Forecasting Case Study: ML-driven forecasts for a manufacturer with promotions
Using cutting-edge machine learning models, SupChains delivered a forecasting model that helped an international manufacturer to reduce their forecast error by 20% compared to the benchmark.
The client is now using this model as a baseline for its monthly demand planning process. On top of the extra forecasting accuracy resulting in fewer lost sales and obsolete inventory, this new model reduced the workload of the demand planning team. The model is also used to generate demand scenarios with and without promotions.
The client enjoys being a global industry leader in the automotive industry. It manufactures parts and sells them to distributors and retailers. They partnered with SupChains to develop a forecasting model for one of their geographical markets.
The client’s budgeting exercise is carried out six months before the beginning of the fiscal year. So, they need to forecast demand up to 18 months ahead to plan their supply. To support their budgeting process, they need a granular forecast per channel, region, customer grouping, and product. This granularity allows them to create an accurate financial forecast. Indeed, as for most B2B companies, the same product can be sold at different prices to different channels or clients.
Moreover, part of the clients’ business is promotion-driven: around 20% of their overall sales are made during promotions.
The client could gather seven years of historical sales data (and five years of promotions). At the required aggregation level (channel x region x customer group x product), we face 120,000 unique combinations. The resulting granular demand can be described as low-volume with an erratic long tail. As shown in the figure below, 70% of the periods have 0 sales. Nevertheless, the average monthly sales are around 4.5 units as some products sell up to 1,500 pieces a month.
The manufacturer also runs around ten different types of promotions for sell-in and sell-out (such as “Buy 3 get 4” or various cashback).
In summary, this dataset and forecasting setup is challenging for three reasons: