The Math - What Really Matters?

In our last article, we teed up the difference between known data that need to be captured and the unknown metrics that need to be modeled.? As you can see in the illustrative formula above, the three big areas of unknowns to model are Service Levels, Demand, and Lead Time.

Optimal Service Level Targets

The service level target, or z-score above, represents the in-stock goals for a given item. This is the percentage of time that a given inventory level will accommodate. When utilizing the formula above, we assume that these demand and supply scenarios are normally distributed. This means that the most common scenario is the average, and demand situations further from the average are less common and less likely. Also, the service level target drives an exponential increase in the safety stock required, slight changes to this value can have extreme changes to the safety stock result. If your goal is to improve from 99 to 99.5% in stock, that is not a 0.5% increase in inventory. It could be double. ? General guidance on service level targets are to follow a priority ranking of SKUs for their importance to the business.? This should consider things like how profitable an item is, how much it can be substituted with a similar product, and the criticality to a customer’s operations. ‘A’s are then defined with a high service level target (99.5% SL for our top 10% of SKUs) where ‘D’s are defined with a less aggressive goal (e.g. 95% SL for bottom 40% of SKUs). Sensitivity analysis is important for your team to understand here and a big gap we often see.? What happens if I change the target from 99.5 to 99.8? In some cases it could potentially double the safety stock required, depending on variability assumptions.

Demand and Demand Variability

In supply chain planning, perhaps the most focus is typically put on demand / forecasting as a capability. Supply planning is straightforward math that usually lives behind the scenes.? 20 years ago there was a lot of talk about how this math should work but now its behind the curtains of planning software. Whereas Demand planning is more of an art that requires an understanding of the market, your customer, and the stochastic forecasting methods available.? At its root, demand planning is as simple as guessing the demand for coming days/weeks/months on a timeline that makes sense.? Some challenges we experienced and overcame in our work recently were:

·?????? Demand variability vs. error - in the strict statistical definition above, demand variability represents the magnitude of ‘noise’ in the demand for an item. How often do I deviate from the mean? How far from average are we?? A more intelligent estimate of Demand Variability is attempting to estimate the amount of error we expect to see relative to the forecasted demand.? This is often measured from historical values but this historical forecast error, on the right horizon, is classically the most common gap we see in data available. Because…

·?????? Systems don’t hold forecasted values on the right horizon – Most systems we look at today calculate some variation of a Mean Absolute Percentage Error with a 0 to 1 week lag. ?This lag time is where the issues arise.? Think back to the example of your own home purchasing eggs from the previous article. This is an item you can procure with a 15-minute lead time, to drive to the store. Thus, your personal ‘forecasting’ task is to guess how many eggs you will consume in the next week, when you are already in that week.? You may already know what meals you plan to have, how many people will be home each morning, if you have guests coming over, etc.? Now think about forecasting the demand for items that have a two-week lead time. Your ‘forecasting’ issue has just become harder because of the planning horizon you are looking at.? This is how forecasts need to be held, with variability by product and aligned with their cycle time and lead time.? As a result, we often see businesses underestimate demand variability for long-horizon items and overestimate variability for easy-to-procure items, causing shortages and overages at the same time. Historical forecasts for items need to be held and referenced to their own LT and order cycle.? We found that this was an error for over 75% of the items planned. ?

·?????? Estimating future error, not just measuring past error – the other major thing to think about is that the goals of supply chain planning are to estimate the future performance not to model the past.? However, the past is likely a great tool.? We see companies become very focused on measuring history at great detail and for long periods of time, without the ability to connect with sales, communicate where risk could be, and incorporate that into the system. This has been a major concern when we think about how wild and unpredictable demand has been for the past handful of years. An item could have had stable supply and demand for decades, but a long-run analysis including 2021-2023 would show it with extreme volatility. Because of this, sales input and customer connection are a more valuable investment than they have been in the past. Supply chain leadership need to make a case for the operational efficiencies and dollar value improvements that can be made with a better connection of demand planners with sales and/or customers.

Lead Time & Lead Time Variability

The second focus for our analysis is typically on lead time and supply disruptions. Historically, this is a large concern and a big focus.? Recently, this is an enormous-disasterous-mess! and where we see a lot of classic misunderstandings come to rest.? The first challenge is modeling the actual lead time.? We can estimate this from simple PO data to show when an order was placed and when a product was received and available in inventory.? There are often misconceptions about an order being placed/confirmed/shipped/received.? But the logical concept we are trying to model is the time it takes from finalizing the decision on an order to when that product is available to be consumed for customer demand.? This cycle has a lot of dependent items within it and cannot be just assumed to be transit or shipping time.? Modeling total lead time has a large impact on the way target safety stock levels are calculated and should be continually monitored.? We are seeing systems that are automatically flagging deviations, positive or negative, to help buyers prioritize where a product or vendor may be modeled with an incorrect lead time.

Modeling the variability of lead time from a supplier is another major challenge right now, given the supply disruption risk that we have felt over the past few years.? The formula above is designed to model deviations from an average lead time on a normal distribution.? If I expect a confirmed order in my purchasing system to be available on my shelves in 4 days, how often is it actually available in 5, 6, or 7 days.? This could be due to a supplier system delay, an internal approval process, or traditional shipping delays.? The more complex scenario is when there is a project risk of a product not being available from a supplier at all.?? Just like your business is modeling the service level target discussed earlier, your suppliers do the same.? This is an area where leadership needs to understand the implied risk of a vendor and how it is being modeled.? The classic calculations of safety stock take this risk of supplier stock-outs or supplier shutdowns as a time variable: how many days do we expect to see a delay, as measured by standard deviation from a mean.? The key here is to understand the amount of variability that you expect safety stock to cover, vs. the amount of variability that would be handled by alternative sourcing error.? For example, if you expect an order to arrive in a 7d lead time, if it takes 8 days to arrive you would expect to be able to cover that one additional day.? Whether you knew it would be a longer cycle time when you placed an order or if you didn’t learn about the delay until the 7th day, you would still expect safety stock to account for this volatility.? However, if you placed an order with a 7d lead time and it was delayed 4 weeks due to a shutdown or backorder situation, you would not be expecting your inventory to cover this risk.? This would be an exceptional situation where buyers/planners would find alternative products or sourcing channels.? So, in the historical calculation, you should similarly not model safety stock with this value.? Data cleansing of historical lead times outside of the ‘should safety stock cover this’ range will allow your planning solution to estimate the right level of volatility better.

Summary

With the foundational steps above, we are able to run relatively simple calculations at a large scale to estimate a reasonable target safety stock level.? We then compare these targets to actuals or historical inventory values to show (with the benefit of hindsight) where to focus your inventory optimization efforts and what the rough magnitude of the opportunity might be.? This is likely the most valuable part of our exercise with most clients because it provides an answer to the three biggest questions ahead of taking on a change:

1.?????? What is this worth? Overall inventory reduction $

2.?????? Where do I start? Heatmap of value by location and product area

3.?????? What action is needed? Line item analysis to ID overstock items and begin root cause analysis

In our upcoming articles we will discuss the practical application of these three challenges and provide real world examples of the supply chain theory vs. a team’s goals and daily actions.