Our guiding idea
The traditional approach based on historical sales is of no use where no such history exists, as is the case with new products. It is therefore necessary to take a different approach, starting perhaps from the idea that every product actually involves a mix of already existing features. By combining factors such as historical sales of similar products, stock levels, prices, product features, scheduling, as well as marketing and discount campaigns, it is possible to devise a predictive model able to determine the correct quantity of new products to be ordered by analogy.
Not just a theory
The solution is based on Open Source technologies specialised in Big Data processing. The data is collected in a central Data Lake using Cloudera distribution and is processed by attribute-based algorithms that make predictions based on the data. Starting from the features of the product and by exploiting the history of similar products, forecasts are calculated using ARIMA models and neural networks. The results are displayed to users using monitoring dashboards. To make the results more robust and reliable, a data pre-processing stage was added, during which the data is cleaned and standardised to eliminate exogenous events involving an increase or decrease in sales.