def seasonal_error(
self, time_series: Union[pd.Series, pd.DataFrame], forecast: Forecast
) -> float:
r"""
.. math::
seasonal_error = mean(|Y[t] - Y[t-m]|)
where m is the seasonal frequency
https://www.m4.unic.ac.cy/wp-content/uploads/2018/03/M4-Competitors-Guide.pdf
"""
# Remove the prediction range
# If the prediction range is not in the end of the time series,
# everything after the prediction range is truncated
forecast_date = pd.Timestamp(forecast.start_date, freq=forecast.freq)
date_before_forecast = forecast_date - 1 * forecast_date.freq
ts = time_series[:date_before_forecast]
# Check if the length of the time series is larger than the seasonal frequency
seasonality = (
self.seasonality if self.seasonality else get_seasonality(forecast.freq)
)
if seasonality < len(ts):
forecast_freq = seasonality
else:
# edge case: the seasonal freq is larger than the length of ts
# revert to freq=1
# logging.info('The seasonal frequency is larger than the length of the time series. Reverting to freq=1.')
forecast_freq = 1
y_t = np.ma.masked_invalid(ts.values[:-forecast_freq])
y_tm = np.ma.masked_invalid(ts.values[forecast_freq:])
seasonal_mae = np.mean(abs(y_t - y_tm))
return seasonal_mae if seasonal_mae is not np.ma.masked else np.nan
def seasonal_error(self, past_data: np.ndarray, forecast: Forecast) -> float:
r"""
.. math::
seasonal_error = mean(|Y[t] - Y[t-m]|)
where m is the seasonal frequency
https://www.m4.unic.ac.cy/wp-content/uploads/2018/03/M4-Competitors-Guide.pdf
"""
# Check if the length of the time series is larger than the seasonal frequency
seasonality = (
self.seasonality if self.seasonality else get_seasonality(forecast.freq)
)
if seasonality < len(past_data):
forecast_freq = seasonality
else:
# edge case: the seasonal freq is larger than the length of ts
# revert to freq=1
# logging.info('The seasonal frequency is larger than the length of the time series. Reverting to freq=1.')
forecast_freq = 1
y_t = past_data[:-forecast_freq]
y_tm = past_data[forecast_freq:]
seasonal_mae = np.mean(abs(y_t - y_tm))
return seasonal_mae if seasonal_mae is not np.ma.masked else np.nan
def __init__(self, loss_function: str, freq: str, *args, **kwargs) -> None:
super(NBEATSTrainingNetwork, self).__init__(*args, **kwargs)
self.loss_function = loss_function
self.freq = freq
self.periodicity = get_seasonality(self.freq)
if self.loss_function == "MASE":
assert self.periodicity < self.context_length + self.prediction_length, (
"If the 'periodicity' of your data is less than 'context_length' + 'prediction_length' "
"the seasonal_error cannot be calculated and thus 'MASE' cannot be used for optimization."
)
def test_get_seasonality(freq, expected_seasonality):
assert get_seasonality(freq) == expected_seasonality