Achieving Your Target Service Levels with Safety Stock – What Would You Do?

Let’s say that inventory item XYZ is one of your items. It’s a purchased component and an “A” item.  XYZ’s current on-time-delivery (OTD) performance is 80%, and its target OTD is 98%. We need to do something – but what? For starters, let’s get some more data.

Here’s what XYZ’s daily historical demand, or usage, looks like, going back almost 18 months:

A visual test of this data suggests at least 4 unusual demand spikes (the points at or exceeding 200), but that’s subjective. In this case, a good statistical test for outliers confirms that only 3 points are unusual.

So did these spikes cause or contribute to the current 80% OTD? That could be hard to ascertain. What we do know is that achieving OTD involves a combination of demand visibility (backlog), forecast accuracy and safety stock. Of course, if you’re reading this safety-stock blog, you likely don’t have the luxuries of demand visibility and forecast accuracy at the item level. Much, most or all of your demand variation is unforeseen far enough in advance to plan for, and therefore effectively random: You know it will happen, but you don’t know when. This means you will have to carry inventory continually to insure against OTD failures – safety stock.

Safety stock is in itself a forecast. It assumes that historical random demand variation predicts future random behavior. So in determining optimal safety stock, it is critical to isolate and use the portion of historical variation that is effectively random.

We might investigate XYZ’s demand spikes. Let’s say we find that the 3 largest spikes were due to a special cause that we could forecast well in advance, by means of better forecasting techniques, more effective S&OP communication, etc. Let’s see what XYZ’s daily historical demand looks like without these 3 spikes:

Now our subjective visual test of this data suggests a positive trend. In fact, there is no significant trend, which emphasizes the importance of using objective data analysis. Our eyes also tell us that demand behavior changed in about the mid-December 2011 time frame, becoming more intermittent and spiky. In any case, our goal in “cleaning” this historical demand data is to obtain random variation with unpredictable timing, so that we may use this data for setting safety stock that will optimally achieve our service-level targets.

Does this data need more analysis? Is it ready to be used in a correct safety-stock-optimization analysis? Do we need more information?

Really, what would you do?