline for end-to-end forecasting.
Trains a forecast model end-to-end:
1. checks input data
2. runs cross-validation to select optimal hyperparameters e.g. best model
3. evaluates best model on test set
4. provides forecast of best model (re-trained on all data) into the future
Returns forecasts with methods to plot and see diagnostics.
Also returns the fitted pipeline and CV results.
Provides a high degree of customization over training and evaluation parameters:
1. model
2. cross validation
3. evaluation
4. forecast horizon
See test cases for examples.
Parameters
----------
df : `pandas.DataFrame`
Timeseries data to forecast.
Contains columns [`time_col`, `value_col`], and optional regressor columns
Regressor columns should include future values for prediction
time_col : `str`, default TIME_COL in constants.py
name of timestamp column in df
value_col : `str`, default VALUE_COL in constants.py
name of value column in df (the values to forecast)
date_format : `str` or None, default None
strftime format to parse time column, eg ``%m/%d/%Y``.
Note that ``%f`` will parse all the way up to nanoseconds.
If None (recommended), inferred by `pandas.to_datetime`.
tz : `str` or None, default None
Passed to `pandas.tz_localize` to localize the timestamp
freq : `str` or None, default None
Frequency of input data. Used to generate future dates for prediction.
Frequency strings can have multiples, e.g. '5H'.
See https://pandas.pydata.org/pandas-docs/stable/user_guide/timeseries.html#offset-aliases
for a list of frequency aliases.
If None, inferred by `pandas.infer_freq`.
Provide this parameter if ``df`` has missing timepoints.
train_end_date : `datetime.datetime`, optional, default None
Last date to use for fitting the model. Forecasts are generated after this date.
If None, it is set to the last date with a non-null value in
``value_col`` of ``df``.
anomaly_info : `dict` or `list` [`dict`] or None, default None
Anomaly adjustment info. Anomalies in ``df``
are corrected before any forecasting is done.
If None, no adjustments are made.
A dictionary containing the parameters to
`~greykite.common.features.adjust_anomalous_data.adjust_anomalous_data`.
See that function for details.
The possible keys are:
``"value_col"`` : `str`
The name of the column in ``df`` to adjust. You may adjust the value
to forecast as well as any numeric regressors.
``"anomaly_df"`` : `pandas.DataFrame`
Adjustments to correct the anomalies.
``"start_date_col"``: `str`, default START_DATE_COL
Start date column in ``anomaly_df``.
``"end_date_col"``: `str`, default END_DATE_COL
End date column in ``anomaly_df``.
``"adjustment_delta_col"``: `str` or None, default None
Impact column in ``anomaly_df``.
``"filter_by_dict"``: `dict` or None, default None
Used to filter ``anomaly_df`` to the relevant anomalies for
the ``value_col`` in this dictionary.
Key specifies the column name, value specifies the filter value.
``"filter_by_value_col""``: `str` or None, default None
Adds ``{filter_by_value_col: value_col}`` to ``filter_by_dict``
if not None, for the ``value_col`` in this dictionary.
``"adjustment_method"`` : `str` ("add" or "subtract"), default "add"
How to make the adjustment, if ``adjustment_delta_col`` is provided.
Accepts a list of such dictionaries to adjust multiple columns in ``df``.
pipeline : `sklearn.pipeline.Pipeline` or None, default None
Pipeline to fit. The final named step must be called "estimator".
If None, will use the default Pipeline from
`~greykite.framework.pipeline.utils.get_basic_pipeline`.
regressor_cols : `list` [`str`] or None, default None
A list of regressor columns used in the training and prediction DataFrames.
It should contain only the regressors that are being used in the grid search.
If None, no regressor columns are used.
Regressor columns that are unavailable in ``df`` are dropped.
lagged_regressor_cols : `list` [`str`] or None, default None
A list of additional columns needed for lagged regressors in the training and prediction DataFrames.
This list can have overlap with ``regressor_cols``.
If None, no additional columns are added to the DataFrame.
Lagged regressor columns that are unavailable in ``df`` are dropped.
estimator : instance of an estimator that implements `greykite.algo.models.base_forecast_estimator.BaseForecastEstimator`
Estimator to use as the final step in the pipeline.
Ignored if ``pipeline`` is provided.
forecast_horizon : `int` or None, default None
Number of periods to forecast into the future. Must be > 0.
If None, default is determined from input data frequency
coverage : `float` or None, default=0.95
Intended coverage of the prediction bands (0.0 to 1.0)
If None, the upper/lower predictions are not returned
Ignored if `pipeline` is provided. Uses coverage of the ``pipeline`` estimator instead.
test_horizon : `int` or None, default None
Numbers of periods held back from end of df for test.
The rest is used for cross validation.
If None, default is forecast_horizon. Set to 0 to skip backtest.
periods_between_train_test : `int` or None, default None
Number of periods for the gap between train and test data.
If None, default is 0.
agg_periods : `int` or None, default None
Number of periods to aggregate before evaluation.
Model is fit and forecasted on the dataset's original frequency.
Before evaluation, the actual and forecasted values are aggregated,
using rolling windows of size ``agg_periods`` and the function
``agg_func``. (e.g. if the dataset is hourly, use ``agg_periods=24, agg_func=np.sum``,
to evaluate performance on the daily totals).
If None, does not aggregate before evaluation.
Currently, this is only used when calculating CV metrics and
the R2_null_model_score metric in backtest/forecast. No pre-aggregation
is applied for the other backtest/forecast evaluation metrics.
agg_func : callable or None, default None
Takes an array and returns a number, e.g. np.max, np.sum.
Defines how to aggregate rolling windows of actual and predicted values
before evaluation.
Ignored if ``agg_periods`` is None.
Currently, this is only used when calculating CV metrics and
the R2_null_model_score metric in backtest/forecast. No pre-aggregation
is applied for the other backtest/forecast evaluation metrics.
score_func : `str` or callable, default ``EvaluationMetricEnum.MeanAbsolutePercentError.name``
Score function used to select optimal model in CV.
If a callable, takes arrays ``y_true``, ``y_pred`` and returns a float.
If a string, must be either a
`~greykite.common.evaluation.EvaluationMetricEnum` member name
or `~greykite.common.constants.FRACTION_OUTSIDE_TOLERANCE`.
score_func_greater_is_better : `bool`, default False
True if ``score_func`` is a score function, meaning higher is better,
and False if it is a loss function, meaning lower is better.
Must be provided if ``score_func`` is a callable (custom function).
Ignored if ``score_func`` is a string, because the direction is known.
cv_report_metrics : `str`, or `list` [`str`], or None, default `~greykite.common.constants.CV_REPORT_METRICS_ALL`
Additional metrics to compute during CV, besides the one specified by ``score_func``.
- If the string constant `greykite.framework.constants.CV_REPORT_METRICS_ALL`,
computes all metrics in ``EvaluationMetricEnum``. Also computes
``FRACTION_OUTSIDE_TOLERANCE`` if ``relative_error_tolerance`` is not None.
The results are reported by the short name (``.get_metric_name()``) for ``EvaluationMetricEnum``
members and ``FRACTION_OUTSIDE_TOLERANCE_NAME`` for ``FRACTION_OUTSIDE_TOLERANCE``.
These names appear in the keys of ``forecast_result.grid_search.cv_results_``
returned by this function.
- If a list of strings, each of the listed metrics is computed. Valid strings are
`~greykite.common.evaluation.EvaluationMetricEnum` member names
and `~greykite.common.constants.FRACTION_OUTSIDE_TOLERANCE`.
For example::
["MeanSquaredError", "MeanAbsoluteError", "MeanAbsolutePercentError", "MedianAbsolutePercentError", "FractionOutsideTolerance2"]
- If None, no additional metrics are computed.
null_model_params : `dict` or None, default None
Defines baseline model to compute ``R2_null_model_score`` evaluation metric.
``R2_null_model_score`` is the improvement in the loss function relative
to a null model. It can be used to evaluate model quality with respect to
a simple baseline. For details, see
`~greykite.common.evaluation.r2_null_model_score`.
The null model is a `~sklearn.dummy.DummyRegressor`,
which returns constant predictions.
Valid keys are "strategy", "constant", "quantile".
See `~sklearn.dummy.DummyRegressor`. For example::
null_model_params = {
"strategy": "mean",
}
null_model_params = {
"strategy": "median",
}
null_model_params = {
"strategy": "quantile",
"quantile": 0.8,
}
null_model_params = {
"strategy": "constant",
"constant": 2.0,
}
If None, ``R2_null_model_score`` is not calculated.
Note: CV model selection always optimizes ``score_func`, not
the ``R2_null_model_score``.
relative_error_tolerance : `float` or None, default None
Threshold to compute the ``Outside Tolerance`` metric,
defined as the fraction of forecasted values whose relative
error is strictly greater than ``relative_error_tolerance``.
For example, 0.05 allows for 5% relative error.
If `None`, the metric is not computed.
hyperparameter_grid : `dict`, `list` [`dict`] or None, default None
Sets properties of the steps in the pipeline,
and specifies combinations to search over.
Should be valid input to `sklearn.model_selection.GridSearchCV` (param_grid)
or `sklearn.model_selection.RandomizedSearchCV` (param_distributions).
Prefix transform/estimator attributes by the name of the step in the pipeline.
See details at: https://scikit-learn.org/stable/modules/compose.html#nested-parameters
If None, uses the default pipeline parameters.
hyperparameter_budget : `int` or None, default None
Max number of hyperparameter sets to try within the ``hyperparameter_grid`` search space
Runs a full grid search if ``hyperparameter_budget`` is sufficient to exhaust full
``hyperparameter_grid``, otherwise samples uniformly at random from the space.
If None, uses defaults:
* full grid search if all values are constant
* 10 if any value is a distribution to sample from
n_jobs : `int` or None, default `~greykite.framework.constants.COMPUTATION_N_JOBS`
Number of jobs to run in parallel
(the maximum number of concurrently running workers).
``-1`` uses all CPUs. ``-2`` uses all CPUs but one.
``None`` is treated as 1 unless in a `joblib.Parallel` backend context
that specifies otherwise.
verbose : `int`, default 1
Verbosity level during CV.
if > 0, prints number of fits
if > 1, prints fit parameters, total score + fit time
if > 2, prints train/test scores
cv_horizon : `int` or None, default None
Number of periods in each CV test set
If None, default is ``forecast_horizon``.
Set either ``cv_horizon`` or ``cv_max_splits`` to 0 to skip CV.
cv_min_train_periods : `int` or None, default None
Minimum number of periods for training each CV fold.
If cv_expanding_window is False, every training period is this size
If None, default is 2 * ``cv_horizon``
cv_expanding_window : `bool`, default False
If True, training window for each CV split is fixed to the first available date.
Otherwise, train start date is sliding, determined by ``cv_min_train_periods``.
cv_use_most_recent_splits: `bool`, default False
If True, splits from the end of the dataset are used.
Else a sampling strategy is applied. Check
`~greykite.sklearn.cross_validation.RollingTimeSeriesSplit._sample_splits`
for details.
cv_periods_between_splits : `int` or None, default None
Number of periods to slide the test window between CV splits
If None, default is ``cv_horizon``
cv_periods_between_train_test : `int` or None, default None
Number of periods for the gap between train and test in a CV split.
If None, default is ``periods_between_train_test``.
cv_max_splits : `int` or None, default 3
Maximum number of CV splits.
Given the above configuration, samples up to max_splits train/test splits,
preferring splits toward the end of available data. If None, uses all splits.
Set either ``cv_horizon`` or ``cv_max_splits`` to 0 to skip CV.
Returns
-------
forecast_result : :class:`~greykite.framework.pipeline.pipeline.ForecastResult`
Forecast result. See :class:`~greykite.framework.pipeline.pipeline.ForecastResult`
for details.
* If ``cv_horizon=0``, ``forecast_result.grid_search.best_estimator_``
and ``forecast_result.grid_search.best_params_`` attributes are defined
according to the provided single set of parameters. There must be a single
set of parameters to skip cross-validation.
* If ``test_horizon=0``, ``forecast_result.backtest`` is None.
"""
if hyperparameter_grid is None or hyperparameter_grid == []:
hyperparameter_grid = {}
# When hyperparameter_grid is a singleton list, unlist it
if isinstance(hyperparameter_grid, list) and len(hyperparameter_grid) == 1:
hyperparameter_grid = hyperparameter_grid[0]
# Loads full dataset
ts = UnivariateTimeSeries()
ts.load_data(
df=df,
time_col=time_col,
value_col=value_col,
freq=freq,
date_format=date_format,
tz=tz,
train_end_date=train_end_date,
regressor_cols=regressor_cols,
lagged_regressor_cols=lagged_regressor_cols,
anomaly_info=anomaly_info)
# Splits data into training and test sets. ts.df uses standardized column names
if test_horizon == 0:
train_df = ts.fit_df
train_y = ts.fit_y
test_df = pd.DataFrame(columns=list(df.columns))
else:
# Make sure to refit best_pipeline appropriately
train_df, test_df, train_y, test_y = train_test_split(
ts.fit_df,
ts.fit_y,
train_size=ts.fit_df.shape[0] - test_horizon - periods_between_train_test,
test_size=test_horizon + periods_between_train_test,
shuffle=False) # this is important since this is timeseries forecasting!
log_message(f"Train size: {train_df.shape[0]}. Test size: {test_df.shape[0]}", LoggingLevelEnum.INFO)
# Defines default training pipeline
if pipeline is None:
pipeline = get_basic_pipeline(
estimator=estimator,
score_func=score_func,
score_func_greater_is_better=score_func_greater_is_better,
agg_periods=agg_periods,
agg_func=agg_func,
relative_error_tolerance=relative_error_tolerance,
coverage=coverage,
null_model_params=null_model_params,
regressor_cols=ts.regressor_cols,
lagged_regressor_cols=ts.lagged_regressor_cols)
# Searches for the best parameters, and refits model with selected parameters on the entire training set
if cv_horizon == 0 or cv_max_splits == 0:
# No cross-validation. Only one set of hyperparameters is allowed.
try:
if len(ParameterGrid(hyperparameter_grid)) > 1:
raise ValueError(
"CV is required to identify the best model because there are multiple options "
"in `hyperparameter_grid`. Either provide a single option or set `cv_horizon` and `cv_max_splits` "
"to nonzero values.")
except TypeError: # Parameter value is not iterable
raise ValueError(
"CV is required to identify the best model because `hyperparameter_grid` contains "
"a distribution. Either remove the distribution or set `cv_horizon` and `cv_max_splits` "
"to nonzero values.")
# Fits model to entire train set. Params must be set manually since it's not done by grid search
params = {k: v[0] for k, v in hyperparameter_grid.items()} # unpack lists, `v` is a singleton list with the parameter value
best_estimator = pipeline.set_params(**params).fit(train_df, train_y)
# Wraps this model in a dummy RandomizedSearchCV object to return the backtest model
grid_search = get_hyperparameter_searcher(
hyperparameter_grid=hyperparameter_grid,
model=pipeline,
cv=None, # no cross-validation
hyperparameter_budget=hyperparameter_budget,
n_jobs=n_jobs,
verbose=verbose,
score_func=score_func,
score_func_greater_is_better=score_func_greater_is_better,
cv_report_metrics=cv_report_metrics,
agg_periods=agg_periods,
agg_func=agg_func,
relative_error_tolerance=relative_error_tolerance)
# Sets relevant attributes. Others are undefined (cv_results_, best_score_, best_index_, scorer_, refit_time_)
grid_search.best_estimator_ = best_estimator
grid_search.best_params_ = params
grid_search.n_splits_ = 0
else:
# Defines cross-validation splitter
cv = RollingTimeSeriesSplit(
forecast_horizon=cv_horizon,
min_train_periods=cv_min_train_periods,
expanding_window=cv_expanding_window,
use_most_recent_splits=cv_use_most_recent_splits,
periods_between_splits=cv_periods_between_splits,
periods_between_train_test=cv_periods_between_train_test,
max_splits=cv_max_splits)
# Defines grid search approach for CV
grid_search = get_hyperparameter_searcher(
hyperparameter_grid=hyperparameter_grid,
model=pipeline,
cv=cv,
hyperparameter_budget=hyperparameter_budget,
n_jobs=n_jobs,
verbose=verbose,
score_func=score_func,
score_func_greater_is_better=score_func_greater_is_better,
cv_report_metrics=cv_report_metrics,
agg_periods=agg_periods,
agg_func=agg_func,
relative_error_tolerance=relative_error_tolerance)
grid_search.fit(train_df, train_y)
best_estimator = grid_search.best_estimator_
# Evaluates historical performance, fits model to all data (train+test)
if test_horizon > 0:
backtest_train_end_date = train_df[TIME_COL].max()
# Uses pd.date_range because pd.Timedelta does not work for complicated frequencies e.g. "W-MON"
backtest_test_start_date = pd.date_range(
start=backtest_train_end_date,
periods=periods_between_train_test + 2, # Adds 2 as start parameter is inclusive
freq=ts.freq)[-1]
backtest = get_forecast(
df=ts.fit_df, # Backtest needs to happen on fit_df, not on the entire df
trained_model=best_estimator,
train_end_date=backtest_train_end_date,
test_start_date=backtest_test_start_date,
forecast_horizon=test_horizon,
xlabel=time_col,
ylabel=value_col,
relative_error_tolerance=relative_error_tolerance)
best_pipeline = clone(best_estimator) # Copies optimal parameters
best_pipeline.fit(ts.fit_df, ts.y) # Refits this model on entire training dataset
else:
backtest = None # Backtest training metrics are the same as forecast training metrics
best_pipeline = best_estimator # best_model is already fit to all data
# Makes future predictions
periods = forecast_horizon + periods_between_train_test
future_df = ts.make_future_dataframe(
periods=periods,
include_history=True)
forecast_train_end_date = ts.train_end_date
# Uses pd.date_range because pd.Timedelta does not work for complicated frequencies e.g. "W-MON"
forecast_test_start_date = pd.date_range(
start=forecast_train_end_date,
periods=periods_between_train_test + 2, # Adds 2 as start parameter is inclusive
freq=ts.freq)[-1]
forecast = get_forecast(
df=future_df,
trained_model=best_pipeline,
train_end_date=forecast_train_end_date,
test_start_date=forecast_test_start_date,
forecast_horizon=forecast_horizon,
xlabel=time_col,
ylabel=value_col,
relative_error_tolerance=relative_error_tolerance)
result = ForecastResult(
timeseries=ts,
grid_search=grid_search,
model=best_pipeline,
backtest=backtest,
forecast=forecast
)
return result
def test_make_future_dataframe():
"""Checks future dataframe creation"""
ts = UnivariateTimeSeries()
df = pd.DataFrame({
TIME_COL: [
dt(2018, 1, 1, 3, 0, 0),
dt(2018, 1, 1, 4, 0, 0),
dt(2018, 1, 1, 5, 0, 0),
dt(2018, 1, 1, 6, 0, 0),
dt(2018, 1, 1, 7, 0, 0)
],
VALUE_COL: [1, None, 3, None, None],
})
with pytest.warns(UserWarning) as record:
ts.load_data(df, TIME_COL, VALUE_COL, regressor_cols=None)
assert f"{ts.original_value_col} column of the provided TimeSeries contains null " \
f"values at the end. Setting 'train_end_date' to the last timestamp with a " \
f"non-null value ({ts.train_end_date})." in record[0].message.args[0]
# test regressor_cols from load_data
assert ts.regressor_cols == []
# tests last_date_for_val from load_data
assert ts.last_date_for_val == dt(2018, 1, 1, 5, 0, 0)
assert ts.train_end_date == dt(2018, 1, 1, 5, 0, 0)
# tests last_date_for_reg from load_data
assert ts.last_date_for_reg is None
# tests fit_df from load_data
result_fit_df = ts.fit_df.reset_index(
drop=True) # fit_df's index is time_col
expected_fit_df = df[df[TIME_COL] <= dt(2018, 1, 1, 5, 0, 0)]
assert_frame_equal(result_fit_df, expected_fit_df)
# tests fit_y
result_fit_y = ts.fit_y.reset_index(drop=True) # fit_y's index is time_col
expected_fit_y = expected_fit_df[VALUE_COL]
assert result_fit_y.equals(expected_fit_y)
# with history, default value for periods
result = ts.make_future_dataframe(periods=None, include_history=True)
expected = pd.DataFrame({
TIME_COL:
pd.date_range(start=dt(2018, 1, 1, 3, 0, 0), periods=33, freq="H"),
VALUE_COL:
np.concatenate((result_fit_y, np.repeat(np.nan, 30)))
})
expected.index = expected[TIME_COL]
expected.index.name = None
assert_frame_equal(result, expected)
# without history
result = ts.make_future_dataframe(periods=10, include_history=False)
expected = pd.DataFrame({
TIME_COL:
pd.date_range(start=dt(2018, 1, 1, 6, 0, 0), periods=10, freq="H"),
VALUE_COL:
np.repeat(np.nan, 10)
})
expected.index = expected[TIME_COL]
expected.index.name = None
expected.index.freq = "H"
assert_frame_equal(result, expected)
from greykite.framework.input.univariate_time_series import UnivariateTimeSeries
from greykite.framework.constants import MEAN_COL_GROUP, OVERLAY_COL_GROUP
from greykite.common.constants import TIME_COL
from greykite.common.data_loader import DataLoader
from greykite.common.viz.timeseries_plotting import add_groupby_column, plot_multivariate, plot_univariate
# Loads dataset into pandas DataFrame
dl = DataLoader()
df = dl.load_peyton_manning()
df.rename(columns={"y": "log(pageviews)"},
inplace=True) # uses a more informative name
# plots dataset
ts = UnivariateTimeSeries()
ts.load_data(df=df, time_col="ts", value_col="log(pageviews)", freq="D")
fig = ts.plot()
plotly.io.show(fig)
# %%
# Yearly seasonality
# ------------------
# Because the observations are at daily frequency,
# it is possible to see yearly, quarterly, monthly, and weekly seasonality.
# The name of the seasonality refers to the length of one cycle. For example,
# yearly seasonality is a pattern that repeats once a year.
#
# .. tip::
# It's helpful to start with the longest cycle to see the big picture.
#
# To examine yearly seasonality, plot the average value by day of year.
def test_get_quantiles_and_overlays():
"""Tests get_quantiles_and_overlays"""
dl = DataLoaderTS()
peyton_manning_ts = dl.load_peyton_manning_ts()
# no columns are requested
with pytest.raises(
ValueError,
match=
"Must enable at least one of: show_mean, show_quantiles, show_overlays."
):
peyton_manning_ts.get_quantiles_and_overlays(
groupby_time_feature="doy")
# show_mean only
grouped_df = peyton_manning_ts.get_quantiles_and_overlays(
groupby_time_feature="dow",
show_mean=True,
mean_col_name="custom_name")
assert_equal(
grouped_df.columns,
pd.MultiIndex.from_arrays([[MEAN_COL_GROUP], ["custom_name"]],
names=["category", "name"]))
assert grouped_df.index.name == "dow"
assert grouped_df.shape == (7, 1)
assert grouped_df.index[0] == 1
# show_quantiles only (bool)
grouped_df = peyton_manning_ts.get_quantiles_and_overlays(
groupby_sliding_window_size=180, show_quantiles=True)
assert_equal(
grouped_df.columns,
pd.MultiIndex.from_arrays(
[[QUANTILE_COL_GROUP, QUANTILE_COL_GROUP], ["Q0.1", "Q0.9"]],
names=["category", "name"]))
assert grouped_df.index.name == "ts_downsample"
assert grouped_df.shape == (17, 2)
assert grouped_df.index[0] == pd.Timestamp(2007, 12, 10)
# show_quantiles only (list)
custom_col = pd.Series(
np.random.choice(list("abcd"), size=peyton_manning_ts.df.shape[0]))
grouped_df = peyton_manning_ts.get_quantiles_and_overlays(
groupby_custom_column=custom_col,
show_quantiles=[0, 0.25, 0.5, 0.75, 1],
quantile_col_prefix="prefix")
assert_equal(
grouped_df.columns,
pd.MultiIndex.from_arrays(
[[QUANTILE_COL_GROUP] * 5,
["prefix0", "prefix0.25", "prefix0.5", "prefix0.75", "prefix1"]],
names=["category", "name"]))
assert grouped_df.index.name == "groups"
assert grouped_df.shape == (4, 5)
assert grouped_df.index[0] == "a"
# checks quantile computation
df = peyton_manning_ts.df.copy()
df["custom_col"] = custom_col.values
quantile_df = df.groupby("custom_col")[VALUE_COL].agg(
[np.nanmin, np.nanmedian, np.nanmax])
assert_equal(grouped_df["quantile"]["prefix0"],
quantile_df["nanmin"],
check_names=False)
assert_equal(grouped_df["quantile"]["prefix0.5"],
quantile_df["nanmedian"],
check_names=False)
assert_equal(grouped_df["quantile"]["prefix1"],
quantile_df["nanmax"],
check_names=False)
# show_overlays only (bool), no overlay label
grouped_df = peyton_manning_ts.get_quantiles_and_overlays(
groupby_time_feature="doy", show_overlays=True)
assert_equal(
grouped_df.columns,
pd.MultiIndex.from_arrays(
[[OVERLAY_COL_GROUP] * 9, [f"overlay{i}" for i in range(9)]],
names=["category", "name"]))
assert grouped_df.index.name == "doy"
assert grouped_df.shape == (366, 9)
assert grouped_df.index[0] == 1
# show_overlays only (int below the available number), time feature overlay label
np.random.seed(123)
grouped_df = peyton_manning_ts.get_quantiles_and_overlays(
groupby_time_feature="doy",
show_overlays=4,
overlay_label_time_feature="year")
assert_equal(
grouped_df.columns,
pd.MultiIndex.from_arrays(
[[OVERLAY_COL_GROUP] * 4, ["2007", "2011", "2012", "2014"]],
names=["category", "name"]))
assert grouped_df.index.name == "doy"
assert grouped_df.shape == (366, 4)
assert grouped_df.index[0] == 1
# show_overlays only (int above the available number), custom overlay label
grouped_df = peyton_manning_ts.get_quantiles_and_overlays(
groupby_time_feature="dom",
show_overlays=200,
overlay_label_custom_column=custom_col)
assert_equal(
grouped_df.columns,
pd.MultiIndex.from_arrays(
[[OVERLAY_COL_GROUP] * 4, ["a", "b", "c", "d"]],
names=["category", "name"]))
assert grouped_df.index.name == "dom"
assert grouped_df.shape == (31, 4)
assert grouped_df.index[0] == 1
# show_overlays only (list of indices), sliding window overlay label
grouped_df = peyton_manning_ts.get_quantiles_and_overlays(
groupby_time_feature="dom",
show_overlays=[0, 4],
overlay_label_sliding_window_size=365 * 2)
assert_equal(
grouped_df.columns,
pd.MultiIndex.from_arrays(
[[OVERLAY_COL_GROUP] * 2,
["2007-12-10 00:00:00", "2015-12-08 00:00:00"]],
names=["category", "name"]))
assert grouped_df.index.name == "dom"
assert grouped_df.shape == (31, 2)
assert grouped_df.index[0] == 1
# show_overlays only (np.ndarray), sliding window overlay label
grouped_df = peyton_manning_ts.get_quantiles_and_overlays(
groupby_time_feature="dom",
show_overlays=np.arange(0, 6, 2),
overlay_label_sliding_window_size=365 * 2)
assert_equal(
grouped_df.columns,
pd.MultiIndex.from_arrays(
[[OVERLAY_COL_GROUP] * 3,
[
"2007-12-10 00:00:00", "2011-12-09 00:00:00",
"2015-12-08 00:00:00"
]],
names=["category", "name"]))
assert grouped_df.index.name == "dom"
assert grouped_df.shape == (31, 3)
assert grouped_df.index[0] == 1
# show_overlays only (list of column names), sliding window overlay label
grouped_df = peyton_manning_ts.get_quantiles_and_overlays(
groupby_time_feature="dom",
show_overlays=["2007-12-10 00:00:00", "2015-12-08 00:00:00"],
overlay_label_sliding_window_size=365 * 2)
assert_equal(
grouped_df.columns,
pd.MultiIndex.from_arrays(
[[OVERLAY_COL_GROUP] * 2,
["2007-12-10 00:00:00", "2015-12-08 00:00:00"]],
names=["category", "name"]))
assert grouped_df.index.name == "dom"
assert grouped_df.shape == (31, 2)
assert grouped_df.index[0] == 1
# Show all 3 (no overlay label)
grouped_df = peyton_manning_ts.get_quantiles_and_overlays(
groupby_sliding_window_size=50, # 50 per group (50 overlays)
show_mean=True,
show_quantiles=[0.05, 0.5, 0.95], # 3 quantiles
show_overlays=True)
assert_equal(
grouped_df.columns,
pd.MultiIndex.from_arrays(
[[MEAN_COL_GROUP] + [QUANTILE_COL_GROUP] * 3 +
[OVERLAY_COL_GROUP] * 50, ["mean", "Q0.05", "Q0.5", "Q0.95"] +
[f"overlay{i}" for i in range(50)]],
names=["category", "name"]))
assert grouped_df.index.name == "ts_downsample"
assert grouped_df.shape == (60, 54)
assert grouped_df.index[-1] == pd.Timestamp(2016, 1, 7)
# Show all 3 (with overlay label).
# Pass overlay_pivot_table_kwargs.
grouped_df = peyton_manning_ts.get_quantiles_and_overlays(
groupby_sliding_window_size=180,
show_mean=True,
show_quantiles=[0.05, 0.5, 0.95], # 3 quantiles
show_overlays=True,
overlay_label_time_feature="dow", # 7 possible values
aggfunc="median")
assert_equal(
grouped_df.columns,
pd.MultiIndex.from_arrays(
[[MEAN_COL_GROUP] + [QUANTILE_COL_GROUP] * 3 +
[OVERLAY_COL_GROUP] * 7,
[
"mean", "Q0.05", "Q0.5", "Q0.95", "1", "2", "3", "4", "5",
"6", "7"
]],
names=["category", "name"]))
assert grouped_df.index.name == "ts_downsample"
assert grouped_df.shape == (17, 11)
assert grouped_df.index[-1] == pd.Timestamp(2015, 10, 29)
assert np.linalg.norm(
grouped_df[OVERLAY_COL_GROUP].mean()) > 1.0 # not centered
with pytest.raises(
TypeError,
match="pivot_table\\(\\) got an unexpected keyword argument 'aggfc'"
):
peyton_manning_ts.get_quantiles_and_overlays(
groupby_sliding_window_size=180,
show_mean=True,
show_quantiles=[0.05, 0.5, 0.95],
show_overlays=True,
overlay_label_time_feature="dow",
aggfc=np.nanmedian) # unrecognized parameter
# center_values with show_mean=True
centered_df = peyton_manning_ts.get_quantiles_and_overlays(
groupby_sliding_window_size=180,
show_mean=True,
show_quantiles=[0.05, 0.5, 0.95],
show_overlays=True,
overlay_label_time_feature="dow",
aggfunc="median",
center_values=True)
assert np.linalg.norm(centered_df[[MEAN_COL_GROUP, OVERLAY_COL_GROUP
]].mean()) < 1e-8 # centered at 0
assert_equal(
centered_df[QUANTILE_COL_GROUP],
grouped_df[QUANTILE_COL_GROUP] - grouped_df[MEAN_COL_GROUP].mean()[0])
# center_values with show_mean=False
centered_df = peyton_manning_ts.get_quantiles_and_overlays(
groupby_sliding_window_size=180,
show_mean=False,
show_quantiles=[0.05, 0.5, 0.95],
show_overlays=True,
overlay_label_time_feature="dow",
aggfunc="median",
center_values=True)
assert np.linalg.norm(centered_df[[OVERLAY_COL_GROUP
]].mean()) < 1e-8 # centered at 0
overall_mean = peyton_manning_ts.df[VALUE_COL].mean()
assert_equal(centered_df[QUANTILE_COL_GROUP],
grouped_df[QUANTILE_COL_GROUP] - overall_mean)
# new value_col
df = generate_df_with_reg_for_tests(freq="D", periods=700)["df"]
ts = UnivariateTimeSeries()
ts.load_data(df=df)
grouped_df = ts.get_quantiles_and_overlays(
groupby_time_feature="dow",
show_mean=True,
show_quantiles=True,
show_overlays=True,
overlay_label_time_feature="woy",
value_col="regressor1")
df_dow = add_groupby_column(df=ts.df,
time_col=TIME_COL,
groupby_time_feature="dow")
dow_mean = df_dow["df"].groupby("dow").agg(
mean=pd.NamedAgg(column="regressor1", aggfunc=np.nanmean))
assert_equal(grouped_df["mean"], dow_mean, check_names=False)
def test_make_future_dataframe_with_regressor():
"""Checks future dataframe creation"""
ts = UnivariateTimeSeries()
df = pd.DataFrame({
TIME_COL: [
dt(2018, 1, 1, 3, 0, 0),
dt(2018, 1, 1, 4, 0, 0),
dt(2018, 1, 1, 5, 0, 0),
dt(2018, 1, 1, 6, 0, 0),
dt(2018, 1, 1, 7, 0, 0),
dt(2018, 1, 1, 8, 0, 0)
],
VALUE_COL: [1, None, 3, None, None, None],
"regressor1": [0.01, None, 0.014, 0.016, 0.017, None],
"regressor2": [0.11, 0.112, 0.114, 0.116, None, None],
"regressor3": [0.21, 0.212, 0.214, 0.216, 0.217, None]
})
regressor_cols = [
col for col in df.columns if col not in [TIME_COL, VALUE_COL]
]
with pytest.warns(Warning) as record:
ts.load_data(df, TIME_COL, VALUE_COL, regressor_cols=regressor_cols)
assert "y column of the provided TimeSeries contains null " \
"values at the end" in record[0].message.args[0]
# test regressor_cols from load_data
assert ts.regressor_cols == ["regressor1", "regressor2", "regressor3"]
# tests last_date_for_fit from load_data (same as without regressor)
assert ts.train_end_date == dt(2018, 1, 1, 5, 0, 0)
# tests last_date_for_reg from load_data
assert ts.last_date_for_reg == dt(2018, 1, 1, 7, 0, 0)
# tests fit_df from load_data
result_fit_df = ts.fit_df.reset_index(drop=True) # fit_df's index is x_col
expected_fit_df = df[df[TIME_COL] <= dt(2018, 1, 1, 5, 0, 0)]
assert_frame_equal(result_fit_df, expected_fit_df)
# tests fit_y
result_fit_y = ts.fit_y.reset_index(drop=True) # fit_y's index is x_col
expected_fit_y = expected_fit_df[VALUE_COL]
assert result_fit_y.equals(expected_fit_y)
# with history
result = ts.make_future_dataframe(
periods=2, include_history=True).reset_index(drop=True)
expected = pd.DataFrame({
TIME_COL:
pd.date_range(start=dt(2018, 1, 1, 3, 0, 0), periods=5, freq="H"),
VALUE_COL: [1, None, 3, None, None],
"regressor1": [0.01, None, 0.014, 0.016, 0.017],
"regressor2": [0.11, 0.112, 0.114, 0.116, None],
"regressor3": [0.21, 0.212, 0.214, 0.216, 0.217]
})
assert_frame_equal(result, expected)
# without history
result = ts.make_future_dataframe(
periods=2, include_history=False).reset_index(drop=True)
expected = pd.DataFrame({
TIME_COL:
pd.date_range(start=dt(2018, 1, 1, 6, 0, 0), periods=2, freq="H"),
VALUE_COL:
np.repeat(np.nan, 2),
"regressor1": [0.016, 0.017],
"regressor2": [0.116, None],
"regressor3": [0.216, 0.217]
})
assert result.equals(expected)
# user doesn't request any future periods
result = ts.make_future_dataframe(
periods=None, include_history=False).reset_index(drop=True)
expected = pd.DataFrame({
TIME_COL:
pd.date_range(start=dt(2018, 1, 1, 6, 0, 0), periods=2, freq="H"),
VALUE_COL:
np.repeat(np.nan, 2),
"regressor1": [0.016, 0.017],
"regressor2": [0.116, None],
"regressor3": [0.216, 0.217]
})
assert result.equals(expected)
# user requests fewer than the available periods
result = ts.make_future_dataframe(
periods=1, include_history=False).reset_index(drop=True)
expected = pd.DataFrame({
TIME_COL:
pd.date_range(start=dt(2018, 1, 1, 6, 0, 0), periods=1, freq="H"),
VALUE_COL:
np.repeat(np.nan, 1),
"regressor1": [0.016],
"regressor2": [0.116],
"regressor3": [0.216]
})
assert result.equals(expected)
# user requests more than 2 periods
with pytest.warns(Warning) as record:
result = ts.make_future_dataframe(
periods=4, include_history=False).reset_index(drop=True)
expected = pd.DataFrame({
TIME_COL:
pd.date_range(start=dt(2018, 1, 1, 6, 0, 0), periods=2, freq="H"),
VALUE_COL:
np.repeat(np.nan, 2),
"regressor1": [0.016, 0.017],
"regressor2": [0.116, None],
"regressor3": [0.216, 0.217]
})
assert result.equals(expected)
assert "Provided periods '4' is more than allowed ('2') due to the length of regressor columns. " \
"Using '2'." in record[0].message.args[0]
def test_train_end_date_with_regressors():
"""Tests make_future_dataframe and train_end_date with regressors"""
data = generate_df_with_reg_for_tests(freq="D",
periods=30,
train_start_date=datetime.datetime(
2018, 1, 1),
remove_extra_cols=True,
mask_test_actuals=True)
regressor_cols = ["regressor1", "regressor2", "regressor_categ"]
keep_cols = [TIME_COL, VALUE_COL] + regressor_cols
df = data["df"][keep_cols].copy()
# Setting NaN values at the end
df.loc[df.tail(2).index, "regressor1"] = np.nan
df.loc[df.tail(4).index, "regressor2"] = np.nan
df.loc[df.tail(6).index, "regressor_categ"] = np.nan
df.loc[df.tail(8).index, VALUE_COL] = np.nan
# default train_end_date, default regressor_cols
with pytest.warns(UserWarning) as record:
ts = UnivariateTimeSeries()
ts.load_data(df,
TIME_COL,
VALUE_COL,
train_end_date=None,
regressor_cols=None)
assert f"{ts.original_value_col} column of the provided TimeSeries contains null " \
f"values at the end. Setting 'train_end_date' to the last timestamp with a " \
f"non-null value ({ts.train_end_date})." in record[0].message.args[0]
assert ts.train_end_date == dt(2018, 1, 22)
assert ts.fit_df.shape == (22, 2)
assert ts.last_date_for_val == df[
df[VALUE_COL].notnull()][TIME_COL].max()
assert ts.last_date_for_reg is None
result = ts.make_future_dataframe(periods=10, include_history=True)
expected = pd.DataFrame({
TIME_COL:
pd.date_range(start=dt(2018, 1, 1), periods=32, freq="D"),
VALUE_COL:
np.concatenate([ts.fit_y, np.repeat(np.nan, 10)])
})
expected.index = expected[TIME_COL]
expected.index.name = None
assert_frame_equal(result, expected)
# train_end_date later than last date in df, all available regressor_cols
with pytest.warns(UserWarning) as record:
ts = UnivariateTimeSeries()
train_end_date = dt(2018, 2, 10)
ts.load_data(df,
TIME_COL,
VALUE_COL,
train_end_date=train_end_date,
regressor_cols=regressor_cols)
assert f"Input timestamp for the parameter 'train_end_date' " \
f"({train_end_date}) either exceeds the last available timestamp or" \
f"{VALUE_COL} column of the provided TimeSeries contains null " \
f"values at the end. Setting 'train_end_date' to the last timestamp with a " \
f"non-null value ({ts.train_end_date})." in record[0].message.args[0]
assert ts.last_date_for_val == dt(2018, 1, 22)
assert ts.last_date_for_reg == dt(2018, 1, 28)
result = ts.make_future_dataframe(periods=10, include_history=False)
expected = df.copy()[22:28]
expected.loc[expected.tail(6).index, VALUE_COL] = np.nan
expected.index = expected[TIME_COL]
expected.index.name = None
assert_frame_equal(result, expected)
# train_end_date in between last date in df and last date before null
# user passes no regressor_cols
with pytest.warns(UserWarning) as record:
ts = UnivariateTimeSeries()
train_end_date = dt(2018, 1, 25)
regressor_cols = []
ts.load_data(df,
TIME_COL,
VALUE_COL,
train_end_date=train_end_date,
regressor_cols=regressor_cols)
assert f"Input timestamp for the parameter 'train_end_date' " \
f"({train_end_date}) either exceeds the last available timestamp or" \
f"{VALUE_COL} column of the provided TimeSeries contains null " \
f"values at the end. Setting 'train_end_date' to the last timestamp with a " \
f"non-null value ({ts.train_end_date})." in record[0].message.args[0]
assert ts.train_end_date == dt(2018, 1, 22)
assert ts.last_date_for_reg is None
result = ts.make_future_dataframe(periods=10, include_history=True)
expected = pd.DataFrame({
TIME_COL:
pd.date_range(start=dt(2018, 1, 1), periods=32, freq="D"),
VALUE_COL:
np.concatenate([ts.fit_y, np.repeat(np.nan, 10)])
})
expected.index = expected[TIME_COL]
expected.index.name = None
assert_frame_equal(result, expected)
# train end date equal to last date before null
# user requests a subset of the regressor_cols
with pytest.warns(UserWarning) as record:
ts = UnivariateTimeSeries()
train_end_date = dt(2018, 1, 22)
regressor_cols = ["regressor2"]
ts.load_data(df,
TIME_COL,
VALUE_COL,
train_end_date=train_end_date,
regressor_cols=regressor_cols)
assert ts.train_end_date == dt(2018, 1, 22)
assert ts.last_date_for_reg == dt(2018, 1, 26)
result = ts.make_future_dataframe(periods=10, include_history=True)
assert "Provided periods '10' is more than allowed ('4') due to the length of " \
"regressor columns. Using '4'." in record[0].message.args[0]
expected = ts.df.copy()[[TIME_COL, VALUE_COL, "regressor2"]]
expected = expected[expected.index <= ts.last_date_for_reg]
assert_frame_equal(result, expected)
# train_end_date smaller than last date before null
# user requests regressor_cols that does not exist in df
with pytest.warns(UserWarning) as record:
ts = UnivariateTimeSeries()
train_end_date = dt(2018, 1, 20)
regressor_cols = ["regressor1", "regressor4", "regressor5"]
ts.load_data(df,
TIME_COL,
VALUE_COL,
train_end_date=train_end_date,
regressor_cols=regressor_cols)
assert ts.train_end_date == dt(2018, 1, 20)
assert ts.last_date_for_reg == dt(2018, 1, 28)
assert (f"The following columns are not available to use as "
f"regressors: ['regressor4', 'regressor5']"
) in record[0].message.args[0]
result = ts.make_future_dataframe(periods=10, include_history=True)
expected = ts.df.copy()[[TIME_COL, VALUE_COL, "regressor1"]]
expected = expected[expected.index <= ts.last_date_for_reg]
assert_frame_equal(result, expected)