How Do You Optimize Your Inventory for Highly Variable Demand?

August 31, 2017

LMI Staff

Some items are unforecastable. Either infrequent demand or frequent highly variable demand makes them impossible to accurately predict. The Defense Logistics Agency (DLA) faced such a quandary.

LMI’s Advice: Stop Forecasting

Instead, LMI developed a new tool for DLA—PNG. This tool has a set of capabilities that employs risk management to optimize inventory.

“For some items, customer demand is so variable, you can’t draw a trend line. If you try to forecast, the error will be large,” explained Michelle Kordell, a director in LMI’s Logistics Analysis group. “You end up with the wrong items on the shelf—too much of what you don’t need or backorders on what you do need.” DLA found itself in precisely this situation—overinvestment in unused inventory, long wait times for critical parts, and buyers issuing procurement requests (PRs) but still not having the right stock on hand.

PNG is similar to a productivity software suite in that it is comprised of two elements: Peak Policy and NextGen.

LMI spent a decade and $6.5 million developing these two new models for DLA. Peak Policy optimizes inventory levels for items with infrequent demand, while NextGen solves the challenge of optimizing inventory for items with frequent demand but high variability in quantity.

Balancing Wait Times and Investment

PNG employs a risk-based hedging strategy to set inventory levels using demand history and item characteristics. LMI inputs five years of demand data for each item into the models. Item characteristics include demand frequency, price, lead-time, and asset position. The models generate tradeoff curves that balance the risk of being out of stock against over investing. The curves plot the tradeoffs among wait time, inventory investments, and procurement requests. The supply chain manager picks the point on the curve best suited to mission and budget, and PNG calculates minimum and maximum levels for each part.

DLA rolled out PNG in January 2013 to establish minimum and maximum levels for a subset of unforecastable Class IX weapon system repair parts, which comprise 485,000 parts—340,000 in Peak and 145,000 in NextGen—representing a baseline of on-hand inventory worth roughly $3.4 billion.  For this initial population, DLA selected points on the curve that target a reduction in its inventory by $180 million and trimming procurement workload by 50 percent—a 27:1 return on investment for the initial implementation.

Inventory Already Down 12 Percent

The implementation is already generating significant benefits. “As of July 2013, the on-hand inventory value for the same population was down 12 percent over the 2012 average,” reported Kordell. PR cancellations are down 40 percent in the first four months of implementation, and monthly PR generation is down 10 percent in the most recent two months.

DLA achieved this inventory reduction despite the fact that PNG causes a short-term spike in new PRs. “Initially, there’s an increase in new procurement requests because you weren’t putting the right items on the shelf,” she explained. “We saw that for the first two months, but now it’s coming down, and PRs are already below the baseline.”

Why was this capability unavailable 10 years ago? “We didn’t have the computing power a decade ago,” Kordell said. “The computing power necessary to generate the optimization choices and develop the tradeoff curves is significant.”

The math algorithms are also new. LMI’s team tied together two disparate mathematical calculations into a new risk-management-based algorithm. “We also made sure that there were no other existing methodologies that could be improved. We didn’t just throw away forecasting. We first exhausted every forecasting methodology available,” she said.

Rigorous Testing Proves Models

After developing the algorithms and model, LMI rigorously tested PNG by analyzing volumes of historical DLA data. The team assessed PNG outputs and compared the models’ suggested max and min levels to demand that DLA experienced. “We also ran PNG against many other scenarios to see how we compared to what the others would have forecasted. We did very well,” she added.  

For the initial implementation, LMI generates tradeoff curves yearly and min/max levels quarterly. Five DLA supply chains, each with a unique tradeoff curve, currently use PNG, including

  • Aviation
  • Land
  • Maritime
  • Industrial Hardware
  • General and Industrial/Construction and Equipment.

DLA has moved to monthly generation of min/max levels to account for changes in input information.  “If there’s a big price or lead-time change, that may impact what we recommend for the min and max levels,” Kordell noted.


For entities seeking to optimize inventory of infrequently demanded items and frequent but highly variable items, Kordell had this advice:

  • Identify which items are suitable for forecasting
  • Stop forecasting unsuitable items
  • Supply LMI with a sample data set
  • Assess potential benefits derived by PNG
  • Remember that automating a bad process won’t improve the supply chain.

PNG: Inventory Control for Highly Variable Demand


White Papers and Case Studies

Please fill out the form below in order to download the document.
Yes, I’d like to receive future LMI news, trends, and updates. LMI will never sell or distribute your email to any third parties.
Privacy Policy 

Contact Us

Yes, I’d like to receive future LMI news, trends, and updates. LMI will never sell or distribute your email to any third parties.
Privacy Policy