def test_AutoARIMA_class():
train, test = wineind[:125], wineind[125:]
mod = AutoARIMA(maxiter=5)
mod.fit(train)
endog = mod.model_.arima_res_.data.endog
assert_array_almost_equal(train, endog)
# update
mod.update(test, maxiter=2)
new_endog = mod.model_.arima_res_.data.endog
assert_array_almost_equal(wineind, new_endog)
def test_pipeline_behavior():
pipeline = Pipeline([("fourier", FourierFeaturizer(m=12)),
("boxcox", BoxCoxEndogTransformer()),
("arima",
AutoARIMA(seasonal=False,
stepwise=True,
suppress_warnings=True,
d=1,
max_p=2,
max_q=0,
start_q=0,
start_p=1,
maxiter=3,
error_action='ignore'))])
# Quick assertions on indexing
assert len(pipeline) == 3
pipeline.fit(train)
preds = pipeline.predict(5)
assert preds.shape[0] == 5
assert pipeline._final_estimator.model_.fit_with_exog_
# Assert that when the n_periods kwarg is set manually and incorrectly for
# the fourier transformer, we get a ValueError
kwargs = {"fourier__n_periods": 10}
with pytest.raises(ValueError) as ve:
pipeline.predict(3, **kwargs)
assert "'n_periods'" in pytest_error_str(ve)
# Assert that we can update the model
pipeline.update(test, maxiter=5)
# And that the fourier transformer was updated properly...
assert pipeline.steps_[0][1].n_ == wineind.shape[0]
class AutoArima:
def __init__(self, args):
self.model = AutoARIMA()
self.seq_len_x = args.seq_len_x
self.out_seq_len = args.out_seq_len
self.args = args
def predict(self, x):
# input [batch, in_seq_len, n]
b, seq_x, n = x.shape
x = np.reshape(x, [-1, seq_x])
n_samples, _ = x.shape
xhat = []
for i in range(n_samples):
y = self.model.fit_predict(x[i], n_periods=self.out_seq_len)
xhat.append(y)
xhat = np.stack(xhat, axis=-1)
xhat = np.reshape(xhat, (b, self.out_seq_len, n))
return xhat # (b, out_len, n)
def test_arima_setup(params, X):
"""Checks if parameters are passed to Auto-Arima correctly"""
coverage = 0.99
model = AutoArimaEstimator(score_func=mean_squared_error,
coverage=coverage,
null_model_params=None,
**params)
# set_params must be able to replicate the init
model2 = AutoArimaEstimator()
model2.set_params(**dict(score_func=mean_squared_error,
coverage=coverage,
null_model_params=None,
**params))
assert model2.__dict__ == model.__dict__
model.fit(X)
direct_model = AutoARIMA(**params)
model_params = model.model.__dict__
direct_model_params = direct_model.__dict__
assert model_params["start_p"] == direct_model_params["start_p"]
assert model_params["d"] == direct_model_params["d"]
assert model_params["start_q"] == direct_model_params["start_q"]
assert model_params["max_p"] == direct_model_params["max_p"]
assert model_params["max_d"] == direct_model_params["max_d"]
assert model_params["max_q"] == direct_model_params["max_q"]
assert model_params["start_P"] == direct_model_params["start_P"]
assert model_params["D"] == direct_model_params["D"]
assert model_params["start_Q"] == direct_model_params["start_Q"]
assert model_params["max_P"] == direct_model_params["max_P"]
assert model_params["max_D"] == direct_model_params["max_D"]
assert model_params["max_Q"] == direct_model_params["max_Q"]
assert model_params["max_order"] == direct_model_params["max_order"]
assert model_params["m"] == direct_model_params["m"]
assert model_params["seasonal"] == direct_model_params["seasonal"]
assert model_params["stationary"] == direct_model_params["stationary"]
assert model_params["information_criterion"] == direct_model_params[
"information_criterion"]
assert model_params["alpha"] == direct_model_params["alpha"]
assert model_params["test"] == direct_model_params["test"]
assert model_params["seasonal_test"] == direct_model_params[
"seasonal_test"]
assert model_params["stepwise"] == direct_model_params["stepwise"]
assert model_params["n_jobs"] == direct_model_params["n_jobs"]
assert model_params["start_params"] == direct_model_params["start_params"]
assert model_params["trend"] == direct_model_params["trend"]
assert model_params["method"] == direct_model_params["method"]
assert model_params["maxiter"] == direct_model_params["maxiter"]
assert model_params["offset_test_args"] == direct_model_params[
"offset_test_args"]
assert model_params["seasonal_test_args"] == direct_model_params[
"seasonal_test_args"]
assert model_params["suppress_warnings"] == direct_model_params[
"suppress_warnings"]
assert model_params["error_action"] == direct_model_params["error_action"]
assert model_params["trace"] == direct_model_params["trace"]
assert model_params["random"] == direct_model_params["random"]
assert model_params["random_state"] == direct_model_params["random_state"]
assert model_params["n_fits"] == direct_model_params["n_fits"]
assert model_params["out_of_sample_size"] == direct_model_params[
"out_of_sample_size"]
assert model_params["scoring"] == direct_model_params["scoring"]
assert model_params["scoring_args"] == direct_model_params["scoring_args"]
assert model_params["with_intercept"] == direct_model_params[
"with_intercept"]
assert model_params["kwargs"] == direct_model_params["kwargs"]
]
)
def test_bad_last_stage(self, stages):
# Will fail since the last stage is not an estimator
with pytest.raises(TypeError) as ve:
Pipeline(stages)
assert "Last step of Pipeline should be" in pytest_error_str(ve)
@pytest.mark.parametrize(
'pipe,kwargs,expected', [
pytest.param(
Pipeline([
("boxcox", BoxCoxEndogTransformer()),
("arima", AutoARIMA())
]),
{},
{"boxcox": {}, "arima": {}}
),
pytest.param(
Pipeline([
("boxcox", BoxCoxEndogTransformer()),
("arima", AutoARIMA())
]),
{"boxcox__lmdba1": 0.001},
{"boxcox": {"lmdba1": 0.001}, "arima": {}}
),
]
)
# Two transformers
[("stage1", BoxCoxEndogTransformer()),
("stage2", FourierFeaturizer(m=12))]
])
def test_bad_last_stage(self, stages):
# Will fail since the last stage is not an estimator
with pytest.raises(TypeError) as ve:
Pipeline(stages)
assert "Last step of Pipeline should be" in pytest_error_str(ve)
@pytest.mark.parametrize('pipe,kwargs,expected', [
pytest.param(
Pipeline([("boxcox", BoxCoxEndogTransformer()),
("arima", AutoARIMA())]), {}, {
"boxcox": {},
"arima": {}
}),
pytest.param(
Pipeline([("boxcox", BoxCoxEndogTransformer()),
("arima", AutoARIMA())]), {"boxcox__lmdba1": 0.001}, {
"boxcox": {
"lmdba1": 0.001
},
"arima": {}
}),
])
def test_get_kwargs(pipe, kwargs, expected):
# Test we get the kwargs we expect
kw = pipe._get_kwargs(**kwargs)
# -*- coding: utf-8 -*-
from sklearn.base import clone
from pmdarima.arima import ARIMA, AutoARIMA
from pmdarima.pipeline import Pipeline
from pmdarima.datasets import load_wineind
from pmdarima.preprocessing import FourierFeaturizer
import pytest
y = load_wineind()
@pytest.mark.parametrize(
'est', [
ARIMA(order=(2, 1, 1), seasonal_order=(0, 0, 0, 1)),
AutoARIMA(seasonal=False, maxiter=3),
Pipeline([
("fourier", FourierFeaturizer(m=12)),
("arima", AutoARIMA(seasonal=False, stepwise=True,
suppress_warnings=True, d=1, max_p=2, max_q=0,
start_q=0, start_p=1,
maxiter=3, error_action='ignore'))
])
]
)
def test_clonable(est):
# fit it, then clone it
est.fit(y)
est2 = clone(est)
assert isinstance(est2, est.__class__)
assert est is not est2
def __init__(self, args):
self.model = AutoARIMA()
self.seq_len_x = args.seq_len_x
self.out_seq_len = args.out_seq_len
self.args = args
class AutoArimaEstimator(BaseForecastEstimator):
"""Wrapper for ``pmdarima.arima.AutoARIMA``.
It currently does not handle the regressor issue when there is
gap between train and predict periods.
Parameters
----------
score_func : callable
see ``BaseForecastEstimator``.
coverage : float between [0.0, 1.0]
see ``BaseForecastEstimator``.
null_model_params : dict with arguments to define DummyRegressor null model, optional, default=None
see ``BaseForecastEstimator``.
regressor_cols: `list` [`str`], optional, default None
A list of regressor columns used during training and prediction.
If None, no regressor columns are used.
See ``AutoArima`` documentation for rest of the parameter descriptions:
* https://alkaline-ml.com/pmdarima/modules/generated/pmdarima.arima.AutoARIMA.html#pmdarima.arima.AutoARIMA
Attributes
----------
model : ``AutoArima`` object
Auto arima model object
fit_df : `pandas.DataFrame` or None
The training data used to fit the model.
forecast : `pandas.DataFrame`
Output of the predict method of ``AutoArima``.
"""
def __init__(
self,
# Null model parameters
score_func: callable = mean_squared_error,
coverage: float = 0.90,
null_model_params: Optional[Dict] = None,
# Additional parameters
regressor_cols: Optional[List[str]] = None,
freq: Optional[float] = None,
# pmdarima fit parameters
start_p: Optional[int] = 2,
d: Optional[int] = None,
start_q: Optional[int] = 2,
max_p: Optional[int] = 5,
max_d: Optional[int] = 2,
max_q: Optional[int] = 5,
start_P: Optional[int] = 1,
D: Optional[int] = None,
start_Q: Optional[int] = 1,
max_P: Optional[int] = 2,
max_D: Optional[int] = 1,
max_Q: Optional[int] = 2,
max_order: Optional[int] = 5,
m: Optional[int] = 1,
seasonal: Optional[bool] = True,
stationary: Optional[bool] = False,
information_criterion: Optional[str] = 'aic',
alpha: Optional[int] = 0.05,
test: Optional[str] = 'kpss',
seasonal_test: Optional[str] = 'ocsb',
stepwise: Optional[bool] = True,
n_jobs: Optional[int] = 1,
start_params: Optional[Dict] = None,
trend: Optional[str] = None,
method: Optional[str] = 'lbfgs',
maxiter: Optional[int] = 50,
offset_test_args: Optional[Dict] = None,
seasonal_test_args: Optional[Dict] = None,
suppress_warnings: Optional[bool] = True,
error_action: Optional[str] = 'trace',
trace: Optional[Union[int, bool]] = False,
random: Optional[bool] = False,
random_state: Optional[Union[int, callable]] = None,
n_fits: Optional[int] = 10,
out_of_sample_size: Optional[int] = 0,
scoring: Optional[str] = 'mse',
scoring_args: Optional[Dict] = None,
with_intercept: Optional[Union[bool, str]] = "auto",
# pmdarima predict parameters
return_conf_int: Optional[bool] = True,
dynamic: Optional[bool] = False):
# Every subclass of BaseForecastEstimator must call super().__init__
super().__init__(
score_func=score_func,
coverage=coverage,
null_model_params=null_model_params)
self.regressor_cols = regressor_cols
self.freq = freq
self.start_p = start_p
self.d = d
self.start_q = start_q
self.max_p = max_p
self.max_d = max_d
self.max_q = max_q
self.start_P = start_P
self.D = D
self.start_Q = start_Q
self.max_P = max_P
self.max_D = max_D
self.max_Q = max_Q
self.max_order = max_order
self.m = m
self.seasonal = seasonal
self.stationary = stationary
self.information_criterion = information_criterion
self.alpha = alpha
self.test = test
self.seasonal_test = seasonal_test
self.stepwise = stepwise
self.n_jobs = n_jobs
self.start_params = start_params
self.trend = trend
self.method = method
self.maxiter = maxiter
self.offset_test_args = offset_test_args
self.seasonal_test_args = seasonal_test_args
self.suppress_warnings = suppress_warnings
self.error_action = error_action
self.trace = trace
self.random = random
self.random_state = random_state
self.n_fits = n_fits
self.out_of_sample_size = out_of_sample_size
self.scoring = scoring
self.scoring_args = scoring_args
self.with_intercept = with_intercept
self.return_conf_int = return_conf_int
self.coverage = coverage
self.dynamic = dynamic
# set by the fit method
self.model = None
self.fit_df = None
# set by the predict method
self.forecast = None
def fit(self, X, y=None, time_col=TIME_COL, value_col=VALUE_COL, **fit_params):
"""Fits ``ARIMA`` forecast model.
Parameters
----------
X : `pandas.DataFrame`
Input timeseries, with timestamp column,
value column, and any additional regressors.
The value column is the response, included in
X to allow transformation by `sklearn.pipeline.Pipeline`
y : ignored
The original timeseries values, ignored.
(The y for fitting is included in ``X``.)
time_col : `str`
Time column name in ``X``
value_col : `str`
Value column name in ``X``
fit_params : `dict`
additional parameters for null model
Returns
-------
self : self
Fitted model is stored in ``self.model``.
"""
X = X.sort_values(by=time_col)
# fits null model
super().fit(X, y=y, time_col=time_col, value_col=value_col, **fit_params)
self.fit_df = X
# fits AutoArima model
self.model = AutoARIMA(
start_p=self.start_p,
d=self.d,
start_q=self.start_q,
max_p=self.max_p,
max_d=self.max_d,
max_q=self.max_q,
start_P=self.start_P,
D=self.D,
start_Q=self.start_Q,
max_P=self.max_P,
max_D=self.max_D,
max_Q=self.max_Q,
max_order=self.max_order,
m=self.m,
seasonal=self.seasonal,
stationary=self.stationary,
information_criterion=self.information_criterion,
alpha=self.alpha,
test=self.test,
seasonal_test=self.seasonal_test,
stepwise=self.stepwise,
n_jobs=self.n_jobs,
start_params=self.start_params,
trend=self.trend,
method=self.method,
maxiter=self.maxiter,
offset_test_args=self.offset_test_args,
seasonal_test_args=self.seasonal_test_args,
suppress_warnings=self.suppress_warnings,
error_action=self.error_action,
trace=self.trace,
random=self.random,
random_state=self.random_state,
n_fits=self.n_fits,
out_of_sample_size=self.out_of_sample_size,
scoring=self.scoring,
scoring_args=self.scoring_args,
with_intercept=self.with_intercept,
return_conf_int=self.return_conf_int,
dynamic=self.dynamic,
regressor_cols=self.regressor_cols
)
# fits auto-arima
if self.regressor_cols is None:
reg_df = None
else:
reg_df = X[self.regressor_cols]
self.model.fit(y=X[[value_col]], X=reg_df)
return self
def predict(self, X, y=None):
"""Creates forecast for the dates specified in ``X``.
Currently does not support the regressor case where there is gap between
train and predict periods.
Parameters
----------
X: `pandas.DataFrame`
Input timeseries with timestamp column and any additional regressors.
Timestamps are the dates for prediction.
Value column, if provided in ``X``, is ignored.
y: ignored.
Returns
-------
predictions: `pandas.DataFrame`
Forecasted values for the dates in ``X``. Columns:
- ``TIME_COL``: dates
- ``PREDICTED_COL``: predictions
- ``PREDICTED_LOWER_COL``: lower bound of predictions
- ``PREDICTED_UPPER_COL``: upper bound of predictions
"""
X = X.sort_values(by=self.time_col_)
# Returns the cached result if applicable
cached_predictions = super().predict(X=X)
if cached_predictions is not None:
return cached_predictions
# Currently does not support the regressor case where
# there is gap between train and predict periods
if self.regressor_cols is None:
fut_reg_df = None
else:
fut_df = X[X[self.time_col_] > self.fit_df[self.time_col_].iloc[-1]]
fut_reg_df = fut_df[self.regressor_cols] # Auto-arima only accepts regressor values beyond `fit_df`
if self.freq is None:
self.freq = pd.infer_freq(self.fit_df[self.time_col_])
if self.freq == "H":
self.freq = self.freq.lower() # np.timedelta recognizes lower case letters
chosen_d = self.model.model_.order[1] # This is the value of the d chosen by auto-arima
forecast_start = int((X[self.time_col_].iloc[0] - self.fit_df[self.time_col_].iloc[0])/np.timedelta64(1, self.freq))
if forecast_start < chosen_d:
append_length = chosen_d - forecast_start # Number of NaNs to append to `pred_df`
forecast_start = chosen_d # Auto-arima can not predict below the chosen d
else:
append_length = 0
forecast_end = int((X[self.time_col_].iloc[-1] - self.fit_df[self.time_col_].iloc[0])/np.timedelta64(1, self.freq))
predictions = self.model.predict_in_sample(
X=fut_reg_df,
start=forecast_start,
end=forecast_end,
dynamic=self.dynamic,
return_conf_int=self.return_conf_int,
alpha=(1-self.coverage)
)
if append_length > 0:
pred_df = pd.DataFrame({
TIME_COL: X[self.time_col_],
PREDICTED_COL: np.append(np.repeat(np.nan, append_length), predictions[0]),
PREDICTED_LOWER_COL: np.append(np.repeat(np.nan, append_length), predictions[1][:, 0]),
PREDICTED_UPPER_COL: np.append(np.repeat(np.nan, append_length), predictions[1][:, 1])
})
else:
pred_df = pd.DataFrame({
TIME_COL: X[self.time_col_],
PREDICTED_COL: predictions[0],
PREDICTED_LOWER_COL: predictions[1][:, 0],
PREDICTED_UPPER_COL: predictions[1][:, 1]
})
self.forecast = pred_df
# Caches the predictions
self.cached_predictions_ = pred_df
return pred_df
def summary(self):
BaseForecastEstimator.summary(self)
# AutoArima summary
return self.model.summary()
def fit(self, X, y=None, time_col=TIME_COL, value_col=VALUE_COL, **fit_params):
"""Fits ``ARIMA`` forecast model.
Parameters
----------
X : `pandas.DataFrame`
Input timeseries, with timestamp column,
value column, and any additional regressors.
The value column is the response, included in
X to allow transformation by `sklearn.pipeline.Pipeline`
y : ignored
The original timeseries values, ignored.
(The y for fitting is included in ``X``.)
time_col : `str`
Time column name in ``X``
value_col : `str`
Value column name in ``X``
fit_params : `dict`
additional parameters for null model
Returns
-------
self : self
Fitted model is stored in ``self.model``.
"""
X = X.sort_values(by=time_col)
# fits null model
super().fit(X, y=y, time_col=time_col, value_col=value_col, **fit_params)
self.fit_df = X
# fits AutoArima model
self.model = AutoARIMA(
start_p=self.start_p,
d=self.d,
start_q=self.start_q,
max_p=self.max_p,
max_d=self.max_d,
max_q=self.max_q,
start_P=self.start_P,
D=self.D,
start_Q=self.start_Q,
max_P=self.max_P,
max_D=self.max_D,
max_Q=self.max_Q,
max_order=self.max_order,
m=self.m,
seasonal=self.seasonal,
stationary=self.stationary,
information_criterion=self.information_criterion,
alpha=self.alpha,
test=self.test,
seasonal_test=self.seasonal_test,
stepwise=self.stepwise,
n_jobs=self.n_jobs,
start_params=self.start_params,
trend=self.trend,
method=self.method,
maxiter=self.maxiter,
offset_test_args=self.offset_test_args,
seasonal_test_args=self.seasonal_test_args,
suppress_warnings=self.suppress_warnings,
error_action=self.error_action,
trace=self.trace,
random=self.random,
random_state=self.random_state,
n_fits=self.n_fits,
out_of_sample_size=self.out_of_sample_size,
scoring=self.scoring,
scoring_args=self.scoring_args,
with_intercept=self.with_intercept,
return_conf_int=self.return_conf_int,
dynamic=self.dynamic,
regressor_cols=self.regressor_cols
)
# fits auto-arima
if self.regressor_cols is None:
reg_df = None
else:
reg_df = X[self.regressor_cols]
self.model.fit(y=X[[value_col]], X=reg_df)
return self
