AutoARIMA

class hana_ml.algorithms.pal.tsa.auto_arima.AutoARIMA(seasonal_period=None, seasonality_criterion=None, d=None, kpss_significance_level=None, max_d=None, seasonal_d=None, ch_significance_level=None, max_seasonal_d=None, max_p=None, max_q=None, max_seasonal_p=None, max_seasonal_q=None, information_criterion=None, search_strategy=None, max_order=None, initial_p=None, initial_q=None, initial_seasonal_p=None, initial_seasonal_q=None, guess_states=None, max_search_iterations=None, method=None, allow_linear=None, forecast_method=None, output_fitted=None, thread_ratio=None, background_size=None, massive=False, group_params=None)

The ARIMA model, a potent tool in time series analysis, can be challenging due to the difficulty in selecting suitable parameters. AutoARIMA automates this selection process. This model includes seven parameters (p, d, q, P, D, Q, and m), where the seasonality (m) can be estimated using the seasonal_decompose() function, and 'd' and 'D' are usually determined first due to information criterion considerations. The optimal values of p, q, P, Q are obtained through two main methods: 'exhaustive search,' which tests all possible combinations but can be time-consuming, and 'stepwise search,' which is more efficient but may not yield the optimal result. The constant part's inclusion depends on the criterion information, mainly when d + D isn't more than 1.

Parameters:

seasonal_periodint, optional

Value of the seasonal period.

Negative: Automatically identify seasonality by means of an auto-correlation scheme.
0 or 1: Non-seasonal.
Others: Seasonal period.

Defaults to -1.

seasonality_criterionfloat, optional

The criterion of the auto-correlation coefficient for accepting seasonality, in the range of (0, 1). The larger it is, the less probable a time series is regarded to be seasonal. Valid only when seasonal_period is negative.

Defaults to 0.2.

Dint, optional

Order of first-differencing.

Others: Uses the specified value as the first-differencing order.
Negative: Automatically identifies the first-differencing order with the KPSS test.

Defaults to -1.

kpss_significance_levelfloat, optional

The significance level for the KPSS test. Supported values are 0.01, 0.025, 0.05, and 0.1. The smaller it is, the more probable a time series is considered as first-stationary, that is, the less probable it needs first-differencing. Valid only when D is negative.

Defaults to 0.05.

max_dint, optional

The maximum value of D when the KPSS test is applied.

Defaults to 2.

seasonal_dint, optional

Order of seasonal-differencing.

Negative: Automatically identifies seasonal-differencing order using the Canova-Hansen test.

Others: Uses the specified value as the seasonal-differencing order.

Defaults to -1.

ch_significance_levelfloat, optional

The significance level for the Canova-Hansen test. Supported values are 0.01, 0.025, 0.05, 0.1, and 0.2. The smaller it is, the more probable a time series is considered seasonal-stationary; that is, the less probable it needs seasonal-differencing.

Valid only when seasonal_d is negative.

Defaults to 0.05.

max_seasonal_dint, optional

The maximum value of seasonal_d when the Canova-Hansen test is applied.

Defaults to 1.

max_pint, optional

The maximum value of AR order p.

Defaults to 5.

max_qint, optional

The maximum value of MA order q.

Defaults to 5.

max_seasonal_pint, optional

The maximum value of SAR order P.

Defaults to 2.

max_seasonal_qint, optional

The maximum value of SMA order Q.

Defaults to 2.

information_criterion{'aicc', 'aic', 'bic'}, optional

The information criterion for order selection.

'aicc': Akaike information criterion with correction (for small sample sizes)
'aic': Akaike information criterion
'bic': Bayesian information criterion

Defaults to 'aicc'.

search_strategy{'exhaustive', 'stepwise'}, optional

Specifies the search strategy for the optimal ARMA model.

'exhaustive': exhaustive traverse.
'stepwise': stepwise traverse.

Defaults to 'stepwise'.

max_orderint, optional

The maximum value of (max_p + max_q + max_seasonal_p + max_seasonal_q). Valid only when search_strategy is 'exhaustive'.

Defaults to 15.

initial_pint, optional

Order p of the user-defined initial model. Valid only when search_strategy is 'stepwise'.

Defaults to 0.

initial_qint, optional

Order q of the user-defined initial model. Valid only when search_strategy is 'stepwise'.

Defaults to 0.

initial_seasonal_pint, optional

Order seasonal_p of the user-defined initial model. Valid only when search_strategy is 'stepwise'.

Defaults to 0.

initial_seasonal_qint, optional

Order seasonal_q of the user-defined initial model. Valid only when search_strategy is 'stepwise'.

Defaults to 0.

guess_statesint, optional

If employing ACF/PACF to guess initial ARMA models, besides the user-defined model:

0: No guess. Besides the user-defined model, uses states (2, 2) (1, 1)m, (1, 0) (1, 0)m, and (0, 1) (0, 1)m meanwhile as starting states.

1: Guesses starting states taking advantage of ACF/PACF.

Valid only when search_strategy is 'stepwise'.

Defaults to 1.

max_search_iterationsint, optional

The maximum iterations for searching optimal ARMA states.

Valid only when search_strategy is 'stepwise'.

Defaults to (max_p + 1) * (max_q + 1) * (max_seasonal_p + 1) * (max_seasonal_q + 1).

method{'css', 'mle', 'css-mle'}, optional

The objective function for numeric optimization

'css': use the conditional sum of squares.
'mle': use the maximized likelihood estimation.
'css-mle': use css to approximate starting values first and then mle to fit.

Defaults to 'css-mle'.

allow_linearbool, optional

Controls whether to check the linear model ARMA(0,0)(0,0)m.

Defaults to True.

forecast_method{'formula_forecast', 'innovations_algorithm'}, optional

Store information for the subsequent forecast method.

'formula_forecast': compute future series via formula.
'innovations_algorithm': apply the innovations algorithm to compute future series, which requires more original information to be stored.

Defaults to 'innovations_algorithm'.

output_fittedbool, optional

Output fitted result and residuals if True.

Defaults to True.

thread_ratiofloat, optional

Adjusts the percentage of available threads to use, from 0 to 1. A value of 0 indicates the use of a single thread, while 1 implies the use of all possible current threads. Values outside the range will be ignored, and this function heuristically determines the number of threads to use.

Defaults to -1.

background_sizeint, optional

Indicates the number of data points used in ARIMA with explanations in the predict function. If you want to use the ARIMA with explanations, you must set background_size to be a positive value or -1 (auto mode) when initializing an ARIMA instance and then set show_explainer=True in the predict function.

Defaults to NULL (no explanations).

massivebool, optional

Specifies whether or not to activate massive mode.

True : massive mode.
False : single mode.

For parameter setting in massive mode, you could use both group_params (please see the example below) or the original parameters. Using original parameters will apply for all groups. However, if you define some parameters of a group, the value of all original parameter settings will not be applicable to such a group.

An example is as follows:

In this example, as a parameter 'output_fitted' is set in group_params for Group_1 & Group_2, parameter setting of 'background_size' is not applicable to Group_1 & Group_2.

Defaults to False.

group_paramsdict, optional

If massive mode is activated (massive is True), input data is divided into different groups with different parameters applied.

An example with group_params is as follows:

Valid only when massive is True and defaults to None.

Attributes:

model_DataFrame

Model content.

fitted_: DataFrame

Fitted values and residuals.

explainer_DataFrame

The decomposition of trend, seasonal, transitory, irregular and reason code of exogenous variables. Only contains value after show_explainer=True in the predict() function.

permutation_importance_DataFrame

The importance of exogenous variables as determined by permutation importance analysis. The attribute only appears when invoking the get_permutation_importance() function after a trained model is obtained, structured as follows:

1st column : PAIR, measure name.

2nd column : NAME, exogenous regressor name.

3rd column : VALUE, the importance of the exogenous regressor.

Methods

`fit`(data[, key, endog, exog, group_key, ...])	Fit the model to the training dataset.
`get_permutation_importance`(data[, model, ...])	Please see Permutation Feature Importance for Time Series for details.
`predict`([data, key, group_key, ...])	Generates time series forecasts based on the fitted model.
`set_conn`(connection_context)	Set connection context for an ARIMA instance.

Examples

Create an AutoARIMA instance:

>>> autoarima = AutoARIMA(search_strategy='stepwise', allow_linear=True, thread_ratio=1.0)

Perform fit():

>>> autoarima.fit(data=df)

Output:

>>> autoarima.model_.collect()
>>> autoarima.fitted_.collect()

Perform predict():

>>> result = autoarima.predict(forecast_method='innovations_algorithm', forecast_length=10)
>>> result.collect()

If you want to see the decomposed result of the predict result, you could set show_explainer = True:

>>> result = autoarima.predict(forecast_method='innovations_algorithm',
                               forecast_length=10,
                               allow_new_index=False,
                               show_explainer=True)

Show the attribute explainer_ of the AutoARIMA instance:

>>> autoarima.explainer_.collect()

fit(data, key=None, endog=None, exog=None, group_key=None, group_params=None, categorical_variable=None)

Fit the model to the training dataset.

Parameters:

dataDataFrame

Input data which at least has two columns: key and endog.

We also support ARIMAX which needs external data (exogenous variables).

keystr, optional

The timestamp column of data. The type of key column should be INTEGER, TIMESTAMP, DATE or SECONDDATE.

In single mode, defaults to the first column of data if the index column of data is not provided. Otherwise, defaults to the index column of data.

In massive mode, defaults to the first non-group key column of data if the index columns of data are not provided. Otherwise, defaults to the second of index columns of data and the first column of index columns is group_key.

endogstr, optional

The endogenous variable, i.e., time series. The type of endog column should be INTEGER, DOUBLE or DECIMAL(p,s).

In single mode, defaults to the first non-key column. In massive mode, defaults to the first non-group_key, non-key column.

exoglist of str, optional

An optional array of exogenous variables. The type of exog column should be INTEGER, DOUBLE or DECIMAL(p,s).

Valid only for Auto ARIMAX.

Defaults to None. Please set this parameter explicitly if you have exogenous variables.

group_keystr, optional

The column of group_key. Data type can be INT or NVARCHAR/VARCHAR. If data type is INT, only parameters set in the group_params are valid.

This parameter is valid only when massive mode is activated (i.e., parameter massive is set as True in class instance initialization).

Defaults to the first column of data if the index columns of data are not provided. Otherwise, defaults to the first column of index columns.

group_paramsdict, optional

If massive mode is activated (massive is set to True in class instance initialization), input data is divided into different groups with different parameters applied.

An example with group_params is as follows:

Valid only when massive is True.

categorical_variablestr or list of str, optional

Specifies INTEGER columns that should be treated as categorical. Other INTEGER columns will be treated as continuous.

Defaults to None.

Returns:

A fitted object of class "AutoARIMA".

get_permutation_importance(data, model=None, key=None, endog=None, exog=None, repeat_time=None, random_state=None, thread_ratio=None, partition_ratio=None, regressor_top_k=None, accuracy_measure=None, ignore_zero=None): Please see Permutation Feature Importance for Time Series for details.

predict(data=None, key=None, group_key=None, group_params=None, forecast_method=None, forecast_length=None, allow_new_index=False, show_explainer=False, thread_ratio=None, top_k_attributions=None, trend_mod=None, trend_width=None, seasonal_width=None)

Generates time series forecasts based on the fitted model.

Parameters:

dataDataFrame, optional

Index and exogenous variables for forecast. For ARIMAX only.

Defaults to None.

keystr, optional

The timestamp column of data. The data type of the key column should be INTEGER, TIMESTAMP, DATE, or SECONDDATE. For ARIMAX only.

In massive mode, defaults to the first non-group key column of data if the index columns of data are not provided. Otherwise, defaults to the second of index columns of data and the first column of index columns is group_key.

group_keystr, optional

The column of group_key. Data type can be INT or NVARCHAR/VARCHAR. This parameter is only valid when massive is True.

Defaults to the first column of data if the index columns of data are not provided. Otherwise, defaults to the first column of index columns.

group_paramsdict, optional

If massive mode is activated (massive is True in class instance initialization), input data is divided into different groups with different parameters applied.

An example with group_params is as follows:

Valid only when self.massive is True.

Defaults to None.

forecast_method{'formula_forecast', 'innovations_algorithm', 'truncation_algorithm'}, optional

Specify the forecast method.

'formula_forecast': forecast via formula.

'innovations_algorithm': apply innovations algorithm to forecast.

'truncation_algorithm': a forecast method much faster than the innovations algorithm when the AR representation of the ARIMA model can be truncated to finite order.

Defaults to 'innovations_algorithm' if, in class initialization, the parameter forecast_method is not set, or set as 'innovations_algorithm'; otherwise defaults to 'formula_forecast'.

forecast_lengthint, optional

Number of points to forecast. Valid only when data is None.

In ARIMAX, the forecast length is the same as the length of the input predict data.

Defaults to None.

allow_new_indexbool, optional

Whether to recalculate and output the index column of the forecast result based on the type of the fitting data's index column.

True: The index column in the forecast result will be recalculated to match the type and sequence of the fitting data's index column.
False: The forecast result will output the original result from HANA PAL, which may use an integer index even if the fitting data's index column is a timestamp.

Defaults to False.

show_explainerbool, optional

Indicates whether to invoke the ARIMA with explanations function in the predict. Only valid when background_size is set when initializing an ARIMA instance.

If True, the contributions of trend, seasonal, transitory, irregular, and exogenous are shown in an attribute called explainer_ of the ARIMA / auto ARIMA instance.

Defaults to False.

thread_ratiofloat, optional

Adjusts the percentage of available threads to use, from 0 to 1. A value of 0 indicates the use of a single thread, while 1 implies the use of all possible current threads. Values outside the range will be ignored and this function heuristically determines the number of threads to use. Valid only when show_explainer is True.

Defaults to -1.

top_k_attributionsint, optional

Specifies the number of attributes with the largest contribution that will be output. 0-contributed attributes will not be output. Valid only when show_explainer is True.

Defaults to 10.

trend_modfloat, optional

The real AR roots with inverse modulus larger than this parameter will be integrated into the trend component. Valid only when show_explainer is True. Cannot be smaller than 0.

Defaults to 0.4.

trend_widthfloat, optional

Specifies the bandwidth of the spectrum of the trend component in units of rad. Valid only when show_explainer is True. Cannot be smaller than 0.

Defaults to 0.035.

seasonal_widthfloat, optional

Specifies the bandwidth of the spectrum of the seasonal component in units of rad. Valid only when show_explainer is True. Cannot be smaller than 0.

Defaults to 0.035.

Returns:

DataFrame 1: Forecasted values.
DataFrame 2 (optional): The explanations with decomposition of trend, seasonal, transitory, irregular, and reason code of exogenous variables. Only valid if show_explainer is True.
DataFrame 3 (optional): Error message. Only valid if massive is True.

set_conn(connection_context)

Set connection context for an ARIMA instance.

Parameters:

connection_contextConnectionContext: The connection to the SAP HANA system.

Returns:

None.

Inherited Methods from PALBase

Besides those methods mentioned above, the AutoARIMA class also inherits methods from PALBase class, please refer to PAL Base for more details.