Today, across almost every category, products have more features and attributes than those available just a few years ago. For instance, a simple kitchen cleaner is not so simple anymore; there are countless combinations to choose from when you consider the brand, fragrance types, price bands and packaging sizes. The electronics industry has similar challenges and complexities.
The trend has arisen as manufacturers and retailers attempt to cater to disparate markets and a wide variety of preferences by offering more choices. While this is a good thing for consumers, it presents a major challenge for the companies that produce and sell these products. As new products are introduced with a greater number of product attributes, there is no direct sales history available to study historical demand patterns. This has made reliable demand forecasting even more difficult.
Why has this been a difficult problem to solve?
When trying to forecast demand based on attributes, traditional regression models for generating statistical forecasts fall short. This is because the models are inherently deficient in understanding how the set of attributes work together to influence demand. There haven’t been too many other choices available to demand planners that could produce reasonable forecasts at a feasible cost — until now.
Unlocking the value of attribute-based forecasting
Without a reasonably good handle on demand, everything else in the supply chain becomes exponentially more challenging. Companies that can predict demand just a little better than their competition stand to gain tremendous advantages in the market; they can plan supply more reliably, rely less on inventory safety stocks, utilize manufacturing capacity more effectively, spend less on expediting – all of which translates into better revenues and profit margins. Better understanding the impact of product attributes on customer demand and improving the quality of their overall demand forecasts is something that benefits both retailers and manufacturers.
Even though historical sales data are not available for new products, it is possible to analyze how different attributes — such as speed, size, brand, price band, category, etc. — from other products have performed in the past. Attribute-based forecasting enables companies to analyze a cross-section of these attribute values and make better demand projections for new product launches.
New technology ushers in new opportunities
Recent advancements in computing power, available memory and machine learning have allowed data and decision scientists to explore advanced neural network-based techniques to study how a combination of product attributes impacts overall product demand. Just as different incoming signals trigger layers of neurons in a biological brain, machine-learning-based computational models simulate multiple product attributes to understand the impact on overall product demand. Such techniques allow for programs to be trained as more data become available, which progressively leads to better predictions over time.
Attribute-based forecasting that uses advanced neural network techniques is still in the research phases in the labs of leading supply chain software companies, but the pace of progress has been very rapid. This likely means that commercial applications will be available soon. Software companies exploring such machine-learning techniques are looking to partner with other forward-thinking manufacturers and retailers to create commercially feasible and scalable applications.
Attribute-based forecasting is just one way machine learning can be leveraged to power your digital supply chain. Learn how it can also enable intelligent segmentation, drive insights from unstructured and structured data, support image recognition for shelf logistics, and help forecast returns.