At a Glance
International Clothing Company
15k employees; 3,000 company-operated locations
To predict sales and sell-out events for mens’ and womens’ new products 12-months in advance using historical sales data and product metadata.
SFL Scientific used metadata and historical sales data to train a machine learning (LightGBM) model on New Core material. This included creating BI tools that automatically forecast and report adjusted demand volumes.
Expertise & Technology
Supply-side forecasts reduce lost profit from understocking and reduce housing costs incurred from overstocking. The retailer wanted the benefit of an optimized supply chain that manages multiple distributors, warehouses, and outlets that have varying needs for products and stock. The SKUs and quantities needed by each distributor varies based upon complex dependant factors such as target consumer, location, and seasonality.
SFL Scientific Solution
The large retailer has a large catalog of products and historical data. The current demand-side forecasts are based on a SAS-based statistical Auto Regressive Integrated Moving Average (ARIMA) model that is geared towards predicting up to 36 months into the future, using a minimum of 6 months of actual sales data.
In order to improve this model in particular for “New Core” products, SFL Scientific looked to leverage the power of machine learning. Since many New Core products have fewer than 6 months of historical data, the current SAS-based model provided could not be used to provide any meaningful prediction; additional external data about each product (e.g. product type, tier, etc) was used to aid in building a forecast.
SFL Scientific used this metadata and historical sales data to train a machine learning (LightGBM) model on New Core material. This included creating a BI tools that automatically forecast and report Adjusted Demand volumes.
SFL Scientific’s model used a window of historical sales as a contextual snapshot, aggregating a set number of previous months as features to infer future sales; to update the model, the historical window is updated to include the most recent month on record while dropping the oldest month previously captured in the window. This “sliding window” maintains a consistent level of model complexity by excluding older data too stale to offer substantive predictive value.
This model performed well in predicting Adjusted Demand, with lowers errors associated with products sold in larger quantities and fewer gaps in historical information, and provided a 15% improvement in forecasting over the SAS-based statistical Auto-Regressive Integrated Moving Average (ARIMA) model.