In [1]:

try:
  import pycaret
except:
  !pip install pycaret-ts-alpha

In [2]:

#### Import libraries ----
from pprint import pprint
from pycaret.datasets import get_data
from pycaret.internal.pycaret_experiment import TimeSeriesExperiment

/usr/local/lib/python3.7/dist-packages/distributed/config.py:20: YAMLLoadWarning: calling yaml.load() without Loader=... is deprecated, as the default Loader is unsafe. Please read https://msg.pyyaml.org/load for full details.
  defaults = yaml.load(f)

In [3]:

#### Get the data ---
y = get_data("airline")

Period
1949-01    112.0
1949-02    118.0
1949-03    132.0
1949-04    129.0
1949-05    121.0
Freq: M, Name: Number of airline passengers, dtype: float64

In [4]:

#### Setup the experiment ----
exp = TimeSeriesExperiment()
exp.setup(data=y, fh=12, seasonal_period=12, session_id=42)

	Description	Value
0	session_id	42
1	Original Data	(144, 1)
2	Missing Values	False
3	Transformed Train Set	(132,)
4	Transformed Test Set	(12,)
5	Fold Generator	ExpandingWindowSplitter
6	Fold Number	3
7	Enforce Prediction Interval	False
8	Seasonal Period Tested	12
9	Seasonality Detected	True
10	Target Strictly Positive	True
11	Target White Noise	No
12	Recommended d	1
13	Recommended Seasonal D	1
14	CPU Jobs	-1
15	Use GPU	False
16	Log Experiment	False
17	Experiment Name	ts-default-name
18	USI	f18d
19	Imputation Type	simple

Out[4]:

<pycaret.internal.pycaret_experiment.time_series_experiment.TimeSeriesExperiment at 0x7f39adfe0f90>

In [5]:

#### Create different types of models ----

# ARIMA model from `pmdarima`
arima_model = exp.create_model("arima")

# ETS and Exponential Smoothing models from `statsmodels`
ets_model = exp.create_model("ets")
exp_smooth_model = exp.create_model("exp_smooth")

# Reduced Regression model using `sklearn` Linear Regression
lr_model = exp.create_model("lr_cds_dt")

	cutoff	MAE	RMSE	MAPE	SMAPE	R2
0	1956-12	38.6824	45.0820	0.0998	0.1051	0.3384
1	1957-12	28.0608	34.6867	0.0751	0.0734	0.6848
2	1958-12	32.1693	38.2681	0.0737	0.0753	0.6724
Mean	NaN	32.9708	39.3456	0.0828	0.0846	0.5652
SD	NaN	4.3731	4.3117	0.0120	0.0145	0.1604

In [6]:

#### Check model types ----
print(type(arima_model))        # <-- sktime `pmdarima` adapter 
print(type(ets_model))          # <-- sktime `statsmodels` adapter
print(type(exp_smooth_model))   # <-- sktime `statsmodels` adapter
print(type(lr_model))           # <-- Your custom sktime compatible model pipeline

<class 'sktime.forecasting.arima.ARIMA'>
<class 'sktime.forecasting.ets.AutoETS'>
<class 'sktime.forecasting.exp_smoothing.ExponentialSmoothing'>
<class 'pycaret.containers.models.time_series.BaseCdsDtForecaster'>

In [7]:

#### Access internal models using `_forecaster` ----
print(type(arima_model._forecaster))
print(type(ets_model._forecaster))
print(type(exp_smooth_model._forecaster))
print(type(lr_model._forecaster))

<class 'pmdarima.arima.arima.ARIMA'>
<class 'statsmodels.tsa.exponential_smoothing.ets.ETSModel'>
<class 'statsmodels.tsa.holtwinters.model.ExponentialSmoothing'>
<class 'sktime.forecasting.compose._pipeline.TransformedTargetForecaster'>

In [8]:

#### What hyperparameters were used to train the model? ----
print(arima_model)

ARIMA(maxiter=50, method='lbfgs', order=(1, 0, 0), out_of_sample_size=0,
      scoring='mse', scoring_args=None, seasonal_order=(0, 1, 0, 12),
      start_params=None, suppress_warnings=False, trend=None,
      with_intercept=True)

In [9]:

#### Access statistical fit properties using underlying `pmdarima`
arima_model._forecaster.summary()

#### Alternately, use sktime's convenient wrapper to do so ---- 
arima_model.summary()

Out[9]:

SARIMAX Results
Dep. Variable:	y	No. Observations:	132
Model:	SARIMAX(1, 0, 0)x(0, 1, 0, 12)	Log Likelihood	-450.590
Date:	Tue, 16 Nov 2021	AIC	907.180
Time:	11:26:51	BIC	915.542
Sample:	0	HQIC	910.576
	- 132
Covariance Type:	opg

	coef	std err	z	P>\|z\|	[0.025	0.975]
intercept	5.7982	2.005	2.892	0.004	1.869	9.727
ar.L1	0.8100	0.061	13.261	0.000	0.690	0.930
sigma2	105.9407	12.533	8.453	0.000	81.377	130.505

Ljung-Box (L1) (Q):	2.30	Jarque-Bera (JB):	1.04
Prob(Q):	0.13	Prob(JB):	0.60
Heteroskedasticity (H):	1.34	Skew:	-0.07
Prob(H) (two-sided):	0.36	Kurtosis:	3.43

Warnings:
[1] Covariance matrix calculated using the outer product of gradients (complex-step).

You can not starts correlating these properties to the forecasts that you see. I will write about it in a subsewquent post.

In [10]:

#### What hyperparameters were used to train the model? ----
print(ets_model)

AutoETS(additive_only=False, allow_multiplicative_trend=False, auto=False,
        bounds=None, callback=None, damped_trend=False, dates=None, disp=False,
        error='add', freq=None, full_output=True, ignore_inf_ic=True,
        information_criterion='aic', initial_level=None, initial_seasonal=None,
        initial_trend=None, initialization_method='estimated', maxiter=1000,
        missing='none', n_jobs=None, restrict=True, return_params=False,
        seasonal='mul', sp=12, start_params=None, trend='add')

In [11]:

#### Access statsitical fit properties using underlying statsmodel
ets_model._forecaster.fit().summary()

#### Alternatively, use sktime's convenient wrapper to do so ---- 
ets_model.summary()

Out[11]:

ETS Results
Dep. Variable:	Number of airline passengers	No. Observations:	132
Model:	ETS(AAM)	Log Likelihood	-488.626
Date:	Tue, 16 Nov 2021	AIC	1013.253
Time:	11:26:51	BIC	1065.143
Sample:	01-31-1949	HQIC	1034.339
	- 12-31-1959	Scale	96.116
Covariance Type:	approx

	coef	std err	z	P>\|z\|	[0.025	0.975]
smoothing_level	0.3734	0.067	5.550	0.000	0.242	0.505
smoothing_trend	3.734e-05	nan	nan	nan	nan	nan
smoothing_seasonal	0.6265	0.067	9.296	0.000	0.494	0.759
initial_level	109.3470	nan	nan	nan	nan	nan
initial_trend	2.6555	nan	nan	nan	nan	nan
initial_seasonal.0	0.9773	nan	nan	nan	nan	nan
initial_seasonal.1	0.8482	nan	nan	nan	nan	nan
initial_seasonal.2	0.9508	nan	nan	nan	nan	nan
initial_seasonal.3	1.0885	nan	nan	nan	nan	nan
initial_seasonal.4	1.1927	nan	nan	nan	nan	nan
initial_seasonal.5	1.2076	nan	nan	nan	nan	nan
initial_seasonal.6	1.1092	nan	nan	nan	nan	nan
initial_seasonal.7	1.0129	nan	nan	nan	nan	nan
initial_seasonal.8	1.0970	nan	nan	nan	nan	nan
initial_seasonal.9	1.1541	nan	nan	nan	nan	nan
initial_seasonal.10	1.0517	nan	nan	nan	nan	nan
initial_seasonal.11	1.0000	nan	nan	nan	nan	nan

Ljung-Box (Q):	41.34	Jarque-Bera (JB):	1.25
Prob(Q):	0.02	Prob(JB):	0.54
Heteroskedasticity (H):	2.21	Skew:	0.11
Prob(H) (two-sided):	0.01	Kurtosis:	3.42

Warnings:
[1] Covariance matrix calculated using numerical (complex-step) differentiation.

In [12]:

#### sktime pipelines are similar to sklearn.
#### Access steps using `named_steps` attribute
print(lr_model._forecaster.named_steps.keys(), "\n\n")

#### Details about the steps ----
pprint(lr_model._forecaster.named_steps)

dict_keys(['conditional_deseasonalise', 'detrend', 'forecast']) 


{'conditional_deseasonalise': ConditionalDeseasonalizer(model='additive', seasonality_test=None, sp=1),
 'detrend': Detrender(forecaster=PolynomialTrendForecaster(degree=1, regressor=None,
                                               with_intercept=True)),
 'forecast': RecursiveTabularRegressionForecaster(estimator=LinearRegression(copy_X=True,
                                                                fit_intercept=True,
                                                                n_jobs=-1,
                                                                normalize=False,
                                                                positive=False),
                                     window_length=10)}