Decision-driven optimization

Quantitative Supply Chain focuses on generating automated high-performance supply chain decisions. The focus is not to deliver numerical artifacts , such as weekly forecasts. Those artifacts are considered as arbitrary internal calculations that just happen to be used to compute the final decisions. We define a decision as an answer to a supply chain problem that can be acted upon and has a tangible, if not physical, consequence for the supply chain itself. From a classic supply chain planning perspective, focusing on supply decisions might appear somewhat surprising, because they are not defined along the usual lines of planning vs. operations. Nevertheless, focusing on decisions dramatically facilitates the actual optimization of the supply chain. In this section, we clarify the concept of a supply chain decision, review the most frequent types of decisions, and characterize key aspects of the decision-driven perspective.

Scoping the eligible decisions

Quantitative Supply Chain takes a highly numerical and statistical stance on supply chain challenges. However, this stance isn’t the appropriate perspective for all challenges. In order to assess whether the quantitative perspective is appropriate for a challenge, the following conditions should be met:

Repeatability : crafting a numerical recipe to solve the challenge takes effort that translates into costs. In order to profitably optimize a supply chain, one needs to make sure that the optimization process itself is not costing more than the benefits expected from it. As a rule of thumb, routine problems, e.g., replenishment, which need to be addressed every day or every week are much better candidates for a quantitative approach than exceptional problems, e.g., expanding to a new country.
Narrow decisions: in order to keep the complexity of the software solution in control, it’s favorable to focus on supply chain challenges that can be addressed by a well-defined typology of decisions, ideally, highly numerical decisions. For example, deciding whether to stop stocking a product altogether because demand is too low to justify this extra burden on the supply chain is a very narrow question, which a highly automated process can readily answer. In contrast, deciding to amend the work practices of a warehouse management team is a very open-ended problem, which is a poor fit for automation.
Historical data: software solutions cannot operate in a vacuum. The knowledge to address the supply chain challenge can be embedded into the software as manually defined rules; however, creating a large body of consistent and performant decision-making rules is a very difficult undertaking. Most modern approaches extensively extract all relevant knowledge from the historical data (sales history, purchase history, etc) and restrict rule entries to well-defined supply chain policies, e.g., MOQs (minimum order quantities), which we certainly don’t want the software to try to extrapolate from the historical data.

As software engineering progresses and, more specifically, as the field of machine learning progresses, the spectrum of decisions that becomes accessible to computer- based systems is broadening every year. For example, the early inventory optimization systems were limited to products with at least several months of sales history, while the newer systems support all products, including those that have not even been sold yet.

Also, sometimes software engineering makes it possible to tackle problems that were considered to be intractable when done manually by a supply chain expert. For example, modern inventory optimization systems can predict which stock records are most likely inaccurate, hence allowing a prioritized recount of the inventory, a feature that outperforms the more traditional approach of linearly recounting all the SKUs.

Examples of supply chain decisions

Supply chains are incredibly diverse, and what constitutes a challenge of primary importance for a given vertical, might only appear as anecdotal in another. In this section, we briefly review typical decisions that are a good fit from the Quantitative Supply Chain perspective.

Purchase orders : deciding the exact quantities to be purchased from each supplier for every product. This decision is refreshed on a daily basis, even if no actual purchase order is expected to happen most days. The purchase order should take into account all the ordering constraints (MOQs) as well as the transportation constraints (e.g., containers). Also, the purchase order might also include opting for a transportation mode (sea vs. air) with the possibility of a transportation mix .
Production orders : deciding the exact quantities to be produced. The production order should take into account all the production constraints that may require minimal production batches. Also, the maximum production capacity might be lower than the market needs of the peak season during the year, in which case, the production should build up stocks ahead of time to cope with the peak season.
Stock balancing : deciding whether units currently held in stock in one location should be moved to another location, typically because the stock balance isn’t aligned any more with the projected future demand differentiated by location. Again, the decision is refreshed on a daily basis, even if on most days, for most products, it’s not economically profitable to move them between locations.
Stock liquidation : deciding whether units currently held in stock should be either destroyed or sold through a secondary - typically highly discounted - channel. Indeed, dead stock can needlessly clutter warehouses and thereby generate costs greater than the economic value of the stock itself. Depending on the vertical, stock can either be liquidated through promotions, specialized channels or pure destruction.
Stocking vs. drop shipping : deciding whether a product is in sufficient demand to justify its being purchased, stored, and served directly or whether it would be better to have the product drop-shipped by a third-party when it is requested. Drop-shipping products typically generates lower margins, but it also incurs less carrying costs. The decision takes the form of defining the exact list of products to be kept in stock, while keeping overall inventory diversity manageable.
Targeted stock counting: deciding whether a SKU should be recounted because of the potential inaccuracy of the electronic record , which may not match the amount of units that are really available on the shelf. This decision is a trade-off between the labor cost associated with the recount operation and the negative impact of phantom inventory on the supply chain performance. In practice, inventory inaccuracies are much greater in publicly accessible retail stores compared to staff-restricted warehouses or plants.

It should be pointed out that specific verticals have their own sets of decisions. The examples below could be considered more context-dependent than those listed above.

Retail store assortment : deciding the exact list of products to be present in each retail store. Sometimes, the full product catalog may vastly exceed the capacity of any given store; hence, each store can only put a subset of the catalog on display. The optimization of the assortment maximizes the performance of the retail store given the store capacity. Also, the challenge is made even more complex in the case of verticals such as luxury goods, as the store will typically have no more than a single unit in stock for each product of the selected assortment.
Opportunistic replacement: deciding when the substitution of a product is acceptable and when it is profitable to proceed with the substitution . For example, a fresh food e-commerce may accept deliveries being performed a few days ahead, a practice which creates the problem of facing a late stock-out for a fresh product that has been already ordered, and thereby alters the original client’s order. In this situation, it might be a more profitable operation for the retailer, and a better service for the client, to substitute a well-chosen alternative product.
Opportunistic divestment : deciding to resell inventory, typically repairable parts, which were originally intended for internal consumption. The stock of repairable parts typically rotates between the two states of serviceable and unserviceable , as parts get serviced, retrieved back, repaired and finally re-serviced. Under specific circumstances such as a drop in demand, the stock of serviceable parts may vastly exceed the company needs. In this case, there is a tradeoff between reselling the part in the aftermarket , typically at a discounted price, in order to recover a fraction of the original inventory value or, alternatively, increasing the risk of not servicing a future part request on time..
Keeping unserviceable stock: deciding to repair immediately an unserviceable but otherwise repairable part, or to postpone the repair and store the part as unserviceable. While repairing parts can be less expensive than buying new parts, the current stock of serviceable parts may be sufficient to cover the demand for a long period of time. Hence, delaying the repair is a tradeoff between offsetting the repair costs to the future - with the possibility of never incurring this cost if the market demand has shifted to alternative parts in the meantime, or increasing the risk of not servicing a future part request on time.
Opportunistic sourcing : deciding when it’s worthwhile to perform a sourcing operation to establish a pricing benchmark for a given part. In some industries, the price of parts is relatively opaque. Uncovering the up-to-date price of a part, possibly a very expensive piece of equipment, can take several days of effort. When operations require thousands of parts, there is a tradeoff between paying for more expensive parts and incurring the manpower costs involved with the sourcing operations.
Preserving bundles : deciding when it’s worthwhile to sell the last unit of a given product as a stand-alone sale or better to preserve this unit for a later sale as part of a bundle. Indeed, there are situations where the availability of bundles , that is combinations of parts or products, is of great importance, while the availability of isolated parts is of lesser importance. Yet, by serving the last part, serviced as a stand-alone part, one can create a stock-out issue for the bigger, more important bundle. Thus, there is a tradeoff between the upside of properly servicing an isolated part now and the downside of facing a later, more impacting, stock-out issue for a bundle.

Until formalized as such, supply chain decisions are usually made rather implicitly , possibly by people, but also by software systems. For example, a Min/Max inventory configuration is implicitly making multiple decisions and not just about the reordered quantity: as long as the Max value is non-zero, the product will be kept in the assortment. Also, no inventory recount happens before triggering a replenishment, which is another implicit decision, etc. Unfortunately, as you can’t optimize what you don’t measure, it’s this lack of formalization of the decisions themselves that typically prevent a systematic improvement of the supply chain performance obtained through those decisions.

Numerical artifacts vs. decisions

When facing complex supply chain problems, practitioners risk confusing the ends and the means. For example, when facing a need for replenishment, establishing a weekly demand forecast associated to a SKU is a mere ingredient required by some, but not all, numerical recipes available to compute the quantity to be reordered. The weekly forecast is only an intermediate calculation, while the ordered quantity is the final decision. From the Quantitative Supply Chain perspective, we refer to those intermediate calculations as numerical artifacts. Quantitative Supply Chain does not dismiss the importance of numerical artifacts; yet, it also emphasizes that those artifacts are just that: disposable, transient numerical expressions that contribute to the final output: supply chain decisions.

As far as numerical optimization is concerned, it is a fallacy is to think that optimizing the numerical artifacts against arbitrary mathematical metrics , e.g. demand forecasts optimized against WMAPE (weighted mean absolute percentage error), somehow mechanically yields financial returns. While this might appear as counter-intuitive, in supply chain, this is usually not the case. Supply chain problems are typically high asymmetric problems. For example in aerospace, a missing 200 USD part can keep a 200-million USD aircraft on the ground. The number of parts to be kept in stock is not necessarily primarily driven by the expected demand: the cost of the part compared to the cost of not having it can completely dominate the stocking decision process.

In contrast, Quantitative Supply Chain emphasizes that, in the end, only decisions truly matter, because they are the only tangible elements that have real and measurable financial consequences on the company. Thus, while it is of primary importance to challenge the performance of the decisions, the supply chain management should also have a healthy dose of skepticism toward KPIs that apply to non-binding, non-committal, transient numerical results such as weekly or monthly demand forecasts.

Constrained decisions, between reality and fiction

Supply chain decisions are typically bound by constraints : answers are only valid if they satisfy a set of numerical constraints. For example, purchase orders may be subject to MOQs (minimum order quantities), which represent a nonlinear constraint. Also, the warehouse has a finite storage capacity - another nonlinear constraint.

Frequently, the constraints are generated from basic economic drivers associated to supply chain operations: taking the current product price point into account; the distribution of a product can only be economically viable if the products are sold packed by pallets and, hence, the product can only be sold with a lot size of, say, 50 units, which represent a loaded pallet.

However, it also happens that the constraints can result from arbitrary organizational rules. For example, a company might have decided that the yearly purchasing budget for a division would be capped to 1 million USD. This budgeting constraint is established long before the sales of the division are actually known. In such a situation, the purchasing decisions are expected to comply with a nonlinear constraint that is the result of a relatively arbitrary budgeting process.

Quantitative Supply Chain tries to reflect to the greatest possible extend the real supply chain constraints, while enabling newer, possibly revised, organizations that can operate without the shackles that were imposed by arbitrary aspects of former processes. Indeed, in supply chain, most of arbitrary constraints are the result of a lack of automation: if the “optimal” budget per division cannot be reliably re-estimated on a daily basis taking into account all the transverse concerns company-wise, then it is natural to fall back on a yearly or quarterly budget instead.

Decisions require prioritization and coordination

Nearly all supply chain decisions are interdependent: every extra unit that is purchased from a supplier is going to take extra space into the warehouse, until the warehouse is full, and then operations grind to a halt. Those dependencies are usually indirect, and difficult to tackle from a numerical perspective, but that does not make them any less important from a supply chain - and even strategic - perspective. If the overall service level is at 99%, which is very good, but the largest client suffers from a 85% service level, because all the stock-outs happen to be concentrated in group of products purchased by this very client, the company is facing a serious risk of losing its largest client.

Prioritizing decisions is typically the most straightforward method to make the most of the shared, but limited, resources within the supply chain. For example, as warehouse storage capacity and working capital are both limited, the goal is not merely to buy one extra unit of stock that happens to be profitable, but to identify the next one unit of stock that happens to be the most profitable unit across the entire product catalog. Treating the stock purchasing decisions in isolation would create the risk of exhausting the warehouse space, or the purchasing budget on low profitability products.

In practice, this prioritization requires changing considerably the analytical software that supports the supply chain. Instead of treating each decision in isolation, as it is the case with primitive supply chain methods, e.g., Min/Max inventory , all decisions need to be put together and ranked against their respective estimated profitability. Such a process is feasible with modern software solutions, but it does require considerably more computing resources compared to early supply chain methods.

Coordinating decisions is required to handle all transversal constraints that apply to supply chain operations. For example, when ordering goods from an overseas supplier, there might be a strong economic incentive to order a full container. Thus, the challenge is not so much choosing quantities per product but choosing quantities that, in aggregate, exactly match the container capacity. Transversal constraints are ubiquitous in supply chain: tuning the assortment of a new collection in fashion, ensuring high service level for clients seeking a list of products within a DIY store, not depleting a central warehouse through oversized orders from one store at the expense of the other stores, etc.

The traditional and highly inefficient way of tackling such coordination concerns consists in performing a two-stage calculation that firstly ignores the coordination concern, and secondly, that revises the initial numerical output in order to fit the concern. Regarding the container example introduced above, first, we can compute the desirable quantities to be ordered, ignoring the container angle entirely; second, we can revise those quantities so that the aggregate actually fits into a container. The main weakness of such a two-stage calculation is that the second stage completely ignores all the economic drivers that went into the calculation of the first stage. In other words, the revision of the results during the second stage can “undo” all the efforts that went into computing profitable decisions in the first stage. Modern software tackles such situations by introducing numerical solvers, which can frontally address such transversal constraints. Once again, those solvers are dramatically more demanding in term of computing resources than their naive two-stage counterparts, but again, considering the computing resources typically available nowadays, this is a non-issue.