The ability to accurately forecast sales ahead of time has long been a dream of many e-commerce businesses. Indeed, sales forecasting is an essential building block of any successful business. Predicting future conditions, managing your workforce, cash flow, and supply chain; all of these are aided by sales forecasting.

However, truly accurate sales forecasting has long been stymied by a reliance on historical sales data. The value of a sales forecast ultimately depends on its accuracy, but traditional forecasting methods utilise sales data to predict the future. Sales can be easily affected by price changes, short-term trends in user preferences, promotion activities, and much more. Meanwhile, inaccurate forecasts can lead to stock issues, such as overstocking.

All of this ultimately damages decision efficiency in e-commerce. As Tom Breur of ‘Data, Analytics and beyond’ argues, “For marketing and business development purposes you (almost) always want information about demand, not your historic ability to meet that demand as reflected in your sales.”

There is a clear need, then, for businesses to develop valuable insights using ultra-accurate sales forecasting—and AI can finally provide the answer. Researchers at Zhejiang University in Hangzhou, China, have developed a method for using Convolutional Neural Networks (CNN) to generate effective features for use in forecasting.

Revolutionary research

“When fed with raw log data, our approach can automatically extract effective features from that and then forecast sales using those extracted features. We test our method on a large real-world dataset from CaiNiao.com and the experimental results validated the effectiveness of our method,” authors Zhao and Wang explain.

As the research article outlines, recent methods of sales forecasting have improved, yet remain a time-consuming endeavour. We are now able to integrate log data into feature engineering, such as page views (PV), page view from search (SPV), user view (UV), user view from search (SUV), selling price (PAY, and gross merchandise volume (GMV). However, “these features are generally case-by-case extracted for specific commercial scenarios and models are difficult to reuse when data or requirements change.”

“Feature learning, rather than manual feature engineering.”

“We use the proposed approach to forecast sales by taking the raw log data and attributes information of commodities as the input. Firstly, we transform the log data and attributes information of commodities [into] a designed Data Frame. Then, we apply Convolutional Neural Network to the Data Frame, where effective features will be extracted at the hidden layers [of data] and subsequently used for sales forecast.”

Neural networks can potentially remove the need for manual feature engineering, making forecasting a much more accurate and less time-consuming prospect. The researchers’ experiment demonstrated “a significant performance improvement.” As Victor Rosenman, CEO of Feedvisor argues:

“Logistics used to be the core competency of retail; today, algorithms constantly crunch data, predict market trends, and respond to market changes in real time. Such advancements are only possible because of AI. All the data from the operations, the market, and the competition can be consolidated and analysed. It can be examined historically and now, with the help of AI technology, forecasted as well.”

AI continues to transform E-commerce. Giving businesses the ability to comprehensively forecast their future sales at significantly increased efficiency will no doubt further the streamlining of operations this technology offers.

Source:
https://arxiv.org/pdf/1708.07946.pdf