The fashion industry is undergoing seismic shifts due to omnichannel and social media disruption that make predicting and managing demand more difficult than ever. Two MIT graduates and supply chain professionals share their research findings on how machine learning can improve forecast accuracy by up to 50%, and lay the foundation for truly agile supply chains.
We Must Have An Answer To Social Media & Omnichannel Disruption
Over the past several years, the fashion industry has been undergoing several challenges. Some of these challenges are supply chain-oriented including long lead times from design to delivery. Other challenges were more related to the microeconomic shifts in the global economy like the emergence of omnichannel and social media. These have led to shorter product lifecycles, obsolescence of the retail calendar, and consequently more volatile and uncertain demand.
The Emerging Need For An Agile Supply Chain In The Fashion Industry
To cope with this changing retail landscape, companies must develop agile supply chains that focus on speed, responsiveness and flexibility. An agile supply chain should enable fashion retailers to respond rapidly to changes in demand, in terms of both volume and variety. Since demand forecasting plays a critical role in supporting upstream supply chain planning decisions, the previously mentioned shifts have made forecasting for fashion products much more challenging.
How Can We Improve Forecast Accuracy To Meet Today’s Strategic Needs?
To answer this question, myself and my research colleague Vicky Chan carried out research at the Massachusetts Institute of Technology as part of our master program in supply chain management. The sponsoring company who participated in this research was a leading US-based footwear retailer. The type of products we investigated were seasonal (first launch) footwear which, in this retailer’s case, represent more than 50% of the SKU count.
The challenge with forecasting demand for these products lies in the lack of historical data
The challenge with forecasting demand for these products lies in the lack of historical data, which traditional time series techniques rely heavily on. For this retailer, the typical lifecycle of a seasonal product is around two to four months. With an order lead time of up to five months from the manufacturer, there is no opportunity to read and react to actual sales data.
The Potential Of Machine Learning And Artificial Intelligence
Besides point-of-sale data (POS), what other types of data can be leveraged to predict demand? We explored the use of product attributes, calendar, lifecycle, price and promotion, as well as store count to build a forecasting model for the seasonal products.
To that end, we see machine learning-based forecasting techniques as potential candidates for demand forecasting for seasonal products. Unlike traditional forecasting methods, machine learning techniques are able to process a large number of predictor variables not confined to sales history, determining the ones that are significant. In building the forecasting model, we explored the use of different machine learning techniques, including regression trees, random forests, k-nearest neighbors (k-NN), linear regression and neural networks. In addition, clustering and classification techniques offer the opportunity to identify similar existing products such that their sales history can be leveraged.
Cluster, Classify and Predict
We analyzed the data to identify significant predictor variables influencing demand for footwear products. Results show that store count, calendar month and lifecycle month (which month in the lifecycle the sale occurs in) are the top three numerical variables impacting demand. Color, material and gender were the top three categorical variables. Two models, a general model and a three-step model, were then built utilizing product, calendar, lifecycle, price and store count attributes to predict demand. The general model directly takes in the variables for prediction, while the three-step model involves clustering and classification to identify similar products, before moving on to prediction.
The machine learning model improved the retailer’s forecast accuracy by more than 50%
The results show that the two forecasting models we developed achieve better forecast accuracy compared to the sponsoring company’s current performance. They both improve the retailer’s forecast accuracy by more than 50%.
While the general model serves as a starting point for easy implementation of the machine learning forecasting framework, the three-step model further offers visibility into the importance of the different underlying factors that impact demand. The project results also demonstrate the value of forecast customization based on product characteristics which offer additional improvement to forecast accuracy. In conclusion, we believe this research not only benefits our sponsoring company, but also other fashion retailers forecasting for seasonal products.
This article is based on research carried by Majd Kharfan and Vicky Chan, two forecasting and supply chain professionals participating in the the Supply Chain Management Master’s course at Massachusetts Institute of Technology (MIT).
The role of machine learning in forecast accuracy will be discussed at IBF’s upcoming Business Planning, Forecasting & S&OP: Best Practices Conference from October 16-19, 2018 in Orlando, F.L. Speakers from NIKE and PUMA will reveal the implications of machine learning for the fashion industry, and how it cuts lead times and meet the challenges of the omnichannel environment. REGISTER NOW FOR EARLY BIRD PRICING.