Digital Transformation
Jun 22, 2018

Inventory optimization in the cognitive era

How AI and machine learning can get the most out of working capital

A recent PwC study on inventory performance reveals that inventories in the US have grown by more than 6% annually since 2010, while inventory turns have steadily decreased from 9.56 to 9.2. In contrast, GDP growth has only been 2.5% each year. So, while manufacturers have focused on managing inventory, they seem to have reached a point of diminishing returns. To combat this, manufacturers are turning to advanced information management solutions, with almost all major manufacturers today licensing one or more inventory management applications.

Although I've been part of many end-to-end supply chain software implementation projects, I've yet to see full-fledged use of all the capabilities of inventory planning software. Planners frequently bypass or overwrite the safety stock numbers the software generates with their own sets of figures. In an age when we have exceptional computing power and highly skilled planners, this shouldn't be happening.

Genpact estimates that only 30%–40% of the world's top CPG companies comply with the stocking policies that best-of-breed platforms recommend.

Calculating optimum stock levels

Illustrated below (figure 1) are the seven steps generally taken to find the correct safety stock parameters for a manufacturing firm.

7 step approach to finalizing stocking policies

The first three steps – data acquisition, SKU classification, and SKU-inventory strategies – are straightforward for CPG companies, leaving little room for error. Their supply chains are primarily make-to-stock and their inventory strategies are either material requirements planning, or reorder point control based.

Most companies falter at step four, where demand analysis comes into play. And that in turn affects step five, the calculation method. 

What's interfering with accurate demand analysis?

Nearly all organizations – and the applications they use – assume that demand and risk are by their very nature arbitrary, so they are represented by random variables that follow a specific probability distribution.

The most common form of distribution assumed is normal distribution, characterized by its mean and standard deviation. The problem is that this assumption of normality often applies only to the top-selling, highest volume items – A class – while products on the long tail – B and C class – generally don't exhibit normal demand behavior.

Moreover, in an environment where demand variability is high and demand distribution is skewed, classic demand and inventory models don't perform well. True, the assumption of normality simplifies the math. But it also results in a significant gap between the real demand pattern and the probability function used to model it. That leads to a misalignment of inventory mix – and therefore service levels – across the network.

To compensate for poor inventory performance that results from non-normal, intermittent, or lumpy demand, planners often manually add inventory to cover for the variability across SKUs. While this excess stock rarely accumulates for fast sellers, it typically accrues in the long tail, with slower moving products. As a result, companies lose working capital and don't get the desired service levels they want.

Non-normal demand is becoming the norm 

Supply chain developments are making it even harder for planners to classify demand. Reasons for this include:

  • Increased product proliferation: Manufacturers are expanding their product portfolios with new packaging options, sizes and colors, which divides demand into smaller buckets
  • More frequent replenishment and daily demand planning: The same SKU may look like a fast-mover with relatively stable demand if demand is measured in monthly time buckets. But from a weekly or daily standpoint, the product may track as a slow mover, with demand that is intermittent or lumpy
  • Additional stocking locations: Demand is increasingly being disaggregated into smaller streams, as the focus of replenishment planning shifts from primary distribution centers to secondary distribution centers and retail shelves

How digital can predict more accurate stock levels

Inventory planners can adopt a three-point approach to reduce overstocking issues.

  • A reliable demand classification technology will automatically adjust all the relevant statistical parameters, seamlessly, across a wide range of SKU behaviors. The technology should also account for all elements of demand uncertainty, including variability caused by order-line frequency and order-line size distribution. Traditional machine learning, and now deep learning algorithms, can automatically classify underlying demand patterns into one of the appropriate demand categories
  • Multi-echelon inventory optimization (MEIO) eliminates the gross approximations of traditional inventory management approaches to optimize very large assortments, including long-tail stock. It optimizes safety stock buffers across the supply chain and considers interdependencies between stages and variables that cause chronic excess inventory, such as long lead times, demand uncertainty, and supply volatility. The goal of MEIO is to continually update and optimize safety stock levels across all the echelons of a supply chain
  • Having a disciplined operational process in place helps eliminate manual interventions and bullwhip behavior. This starts with providing visibility into the critical parameters necessary to validate and approve stocking policies, such as coefficient of variation, inventory segmentation class, and deadstock percentage

We need advanced inventory planning software to solve complex supply chain problems. But when these problems are incorrectly modeled, or when that modeling is based on incorrect data and process assumptions, the software will not be able to deliver the expected value.

About the author

Vikram Sethi

Vikram Sethi

Assistant Vice President, Genpact Supply Chain Practice

Follow Vikram Sethi on LinkedIn