A. Objective: generate forecast at SKU level, by any time horizon (by week, by month) automatically
B. Mechanism:
-
Data structure: N x M matrix. N is time horizon M are volume by ton of SKUs, exclude B2B
-
Forecast function: HoltWinters (HW), Seasonal Linear (SLN), ARIMA. Parameters of model are optimized automatically
-
Data pre-processing: if ( 3 years < historical data) : model read data from year 2015, use HW and SLN .
if ( 2 years < historical data < 3 years) : model read data from year 2016, use HW and SLN .
if ( 1 years < historical data < 2 years) : model read data from year 2017, use SLN .
other : use ARIMA
-
Interate each SKU and generate forecast, then append them to one data frame.