def forecast(x, h=None, **kwargs):
'''
Generate a forecast for the time series x, using ets if x is non-seasonal
or has frequency less than 13, and stlf if x is periodic with frequency
above 13.
Args:
x: an R time series, obtained from converters.ts(), or a Pandas Series
with the correct index (e.g. from converters.sequence_as_series().
h: the forecast horizon
level: A number or list/tuple of prediction interval confidence values.
Default is 80% and 95% intervals.
robust: Default False. If True, missing values are filled before
forecasting and outliers are identified and replaced with tsclean().
lam : BoxCox transformation parameter. The default is R's NULL value.
If NULL, no transformation is applied. Otherwise, a Box-Cox
transformation is applied before forecasting and inverted after.
find_frequency: Default False. If True, function will try to determine
the series frequency from the data.
allow_multiplicative_trend: Default is False. If True, consider models
with a multiplicative trend component. That type of model may grow
explosively.
Returns:
If x is an R ts object, an R forecast is returned. If x is a Pandas
Series, a Pandas Data Frame is returned.
'''
x, is_pandas = converters.to_ts(x)
h = _get_horizon(x, h)
kwargs = converters.translate_kwargs(**kwargs)
out = fc.forecast(x, h=h, **kwargs)
return converters.forecast_out(out, is_pandas)
def forecast(x, h=None, **kwargs):
'''
Generate a forecast for the time series x, using ets if x is non-seasonal
or has frequency less than 13, and stlf if x is periodic with frequency
above 13.
Args:
x: an R time series, obtained from converters.ts(), or a Pandas Series
with the correct index (e.g. from converters.sequence_as_series().
h: the forecast horizon
level: A number or list/tuple of prediction interval confidence values.
Default is 80% and 95% intervals.
robust: Default False. If True, missing values are filled before
forecasting and outliers are identified and replaced with tsclean().
lam : BoxCox transformation parameter. The default is R's NULL value.
If NULL, no transformation is applied. Otherwise, a Box-Cox
transformation is applied before forecasting and inverted after.
find_frequency: Default False. If True, function will try to determine
the series frequency from the data.
allow_multiplicative_trend: Default is False. If True, consider models
with a multiplicative trend component. That type of model may grow
explosively.
Returns:
If x is an R ts object, an R forecast is returned. If x is a Pandas
Series, a Pandas Data Frame is returned.
'''
x, is_pandas = converters.to_ts(x)
h = _get_horizon(x, h)
kwargs = converters.translate_kwargs(**kwargs)
out = fc.forecast(x, h=h, **kwargs)
return converters.forecast_out(out, is_pandas)
def nsdiffs(x, **kwargs):
'''
Estimates the number of seasonal differences to take on the time series,
x, to reach stationarity. For this function, x must be a seasonal series.
Args:
x: an R time series or a Pandas Series
m: Seasonal period. Default is frequency(x). No other value makes sense.
test : Test to use to determine number of seasonal differences.
Default is 'ocsb' for the Osborn-Chui-Smith-Birchenhall test.
The alternative is 'ch' for the Canova-Hansen test.
max_D: Maximum number of seasonal differences to try. Default is 1.
Returns:
The number of seasonal differences to take
'''
x, _ = converters.to_ts(x)
kwargs = converters.translate_kwargs(**kwargs)
return fc.nsdiffs(x, **kwargs)[0]
def ndiffs(x, **kwargs):
'''
Estimates the number of first differences (non-seasonal) to take on the
time series, x, to reach stationarity.
Args:
x: an R time series or a Pandas Series
alpha: Default 0.05, the level of the test used
test : Test to use to determine number of first differences. Default
is 'kpss', for the KPSS test. Other values are 'adf' for augmented
Dickey-Fuller, or 'pp' for Phillips-Perron.
max_d: max number of differences to try. Default is 2.
Returns:
The number of differences to take
'''
x, _ = converters.to_ts(x)
kwargs = converters.translate_kwargs(**kwargs)
return fc.ndiffs(x, **kwargs)[0]
def stl(x, s_window, **kwargs):
'''
Perform a decomposition of the time series x into seasonal, trend and
remainder components using loess. Most of the arguments listed below are
in **kwargs, and all of those arguments have sensible defaults. Usually
only the mandatory s_window paramter has to be set.
Args:
x: an R time series, obtained from converters.ts(), or a Pandas Series
with the correct index (e.g. from converters.sequence_as_series().
s_window : either 'periodic' or the span (in lags) of the
loess window for seasonal extraction, which should be odd.
This has no default.
s_degree : Default 0, should be 0 or 1. Degree of local polynomial
for seasonal extraction.
t_window : The span (in lags) of the loess window for trend extraction,
which should be odd. Default is a sensible, data-dependent value.
See the R docs for the details.
t_degree : Default 0, should be 0 or 1. Degree of local polynomial
for trend extraction.
l_window : Span in lags of the loess window used to low-pass filter each
seasonal subseries. The default is first odd number greater than or
equal to frequency, which is recommmended.
s_jump, t_jump, l_jump : integer parameters (min. 1) to increase speed of
each smoother by skipping data points.
l_degree : Default is t.window, must be 0 or 1. Degree of local polynomial
for subseries low-pass filter.
robust : Default is False. If True, robust loess fitting used.
inner : number of backfitting iterations
outer : number of outer robustness iterations
na_action : Default is na.fail, which means that the user has to fill or
remove any missing values. If used, it must be an object that maps to
an R function, obtained from rpy2.
Returns:
If x is an R ts object, an R object of class 'stl' is returned.
If x is a Pandas Series, a Pandas Data Frame is returned.
'''
x, is_pandas = converters.to_ts(x)
kwargs['s.window'] = s_window
kwargs = converters.translate_kwargs(**kwargs)
out = stats.stl(x, **kwargs)
return converters.decomposition_out(out, is_pandas)
def nsdiffs(x, **kwargs):
'''
Estimates the number of seasonal differences to take on the time series,
x, to reach stationarity. For this function, x must be a seasonal series.
Args:
x: an R time series or a Pandas Series
m: Seasonal period. Default is frequency(x). No other value makes sense.
test : Test to use to determine number of seasonal differences.
Default is 'ocsb' for the Osborn-Chui-Smith-Birchenhall test.
The alternative is 'ch' for the Canova-Hansen test.
max_D: Maximum number of seasonal differences to try. Default is 1.
Returns:
The number of seasonal differences to take
'''
x, _ = converters.to_ts(x)
kwargs = converters.translate_kwargs(**kwargs)
return fc.nsdiffs(x, **kwargs)[0]
def ndiffs(x, **kwargs):
'''
Estimates the number of first differences (non-seasonal) to take on the
time series, x, to reach stationarity.
Args:
x: an R time series or a Pandas Series
alpha: Default 0.05, the level of the test used
test : Test to use to determine number of first differences. Default
is 'kpss', for the KPSS test. Other values are 'adf' for augmented
Dickey-Fuller, or 'pp' for Phillips-Perron.
max_d: max number of differences to try. Default is 2.
Returns:
The number of differences to take
'''
x, _ = converters.to_ts(x)
kwargs = converters.translate_kwargs(**kwargs)
return fc.ndiffs(x, **kwargs)[0]
def stl(x, s_window, **kwargs):
'''
Perform a decomposition of the time series x into seasonal, trend and
remainder components using loess. Most of the arguments listed below are
in **kwargs, and all of those arguments have sensible defaults. Usually
only the mandatory s_window paramter has to be set.
Args:
x: an R time series, obtained from converters.ts(), or a Pandas Series
with the correct index (e.g. from converters.sequence_as_series().
s_window : either 'periodic' or the span (in lags) of the
loess window for seasonal extraction, which should be odd.
This has no default.
s_degree : Default 0, should be 0 or 1. Degree of local polynomial
for seasonal extraction.
t_window : The span (in lags) of the loess window for trend extraction,
which should be odd. Default is a sensible, data-dependent value.
See the R docs for the details.
t_degree : Default 0, should be 0 or 1. Degree of local polynomial
for trend extraction.
l_window : Span in lags of the loess window used to low-pass filter each
seasonal subseries. The default is first odd number greater than or
equal to frequency, which is recommmended.
s_jump, t_jump, l_jump : integer parameters (min. 1) to increase speed of
each smoother by skipping data points.
l_degree : Default is t.window, must be 0 or 1. Degree of local polynomial
for subseries low-pass filter.
robust : Default is False. If True, robust loess fitting used.
inner : number of backfitting iterations
outer : number of outer robustness iterations
na_action : Default is na.fail, which means that the user has to fill or
remove any missing values. If used, it must be an object that maps to
an R function, obtained from rpy2.
Returns:
If x is an R ts object, an R object of class 'stl' is returned.
If x is a Pandas Series, a Pandas Data Frame is returned.
'''
x, is_pandas = converters.to_ts(x)
kwargs['s.window'] = s_window
kwargs = converters.translate_kwargs(**kwargs)
out = stats.stl(x, **kwargs)
return converters.decomposition_out(out, is_pandas)
def tsclean(x, **kwargs):
'''
Identify and replace outliers. Uses loess for non-seasonal series and
an STL decomposition for seasonal series. Optionally fills missing values.
Args:
x: an R time series, obtained from converters.ts(), or a Pandas Series
with the correct index (e.g. from converters.sequence_as_series().
replace_missing: Default True.
If True, use na_interp to fill missing values in x.
lam: optional BoxCox transformation parameter.
Returns:
If x is an R ts object, an R time series is returned. If x is a Pandas
Series, a Pandas Series is returned. In either case, outliers are replaced
and optionally, missing values are filled.
'''
x, is_pandas = converters.to_ts(x)
kwargs = converters.translate_kwargs(**kwargs)
out = fc.tsclean(x, **kwargs)
return converters.series_out(out, is_pandas)
def tsclean(x, **kwargs):
'''
Identify and replace outliers. Uses loess for non-seasonal series and
an STL decomposition for seasonal series. Optionally fills missing values.
Args:
x: an R time series, obtained from converters.ts(), or a Pandas Series
with the correct index (e.g. from converters.sequence_as_series().
replace_missing: Default True.
If True, use na_interp to fill missing values in x.
lam: optional BoxCox transformation parameter.
Returns:
If x is an R ts object, an R time series is returned. If x is a Pandas
Series, a Pandas Series is returned. In either case, outliers are replaced
and optionally, missing values are filled.
'''
x, is_pandas = converters.to_ts(x)
kwargs = converters.translate_kwargs(**kwargs)
out = fc.tsclean(x, **kwargs)
return converters.series_out(out, is_pandas)