Examples

Croston’s Method

To perform Croston’s method on a time series, import the croston module. The module contains adaptations such as Syntetos-Boylan Approximation (sba), Shale-Boylan-Johnston (sbj), Teunter-Syntetos-Babai (tsb) as well as an auto-selection method (auto).

Smoothing factors, alpha and beta, can be set manually.

>>> from intermittent_forecast import croston

>>> ts = np.random.randint(0,2,100)
>>> croston(ts, method='sba', alpha=0.1, beta=0.05)
array([nan, 0.6825, 0.0.6928934, 0.6928934, ...,])

An optimised fit can also be performed which will select values for alpha and beta such that the chosen metric is minimised.

>>> croston(ts, method='sba', opt=True, metric='mse')
array([nan, 0.6999993 , 0.69999972, 0.69999972, ...,])

ADIDA

Aggregate-Disaggregate Intermittent Demand Approach.

The input time series is aggregated into “buckets” to reduce or remove intermittency. One of the forecasting methods from the Croston module can then be applied to the aggregated series, followed by a seasonal or e qual-weighted disaggregation.

>>> from intermittent_forecast import adida

>>> ts = np.random.randint(0,2,100)
>>> adida(ts, size=3, overlapping=True, method='sba', alpha=0.1)
array([nan, nan, nan, nan, 0.2462963, 0.27708333, ...])

IMAPA

Intermittent Multiple Aggregation Prediction Algorithm.

The input time series is aggregated using multiple aggregation sizes. A forecast for each aggregation size is calculated, then either the mean or median is taken to produce a single forecast series.

>>> from intermittent_forecast import imapa

>>> ts = np.random.randint(0,2,100)
>>> imapa(ts, sizes=[1,2,3], combine='mean', overlapping=True)
array([nan, nan, nan, 0.5713449, 0.59967533, ...])