By: Eric Siegel, Founder, Predictive Analytics World for Business
In anticipation of his upcoming conference presentation at Predictive Analytics World for Business Las Vegas, May 31-June 4, 2020, we asked Kevin Feasel, Engineering Manager, Predictive Analytics at ChannelAdvisor, a few questions about their deployment of predictive analytics. Catch a glimpse of his presentation, Forecasting Demand in the e-Commerce Space, and see what’s in store at the PAW Business conference in Las Vegas.
Q: In your work with predictive analytics, what behavior or outcome do your models predict?
A: Our team’s job is to help our customers–mid- to large-sized organizations–forecast demand for products. For example, we estimate quantity sold for products over windows such as 30, 60, and 90 days and also predict whether our customers are likely to run out of stock in the meantime.
Q: How does predictive analytics deliver value at your organization – what is one specific way in which it actively drives decisions or operations?
A: Within ChannelAdvisor, we use predictive analytics to solve several problems. For example, we actively monitor systems using a custom-built anomaly detection engine. That way, we can track if the number or breakdown of errors differs from “normal” and alert teams on failure with fewer false positives. Teams have built monitoring based off of rules of thumb, but these typically lead to a significant number of either false positives or false negatives–or even both. Further, people typically only change the parameters after a significant number of false positives or known false negatives, so the adaptation rate on hand-tuned systems tends to be low. We were able to drop in our anomaly detection engine and immediately see good results not just at the moment of implementation, but also over time as what constitutes normal adjusts over time.
Q: Can you describe a quantitative result, such as the predictive lift of your model or the ROI of an analytics initiative?
A: Our model is accurate approximately 81% of the time for non-trivial products (where trivial products, for example, don’t sell anything during the course of the time frame). We also compare our results to a reasonable default: a simple weighted moving average of past product history. In comparison to this default model, our model has 25% less error.
Q: What surprising discovery or insight have you unearthed in your data?
A: Something we, as a team, did not know when we started was just how extremely long-tail online sales is. Across the board, over 85% of products sell no more than three times a month. We had an understanding that the web makes long-tail product sales much easier, as you don’t need to keep manage retail inventory and shelf space, but the extent to which this was the case surprised us.
Q: Sneak preview: Please tell us a take-away that you will provide during your talk at Predictive Analytics World.
A: Features beat models. We got a reasonable amount of lift from trying out different algorithms early on, but found ourselves chasing noise. It was only after we incorporated additional data sets and reshaped existing data before we started getting results which consistently beat our reasonable default.
—————————–
Don’t miss Kevin’s presentation, Forecasting Demand in the e-Commerce Space, at PAW Business on Wednesday, June 3, 2020 from 4:45 to 5:05 PM. Click here to register for attendance.
By: Eric Siegel, Founder, Predictive Analytics World for Business
This content is restricted to site members. If you are an existing user, please log in on the right (desktop) or below (mobile). If not, register today and gain free access to original content and industry news. See the details here.
 |
 |
 |
