In my previous guest commentary, I shared some of my thoughts on the impact of machine learning on S&OP and described what would be a potential ultimate machine learning solution. In today’s post, I’ll share some concrete examples of machine learning applications used today in the context of S&OP.
Machine learning to boost your sales
The first example is to use machine learning for improving the results of your promotions and improving sales. Promotions planning is an important lever to use in an S&OP process. It allows to stimulate your sales when you have too much available capacity or have a significant impact on the desired product mix. Moreover, it is often an underused powerful lever to achieve your objective and beat your budget in the scope of your S&OP meetings. However, to truly have an optimal positive impact on the bottom line, aspects surrounding promotions need to be thought through carefully. Take the Coca Cola company, for example. It does many types of promotions: it advertises via TV, newspaper, billboards, radio; it offers discounts or bundles like “buy one, get one free”; it chooses the scope of the promotion (e.g., regional or national) and it decides the size of the marketing budget. There are many different possibilities to organize a promotion.
Now, if Coca Cola would like to do a new promotion to boost the sales of its Diet Coke 2 liters, what should it do? Which configuration of promotion will have the biggest impact on the market with the lowest impact on the wallet? Well, by using machine learning, it is possible to train an algorithm to figure out what will be the lift of the future promotion based on past promotion results. Using the latest algorithm in the field, provided you have sufficient data and the results of past promotions, you can get a pretty accurate view of the consequences of your choices while planning for a new promotion. In this way, you can make sure that you only create promotions that will be profitable and maximize your margin.
The second example is to use machine learning to greatly improve forecast accuracy. Since the sales forecast is the input for many important downstream decisions for S&OP, it’s critical to have it as reliable and precise as possible. A bad forecast will cause bullwhip effects and often be the source of overstocks, under-utilization and other unwanted consequences on your supply chain.
People familiar with demand planning know that determining the best level to do your statistical forecasting is a very difficult puzzle. If you go too low, then you are very likely to find spikes in your historical sales, which will be very difficult to forecast. If you go too high, then you will lose a lot of information (e.g., not seeing that Coca Cola Light and Zero have two different trends).
Currently, the puzzle is solved by demand planning solutions. These solutions test the different levels and find out which one is the best. In our experience, this approach is too restrictive and does not bring the best results. Indeed, if you only consider the existing hierarchy, you constrain yourself and only have a few possibilities to test out. This is quite rudimentary. We took another approach: we chose to consider all combinations of our forecasting attributes (for example, all combinations of products, customers and regions) as a cloud of points. We use a very advanced machine-learning algorithm to group those points in clusters. Then, running the statistical forecast on those clusters allows us to have, on average, a forecasting accuracy that is much better than usual.
Those two examples in the field of demand planning clearly show the positive effects of machine learning on S&OP: you get higher sales with forecasts that are more reliable. Apart from the two example described, there are other applications where we use machine learning to boost sales.
Machine learning to improve S&OP adoption
In the market, there are two types of S&OP solutions: analytics and optimization-based. The analytics solutions are easy to use and visually attractive with dashboards. However, they do not bring the real benefits as the plans are suboptimal and not optimized. The gap between a plan obtained by heuristics and optimization can lead to a difference as big as 20 percent on the bottom-line accumulated over 12 months.
The optimization-based solution, when implemented right, brings a much larger ROI but can also be perceived as being more complex to use because of the amount of data required to compute meaningful plans. For example, you need to input the throughputs of the machines, the average cost of the operations, and the average duration of a certain trip. In addition, the data is often not a direct output from the ERP, as a higher level of conceptualization is needed in S&OP. For example, you do not necessarily need the throughputs on individual SKUs as stored in the ERP but on industrial families. The adoption can certainly be longer because of all the data to maintain.
Although it does not directly impact the S&OP process, data maintenance can be a hindrance and is blocking many companies from getting the full ROI out of their S&OP.
But you can use machine learning to overcome that issue. How? By using the past recorded knowledge and machine learning to maintain the future knowledge used in the planning. For example, you can train a machine-learning algorithm on the past duration for going from Paris to Amsterdam with a truck type X to deduce the future duration and then the average costs. You can then build a complete supply chain where all the knowledge is auto-maintained and only needs human intervention occasionally. This approach greatly improves user adoption and maintenance while providing knowledge that is even more accurate, which results in more accurate S&OP plans.
In this sense, even though it did not influence the S&OP process directly, machine learning has significantly increased the perceived value of the process and allows you to reach higher benefits and return on investments.
In my opinion, machine learning and predictive analytics in general are already bringing a lot to S&OP via software and will probably bring even more benefits in the future. I have explained a few examples above where I see this coming; there are, of course, many more. What about you? Where are you seeing the impacts of machine learning on your S&OP? I am happy to continue the discussion and receive your ideas and views on what’s coming next!