The world of replenishment techniques is changing rapidly. After the development and deployment of classical ROP (Re-order Point), MPS (Master Production Scheduling) and DRP (Distribution Requirements Planning) techniques for many years, recently there is a drastic change in the adoption of more advanced techniques for replenishment using TOC (Theory of Constraints) and Machine Learning principles. The TOC principle came from the fashion retail industry, which needed faster techniques to adopt to short life cycle products and rapid changes in consumer buying behavior.
All techniques can be classified into the following categories:
- Classical Techniques using Simple Netting Logic such as ROP, MPS and DRP,
- Heuristic based Techniques such as Dynamic Buffer Management derived from TOC, and
- Advanced Techniques based on Artificial Intelligence such as supervised and unsupervised Machine Learning algorithms
We will explore now each category.
The ROP, MPS and DRP Model
First there is difference between ROP and MPS/DRP techniques. ROP is a static calculation waiting for the inventory levels to drop to the re order point and then use the EOQ (Economic Order Quantity) calculation to reorder. This technique is completely reactive and cannot anticipate future demand changes. The EOQ for example assumes stable demand. Both the MPS and DRP logic are based on future projections and can anticipate based on a demand (mix of forecast and orders) the evolution of the inventory levels of an item at a location.
These models are based on a basic netting logic, netting inventory in a location using cycle and safety stock principles. Safety stock is derived from desired customer service levels as well as the demand and supply variation of the item at this location. Cycle stock is the replenishment cycle, driven by lead time, minimum and incremental order quantities (using the EOQ approach for example). The logic is simple. The demand is put on the location for an item, the on hand is checked and for the ROP at a single point in time and for the MPS/DRP a future projection is calculated and a planned order (ROP) or a set of planned orders (MPS/DRP) is generated. This technique attempts to keep the inventory levels within a set minimum and maximum boundary.
The difference between MPS and DRP is also simple. The MPS works on a single location. The DRP works on a network of locations (static sourcing matric) with lead time and transport constraints connected certain locations with each other. Hence the outcome is:
- In the MPS world parameters such as safety stock are valid per item
- In the DRP world parameters such as safety Stock are valid per item at a location
The Dynamic Buffer Model
The dynamic buffer model is derived from the TOC technique of managing inventory buffer of work to protect the bottleneck resource. The objective is to always keep the bottleneck resource busy with productive work. Used in replenishment and fashion retailing, the productive work objective is replaced with a product availability objective in a store or DC. This is the same use of inventory. The key is the composition of the content of the inventory buffer.
Dynamic buffer management recalculates the “speed” or an item daily and adjusts the inventory buffer in a store or DC to the changing “speed”. The “speed” is described by a daily Pareto analysis, categorizing each item into fast mover, slow mover and non mover (for more details please refer to the Article: Assortment Performance Management).
This model is based on a daily recalculation based on sales at each consumption point (store or DC). One enhancement is to add a forecast to allow the model to anticipate and not just react to the daily sales. In addition, anticipated changes in the assortment due to changes in the range plan, need to be considered. Retailisation and ebp Global have launched an Integrated Assortment Planning tool, which includes a range management capability. This allows such range changes to be planned and executed to the assortment at store level.
One of the keys to effective replenishment is substitution management. In order to satisfy a customer demand, one does not always have the sell the item desired. A substitute product will do as well. Recent developments have added algorithms to the arsenal, which will consider not just the straight replenishment needs of an item, but also the substitution effects. Based on a data driven approach, which Farahat and Lee have developed, they call this technique the approximate similarity transformation. Quote from Farahat: “This algorithm recognizes that there is a relationship between how much retailers stock and their profits. This relationship is complicated, so we replaced it with a simpler one that provides an upper bound on sales, but it is a tight upper bound. Dealing with that simpler, yet approximate, sales function leads eventually to better decisions”.
The Machine Learning Model
Over the last couple of years machine learnings has entered the world of replenishment as well. This is using either the unsupervised (non learning) or the supervised (learning) approach. Both techniques rely on data being either analyzed (unsupervised) or used for training (supervised) to determine the correct replenishment quantity and timing. There are a number of players which have built capability in this space. These vendors are operating today primarily in the Grocery industry for fresh food forecasting and replenishment, some of these vendors are:
- Black Swan
However, these vendors rely on big data, with a heavy emphasis on analyzing and predicting demand, hence they work with a mix of
- Sales History, and
- Causal Factors such as weather etc
Based on those inputs that they operate a push system to the DC and stores in a fixed replenishment network without space constraints. There is a lot to enhance, but the base is the more the worth exploring as the future of replenishment.
The next Frontier of Replenishment
Except Manugistics in the past which used a dynamic deployment logic to redistribute inventory amongst location dynamically and in the short term based on changing demand patterns, no one has addressed the dynamic nature of demand and inventory deployment. Any of the techniques described above are not properly addressing the deployment of inventory within a location network dynamically. Only Retailisation has explored the redistribution of inventory from one store to another, or from one DC to another. In some store and DC networks the under- and over coverage of inventory against demand happens fast. This is due to the nature of demand changing rapidly and unexpected.
Retailisation has developed a technique to analyze this under and over-coverage daily and to propose transfer orders to re-distribute this inventory. Often the replenishment lead times from the DC are longer than the ones from store to store. Hence using stores as mini hubs in the network, from which the redistribution takes place, is a valid approach.
How to select the correct Method?
Any use of the above techniques required an analysis of the sales and inventory data for each item and for each location. Depending in the “speed” of products and network configuration of locations and lead-times, one of the above methods has to be selected. This process involves data analysis and experience in the performance of the different methods.
ebp Global used its standard analysis and selection process to support companies with the above.