PRIVATE AREA LOGIN

Are you a new user? Click to register for free

Have an account ? click for login

Minimum 5 characters

Minimum 8 characters

GN Demand Forecasting: ML for Sales Forecasting

Solutions

Statistical and Machine Learning algorithms for sales forecasting

GN Techonomy has realized, using Statistical and Machine Learning Algorithms, a solution for forecasting product demand – and therefore sales -, combined with Oracle Analytics for the visualization and analysis of results. Starting from taking charge of the monthly sales history detected by the ERP JD Edwards, the system performs a classification of the articles in seasonal or continuous, an analysis and a normalization of sales anomalies and spot orders per item / customer over time and, calculate the monthly forecasts by article with the possibility of splitting the monthly forecasts into weekly. This solution allows the company to plan orders in advance to promptly satisfy requests, avoid pending orders and, at the same time, limit warehouse costs.

Forecast calculation

The model created by GN Techonomy is flexible and adapts to both the series of sales history available and the life cycle of the article. The sales history is reprogrammed in the case of spot orders automatically by the system in order to avoid that a “sporadic” data could influence the forecasts. In fact, the article / customer analysis often highlights “anomalous” situations of spot sales. Therefore, before proceeding with the calculation of the forecasts on the historical sales basis, the system must analyze the spot situations and check if it is an isolated case or if the demand for that article can be considered recurring. The historian is then revised to harmonize anomalous situations in order to provide the forecasting algorithms with a historical-sales base without peak situations. There is also the possibility of managing products with a historical-sales base lower than the classic 24-month period.

As for the new articles, or in any case with a sales history of less than 6 months, the model connects these articles to a series of articles from the past that may be considered similar – we will talk about the map of correlations available to the User. The system analyzes the sales of the first 6 months of these goods and produces a forecast of the demand for the new product.

Articles with historical basis over time: at least 24 monthsArticles with historical basis over time: less than 24 monthsNew articles: less than 6 months
Analysis and Classification of Articles by (seasonal, trend …) Abnormal situations management (spot sales) Use of multiple algorithms on Machine LearningAutomatic search of the historical period Check historical trend (continuous, discontinuous …) Use of multiple algorithms on Machine LearningMap for inserting «reference model items» Sales analysis group of reference items in the initial period Use of multiple algorithms on Machine Learning
→ Monthly forecasts with possible weekly breakdown→ Weekly forecasts→ Weekly forecasts based on the historical-sales of reference products
Forecasts Reliability

At each new calculation of the Forecasts, the system records a history: the current operating forecasts are overwritten but the previous values ​​are recorded in a historical table. For each article it is therefore possible to analyze how the forecasts in a given period have changed. The check has a preventive meaning when analyzing the future and checking if the forecasts recalculated from month to month have substantial differences: constant values ​​indicate sales in line with the forecasts of the first period and, at the same time, good reliability for the future. If the analysis, however, involves significant deviations, then the articles are alarmed as it is necessary to analyze the critical lead times, the safety stock levels and other Plan data. The Machine Learning system created by GN Techonomy analyzes both sales data and forecast trends in order to identify correlations. Reliable forecasts – at least 80% up to 95% – are made thanks to the combined use of different algorithms.

Info

For more information, contact us!

LinkedIn
Machine Learning