I’ve been writing quite a bit about Control Charts with Stages around here because I think they’re such a great visual tool. There are many uses for them (see four of them here) and they’re very easy to create. In addition to helping analyze a process before and after an improvement or change, they can be a great first step to more accurate predictions or forecasting.
There are many different ways to make predictions. If you’re looking to analyze a trend in your data – whether it be sales forecasting or the weather – time series helps you forecast how any variable might change over time. If you’re looking to predict a response using different variables, then predictive analytics, like regression, tree-based methods and many more algorithms might be the right choice for you.
Here is an example of a manufacturer of hand sanitizer that is looking to forecast demand to plan manufacturing capacity. Taking data over the past few years, the business analyst, let’s call him Josh, tries to create a simple trend analysis.
As he’s not sure what type of data he has, he creates a linear model and a quadratic model (both below). Now he’s in a pinch! The linear model predicts sales growth, while the quadratic model predicts a sales decline. The quadratic model has better accuracy measures (as evidenced by comparatively smaller values of MAPE, MAD and MSD), but his sales team has been positive about their sales pipeline all quarter. What should he tell the head of manufacturing?
The business analyst quickly calls his company statistician Cheryl and asks for help. Cheryl explains that using single exponential smoothing can provide short-term forecasts and works well for data without a trend or seasonal component.
Though Josh the analyst has never heard of this, he quickly finds single exponential smoothing in the Minitab Time Series menu and voila! He appears to have a much more reasonable forecast and a more accurate one than the previous two.
He calls Cheryl and shows her the graph. She points out the 95% PI on the graph and explains that with a 95% PI, Josh can be 95% confident that a single response will be contained in the interval.
That means future sales could be higher or lower (based on this projection). She also mentions that while the MAP, MAD and MSD measures are better than his previous models, the lower these measures are high, which means the accuracy could be weak.
Disappointed, but resigned to having built the model the best he can, his phone rings. It’s Cheryl. She reminds him that before simply jumping into analysis, it is important to think about and explore your data. She asks if there’s any way to explain some of the large shifts in his data. And then it hit him.
As a hand sanitizer manufacturer, demand spiked during COVID! Cheryl recommends making a staged control chart. Charting the demand before, during and after COVID, the change in the data makes more sense.
Josh also notices that the demand appears to be trending up, so he decides to apply his forecasting techniques to his post COVID data.
Whereas his original model projected sales at 85 to 86 million units, his new more accurate projections (based on a much lower MAPE, MAD and MSD), now show projections of 85 to 88 million units. This means he could have significantly underproduced had he followed his original model. He doesn’t need to run a smoothing analysis because there appears to be a clear trend, but he does anyway – and finds a less accurate model. He phones up his boss and tells him they need 86 million units to be produced and should plan to expand production to 87 and 88 million over the next few months.
When forecasting, business understanding is just as critical as the forecasting techniques. Whether or not you’re using Time Series or a different predictive model, using a staged control chart where appropriate cannot only help identify a better data set to forecast, but also help explain how and why the forecast came about.
Luckily, Minitab’s Assistant has made it particularly easy for anyone to create and see whether a process is within control limits, to confirm that observation statistically, and to see whether a change in the process results in a change in the process outcome or variation.