How Not to use Machine Learning for Demand Forecasting
Eight years ago ToolsGroup was one of the first supply chain planning software vendors to employ machine learning to improve demand forecasting. As the tag line to a popular commercial goes, by now “we know a thing or two because we’ve seen a thing or two.”
When we first began employing this new technology, it wasn’t on anyone’s radar. Even in 2014 when Gartner wrote a case study about one of our customers using our machine learning to help forecast promotions, it was barely a blip on the horizon. But now we are being deluged by machine learning messages, many of them ridiculous.
In this blog I want to debunk just one of those messages — that machine learning is a universal problem solver — You just turn it on and it addresses your supply chain planning needs. Or as Shaun Snapp wrote recently in a blog entitled How Real is the SAP Machine Learning and Data Science Story, “SAP makes it sound like machine learning works like Skynet from the movie Terminator. But machine learning algorithms don’t work like this at all.”
Nikki Baird, Managing Partner at Retail Systems Research, recently offered a similar observation on Forbes.com, cautioning against the “Silver Bullet Syndrome.” She says, a “company executive reads an article on an airplane that mentioned AI and now insists that the company invest in order to get all the benefits he or she read about in the article. This is AI for AI’s sake, and whenever retailers make technology investments along these lines, they invariably fail.”
I agree. Machine learning has certain strengths. Statistical forecasting approaches have others. Each does better in specific circumstances, so machine learning should be used as a complement to statistical forecasting.
One of our recent blogs described the difference between traditional statistics and machine learning (Click here to read the blog).
Gartner also recently published a report entitled Machine Learning 101 for Supply Chain Leaders (Noha Tohamy, February 2018) that highlights the differences. Tohamy says, “The difference between machine learning and traditional techniques is the role of the technology solution (the computer) in the process of solving a problem. In traditional programming, the computer is given input data, together with prechosen algorithms that are then used to generate an output when given new data. In machine learning, the computer is given a training dataset with inputs and corresponding output. The computer identifies patterns between the inputs and output and chooses the algorithm that would best represent those patterns. Once defined, this algorithm is then used to generate similar output in new data.” (See diagram above)
So traditional statistical forecasting algorithms bring a lot of insights into long term behavior and are able to give very solid results with very small amounts of data. For long term forecasting, they are also good at capturing seasonality from noise and long term trends.
Machine learning offers a unique insights on understanding many sources of data and inferring outcomes. It is very good at analyzing large quantities of complex, dirty and incomplete data and so it offers better visibility in explaining short term demand variations, such as inputs from social media, weather patterns, or IoT. In data rich environments, the ability to consume and leverage ever-increasing amounts of data clearly favors employing machine learning to help sense demand and to adapt supply to shifting consumption and replenishment needs.
For demand shaping events, machine learning also has a big edge on translating noise to signal, and understanding the underlying attributes, allowing it to drive to conclusions such as a promotion’s real impact or a new product launch’s potential sales. But machine learning can also have limitations that need to be constrained, like a tendency to jump from one forecast quantity to something quite different. It can also “see” mistaken causality, confusing random coincidental behavior with causality. Therefore the best combination is to have both technologies embedded in supply chain software, able to work together seamlessly and automatically, giving users the ability to forecast demand without concerning themselves with the underlying technology.
Machine learning is a great technology, if you know a thing or two about how to use it.
Improving on Four Analytic Techniques
Gartner also states that machine learning (ML) can improve upon the capabilities of the four analytics techniques, offering the following examples:
Descriptive analytics: What has happened?
In supply chain performance management, with machine learning, an organization can uncover previously unidentified patterns to improve its understanding of the health of its supply chain.
Descriptive analytics: Why is it happening?
Using root cause analysis, machine learning can point to specific patterns and uncover correlations among variables to reveal the root causes behind an event or a disruption. ML can continue to analyze new data to point to new patterns and potential root causes.
Predictive analytics: What will happen?
For demand sensing, with machine learning, a company can uncover external causal attributes, like flow of traffic, demographics and competitive moves that impact predicted demand. ML will continue to test hypotheses and revise its strategy to generate the most accurate predictors of demand.
Prescriptive analytics: What should I do?
For risk management, machine learning can find correlations between certain attributes and risk exposure. Based on predicting future disruption, ML can recommend mitigation strategies through analyzing new data and self-learning from the effectiveness of previous mitigation strategies.