kats.models.ensemble.kats_ensemble module

Kats ensemble model

Implementation of the Kats ensemble model. It starts from seasonality detection, if seasonality detected, it continues to perform STL decomposition, then fit forecasting models on de-seasonalized components and aggregate; otherwise it simiply leverage individual forecasting models and ensembling. We provided two ensembling methods, weighted average and median ensembling.

class kats.models.ensemble.kats_ensemble.KatsEnsemble(data: kats.consts.TimeSeriesData, params: Dict[str, Any])

Bases: object

Decomposition based ensemble model in Kats This is the holistic ensembling class based on decomposition when seasonality presents

aggregate() → pandas.core.frame.DataFrame

Aggregate the results from predict method

Parameters

None –

Returns

final results in pd.DataFrame

backTestExecutor() → Dict[str, float]

wrapper for back test executor

services which use KatsEnsemble need to write their own backtest wrapper

Parameters

None –

Returns

The dict of backtesting results

static clean_dummy_CI(fcsts: Dict[str, pandas.core.frame.DataFrame], use_zero: bool = True) → Dict[str, pandas.core.frame.DataFrame]

Helper method to clean dummy prediction interval

Parameters

• fcsts – the dict of forecasting results from individual models

• use_zero – flag to use zero to fill nan, default as True

Returns

the cleaned results in a dict

static deseasonalize(data: kats.consts.TimeSeriesData, decomposition_method: str) → Tuple[kats.consts.TimeSeriesData, kats.consts.TimeSeriesData]

STL decomposition to given TimeSeriesData

Static method to perform decomposition on the input data

Parameters

• data – kats.consts.TimeSeriesData, input time series data

• decomposition_method – the specific method for decomposition

Returns

Tuple of seasonal data and de-seasonalized data

fit() → kats.models.ensemble.kats_ensemble.KatsEnsemble

Fit individual forecasting models via calling fitExecutor

This is the fit methdo to fit individual forecasting model

fitExecutor(data: kats.consts.TimeSeriesData, models: kats.models.ensemble.ensemble.EnsembleParams, should_auto_backtest: bool = False) → Tuple[Dict[Any, Any], Optional[Dict[str, float]]]

callable forecast executor

This is native implementation with Python’s multiprocessing fit individual model in models with given data. Services who use KatsEnsemble need to implement their own executor for better performance, if no executor function is given, the native version will be used.

data

kats.consts.TimeSeriesData, given TimeSeriesData, could be original or de-seasonalized

models

EnsembleParams object containing model params in BaseModelParams

should_auto_backtest

boolean flag for additional back testing runs

Returns

Tuple of fitted individual model and weights

forecast(steps: int) → Tuple[Dict[str, pandas.core.frame.DataFrame], Optional[Dict[str, float]]]

Holistic forecast method in Kats ensemble

combine fit and predict methods to produce forecasted results this is especially useful for services which prefer to produce final forecasts without saving the fitted model

Parameters

steps – the length of forecasting horizon

Returns

Tuple of predicted values and weights

forecastExecutor(data: kats.consts.TimeSeriesData, models: kats.models.ensemble.ensemble.EnsembleParams, steps: int, should_auto_backtest: bool = False) → Tuple[Dict[str, pandas.core.frame.DataFrame], Optional[Dict[str, float]]]

Forecast Executor

This is a callable execution function to (1). fit model (2). predict with a given steps (3). back testing (optional)

Parameters

• data – kats.consts.TimeSeriesData, the input time series data as in kats.consts.TimeSeriesData

• models – the ensemble parameters as in EnsembleParams

• steps – the length of forecasting horizon

• should_auto_backtest – flag to automatically perform back test, default as False

Returns

The predicted values from each individual model and weights

plot() → None

plot forecast results

predict(steps: int) → kats.models.ensemble.kats_ensemble.KatsEnsemble

Predit future for each individual model

Parameters

steps – number of steps ahead to forecast

Returns

None

static reseasonalize(sea_data: kats.consts.TimeSeriesData, desea_predict: Dict[str, pandas.core.frame.DataFrame], decomposition_method: str, seasonality_length: int, steps: int) → Dict[str, pandas.core.frame.DataFrame]

Re-seasonalize the time series data

Static method to re-seasonalize the input data

Parameters

• sea_data – kats.consts.TimeSeriesData, the seasonal data from deseasonalize method

• desea_predict – dict of forecasted results for the deseasonalized data for each individual forecasting method

• decomposition_method – the specific method for decomposition

• seasonality_lenth – the length of seasonality

• steps – the length of forecasting horizon

Returns

Dict of re-seasonalized data for each individual forecasting model

static seasonality_detector(data) → bool

Detect seasonalities from given TimeSeriesData

Parameters

data – kats.consts.TimeSeriesData, the input TimeSeriesData

Returns

Flag for the presence of seasonality