Improving Replenishment and Markdown with Demand Forecasting

Challenges

• An omnichannel fashion retailer had no forward-looking view of where inventory would land at season end.
• Markdown and replenishment decisions were made too late — resulting in stockouts during campaigns, or forcing lower margins during clearance sales.
• Many products were manufactured in China with lead times of up to six months, exacerbating any errors.
• The fashion industry was characterized by distinct seasonal patterns and trends, making prediction extremely difficult.

Approach

• Built hierarchical forecasting models with top-down reconciliation, accounting for seasonality and trending products.
• Integrated with inventory, product, store and customer data for full 360-degree analytics.
• Built a clustering model to give new products with limited data a start-up profile.
• Collaborated with domain experts to implement real-world manual “safeguards” in the predictions, detached from the statistical model.
• Developed front-end tool to serve predictions and analytics to buyers, from quick overview to detailed product deep dives.
• Developed scenario simulation, where buyers could plug in assumptions to produce custom forecasts (e.g. “assume that this product will sell as an old model”).

Results

• A dedicated, forward-looking tool answering one simple question: how much inventory do I risk ending up with for clearance sales if I don’t do anything now?
• Ability to reorder or markdown products in a timely manner — resulting in higher sales and better margins.

Technologies & methods

Time-Series Forecasting Machine Learning K-Nearest Neighbours Inventory Optimization SAS AI