Source: Charles Gaydon This data only contains 5 variables of Product_code, Warehouse, Product_Category, Date, Order_demand I showed that it is possible, with trivial models, to lower the mean average forecasting error to only around 20% in terms of volume of command, this for 80% of the total volume ordered. This should prove that there is a predicting potential in this dataset that only waits to be exploited. Again, I the reader wants to continue this work, he or she should use only a selection of the past months to make the forecast. Other ideas for further development : --> use warehouse and category data in the model; --> predict normalized categories of order command (ex: 0 - 1 to 20% -... - 100% to 120%; where 100% is the historical max of a product) and use a classifier instead of a linear model. --> check for AIC, BIC, & AICc scores.