Descriptive, predictive and prescriptive analytics should be combined to optimize your demand planning processes
In early 2016, a 40-year-old Midwest process manufacturer with 40,000 SKUs finally completed their implementation of a new ERP system and quickly worked to combine all of their new data with an array of other operational software, spitting out tons of new data daily. Their goal was a data-driven organization making fact-based decisions in near real time. In particular, they wanted better forecasting and demand planning processes, which in the past had been beset by low accuracy and poor adoption.
Pulling their data into a BI tool in each department, they started generating a new series of reports designed to meet the requirements of each department. But those reports soon generated arguments across their sales, operations, marketing and finance teams. Data were inconsistent across groups, and despite mountains of graphs and tables, no one was clear on how to improve business using the information. After a good six months, the company still struggled to leverage any new insight and began reverting to tried and true Excel reports. Teams were disappointed. The spreadsheet reports took a month to produce and used all the capacity of two IT staff members. When the Finance Manager who had designed the reports resigned, the organization knew it had to figure out a better plan for using their data.
With the evolution of the digital supply chain has come mountains of new data for every company. But there’s no guarantee you can use that data effectively. In fact, there can be as much risk in misusing data as there is in flying blind with no data. In particular, forecasting and demand planning can be a challenge because they require a significant amount of quantitative analysis and subjective management judgment. To really exploit your new data to gain an advantage, you need a clear analytical plan, specific metrics tailored to your situation, and an understanding of what analytics and reporting will help you the best. This blog outlines three types of inter-related analytical tools that you need to use in concert to get the best insight from your data.
Sure, forecasting is the mainstay of demand planning. We generate a prediction from historic actuals, maybe bring in some external factors, allow management to edit the results, and publish new forecasts to drive planning. In all, it works pretty well, and there’s tons of written stuff on this. But what we often miss is a good understanding of the market dynamics driving our predictions, and an understanding of whether our adjustments to the plan are optimal. Integrating descriptive, predictive and prescriptive analytics will help you to improve your forecasting and demand planning processes. Here’s where they help.
Descriptive analytics are the summary measures and metrics that give you insight. Why are our error metrics (MAPE, MAE, etc.) changing so much – or not getting any better? Often, we simply act on error measures. But we really want to know how forecastable our data is and what is causing any bias. Think of descriptive analytics as the means, medians and modes, the histograms and frequency distributions, and the variance measures that you learned in college. They are critical because they tell you the underlying structure of your data. Key descriptive measures for demand planning: Skew and Kurtosis of your sales data. Skew is the measure of lumpiness at a particular end of the distribution. If you have too much skew, you can’t use means effectively and should rely on medians. For example, you ran multiple promotions at different price points for different packages in the past quarter. If sales bumped up large below a certain price point, your forecast will be biased unless you keep the promotion going. Instead, you’ll need to segment products for your forecast or make statistical adjustments to dampen the effect. Either way, you need a precise understanding of the bias and descriptive analytics will get you there.
Predictive analytics, practically, cover any effort to use the analysis of historical data points to anticipate what will most likely happen in the future. Time series forecasting is predictive – but again that’s well covered. However, newer approaches like tree analysis can be used to segment or classify sales results and build predictions from resulting likelihood estimates. These methods are powerful in many situations. They overcome some of the big weaknesses of times series, such as an inability to deduce what factors are driving the forecast. These approaches work by identifying factors that differentiate high versus low sales and incorporate a time dimension in the results to determine future trends.
Predictive analytics can also be used to assess risk around your forecast. Do you have very low accuracy levels compared to some reasonable benchmark? Why? In more than a few situations, we’ve built predictive models to understand the drivers of error versus the forecast. By targeting these drivers, we can reduce elements of error and ultimately improve the forecast. Regression methods are a good choice for building a model that can be translated into a decision tool.
Let’s say all our forecasting and market analysis is done and we’re reasonably happy with the results. We can increase inventory for spiking demand and clear our inventory for under-performers through pricing and promotion. But that’s pretty reactive. Prescriptive analytics are effectively recommendation engines that guide us to the best choice among multiple product actions – given the known situation. Collaboration filters in e-commerce are the most common tools.
In a large ongoing customer program, Halo is using a powerful tool called conjoint analysis to determine how consumers will react to different product features and offers. Add up all the features and their attractiveness, and we get a better view of how to get the most profitable outcome at a customer level. This model is used to derive a decision algorithm to automatically recommend the right offer for the individual case through a CRM system.
In demand planning, we can use prescriptive analysis at a strategic and tactical level. Strategically, most organizations are continuously weighing the trade-offs between revenue growth, profit margins, market share, and service levels. Since most research shows that a majority of new product launches can fail in their first year, it would be ideal to have better insight into whether to invest more, tolerate longer, or divest products based on the company’s goals. Prescriptive analytics include algorithms to make these trade-off analyses – then put them in a modeling and what-if calculator to enable business managers to make decisions quickly in terms that senior management will appreciate.
Overall, descriptive, predictive and prescriptive analytics mean different things – but their real value is combining them to make your organization data-driven. Use descriptive statistics to get a real feel for your business and understand the strengths and weaknesses of your data. Once you know your data, use predictive tools to enable the organization to understand likely future outcomes of your demand planning. And lastly, simplify and automate decision-making through prescriptive analytics. While it’s not yet time to turn over the business to the robots, we can frame up our decisions in a more structured fashion. We can then provide the quantitative information to get a demand consensus to best match supply and demand.
To learn more about the different analytic options available today, check out The Two-minute Guide to Understanding and Selecting the Right Descriptive, Predictive, and Prescriptive Analytics.