AI in Supply Chain: A Novel Risk-based Approach to Inventory Management

Inventory management is a crucial component of the supply chain, but many businesses struggle with finding the balance between having too much inventory and not enough. While machine learning solutions have made progress in demand modeling, high inventory costs still present a major issue for supply chain managers, especially in complex warehouses with slow-moving inventory and highly sporadic demand.

In this article, a novel approach to inventory management demand prediction is presented, which incorporates overstock costs and under-order costs into the decision-making process as a complement to machine-learning time-series modeling. The approach puts cost optimization and business goals at the forefront, allowing supply chain managers to customize the model based on corporate goals and targets.

1. Introduction

Inventory management, a critical component of the supply chain, involves ensuring that the right products are in the right place at the right time. This requires a careful balance between excessive inventory, which ties up capital and creates waste, and having too little inventory, which can lead to stock shortages, lost sales, and customer dissatisfaction. A key component of Inventory management which many businesses struggle with is the inability to anticipate demand, leading to over-ordering or not ordering enough. Some progress has been made in demand modeling by major ERP providers and machine learning solution companies, which is unquestionably an improvement relative to manual methods still employed by many businesses today. However, high inventory costs still present a major issue for supply chain managers who are coping with particularly complex warehouses, slow-moving inventory, and groups of items with highly sporadic demand, which all present a significant challenge for even the most advanced demand forecasting models. Global supply chain consideration for multi-national companies and high lead times for item procurement augment the complexity and accentuate the need for smart inventory planning.

In this article, we will present a novel approach to Inventory Management demand prediction which incorporates overstock costs and under-order costs into the decision-making process, as a complement to machine-learning time-series modeling. By combining statistical methods and the industry’s domain expertise regarding inventory costs, this inventory management system can predict stock levels which are optimal to the business. While achieving demand forecasting accuracy is crucial for the model’s success, this approach puts cost optimization and business goals at the forefront; the willingness to sacrifice theoretical forecast accuracy in favor of maximal business impact is what defines this risk-based approach to demand forecasting.

While cost minimization lies at the core of this approach, supply chain managers may further customize this model to better address corporate goals and targets. Additional business rules may be implemented based on domain best-practices, and the model can be adapted to prioritize other metrics, such as customer service level or other business-specific KPIs. The alignment of this approach with current company priorities and practices comes with the added benefit of being explainable and convincing to management and staff, and the high confidence level in the model is a crucial element to achieving the holy grail of demand forecasting: the ability to automate large portions of the decision-making process, thus allowing human procurement experts to shift their focus to items for which probabilistic forecast methods are at a disadvantage (e.g. due to lack of data) or where the procurement decision has extreme business consequences.

We shall commence this article by presenting a real-life example where demand prediction is a pressing concern for the business, and current inventory management demand forecasting practices fall short in providing an optimal response. We will then explain our novel risk-based approach and detail how this solution can address those scenarios.


2. Case Study: An airline’s warehouse woes 

An airline has a particularly tough time handling inventory management. With an arsenal of dozens of airplane models from varying years, the business must regularly order tens of thousands of unique aircraft parts and accessories which have a long-tail distribution and are extremely different in nature from one another. For instance, an engine is a single item often unique to a particular plane model, which is ordered sporadically, no more than a few times per year, and may remain in inventory for an extended period of time (henceforth: a slow-moving item); conversely, hex head bolts are used to fasten components irrespective of plane type, are consistently replaced in repairs, and are consequently ordered often and in large quantities (a fast-moving item). Items may be required seasonally, sporadically, consistently, or not at all; ordered in bulk or one at a time; come in packages of one or multiples; have a shelf life of a few months or a few decades; and may have many versions or subtypes based on different plane models.    

To try and maintain order in warehouses with such extreme complexity, this business employs multiple teams of procurement managers and planners whose sole role is to examine the inventory level of hundreds of item types and decide how many must be procured monthly. While an experienced procurement manager can achieve consistently high accuracy levels, achieving the required domain expertise takes years, and should an expert leave the company or be temporarily or permanently incapacitated, the more junior planners will struggle to fill the gap in experience, leading to inaccurate predictions or even disruption of inventory management entirely. In addition, the airline noted that human error takes a serious toll: the significant inconsistency between a procurement manager and staff, or between different procurement teams, led to significant variance in inventory levels, resulting in significant losses for the company. 

A partial response to this issue is provided by the company’s ERP system. In general, major ERP providers, such as SAP and Oracle’s NetSuite, focus on creating an all-in-one platform which effectively tackles many of the issues of Supply Chain, such as providing a Graphic User Interface (GUI) which addresses inventory visibility and ease of procurement. While these providers also offer inventory prediction tools, often these methods are basic probabilistic trend-capturing features which cannot incorporate domain-particular nuances; for instance, in Oracle NetSuite’s “Supply and Demand Planning” prospectus1, the customer is offered four demand prediction techniques: linear regression, moving average, seasonal average, and incorporating sales forecasts. While the latter allows for some customization on a macro level, this solution is nonetheless limited to fast-moving items with a consistent trend, and even in those cases may prove faulty: for instance, we see in Figure 1 that the moving average technique used by the airline has trouble capturing the erratic demand for diesel injector seal in 2021, with moving average reacting too slowly to sharp demand changes. A follow-up analysis showed that out of the items which were identified as potentially trending in an 18-month period, fewer than half continued the trend in the following 6 months; a statistical based approach would thus fail to accurately predict demand, leading to poor business results. 


Figure 1: Monthly demand moving averages for diesel injector seals during 2021 


Inaccuracies in high-variance fast-moving items are understandable, and near-perfect accuracy is unattainable by even the best of models, so for a company whose inventory centers around fast-moving items, even generic trend predictions have the potential to make a sizable impact, and some AI-oriented startups claim to have made progress on that front. This being said, fast-moving inventory is still prone to having peaks and rapid changes in market trends, and therefore setting the optimal stock level remains a challenge which often requires advanced trend-capturing methods, and factors such as seasonality and consumer behavior must be considered to attain reasonable accuracy levels. 

This airline, however, considers 86% of their SKUs to be slow-moving or with sporadic demand, as shown in Figure 2. As a result, methods such as regression and moving averages are ineffective, and consequently the procurement managers overrule more than half the decisions made by their ERP, implying a lack of trust in the system for their type of business and leading to a disorganized way of working. They are alone in the struggle: slow-moving items are considered one of the top concerns within the industry. An article by McKinsey states that 10-40% of OEM’s inventories are considered slow-moving, and in some industries – such as airlines or automobile suppliers – that number may be higher still. The scope and impact of this issue indicates that a different approach is required, one that can still effectively lower costs without heavily relying on accuracy. 

Figure 2: Demand distribution for an airline’s purchases in 2022


To summarize, we can pinpoint the following struggles for the airlines, which present ample opportunity for improvement if the approach to inventory management is changed: 

  • Basic trend-capturing predictive methods are insufficient for long-tail, slow-moving behavior. If accuracy is difficult to achieve (regardless of modeling technique) due to industry behavior, an entirely different approach should be considered. 
  • Low confidence in ERP predictions is a direct effect of the above, leading to a low adoption rate and a disorganized way of working. An effective cost-minimizing approach has potential to convince management to accept the AI-based predictions, accepting that the inaccurate predictions come at a relatively small cost. 
  • Manual planning leads to inevitable human error, inconsistency between procurement teams, high level of employee strain, higher costs, and the risk of losing an expert temporarily or indefinitely. Partial automation is unattainable when there is low confidence in model predictions, but an approach which iteratively implements domain expertise into the model will be far better suited to ultimately replace manual procurement predictions for a good portion of item types. 




An AI-based, risk-oriented approach to inventory management is suggested as an optimal solution to the management of complex inventories and is poised to be the future of this critical business function. This approach leverages machine learning algorithms to attempt to calculate inventory level based on prior demand, but also recognizes that not all errors between prediction and actual demand have similar consequence: in some cases, the penalty of understocking an item may be particularly steep relative to overstocking costs – in which case a sensible recommendation may be to order in bulk even if actual demanded is likely to be exceeded – whereas for other types of inventory the reverse may be true. This approach thus focuses on minimizing the errors which are particularly costly, whereas achieving optimal accuracy for lower-impact items is secondary; in other words, the forecast model with the highest possible accuracy is not necessarily that which is best for the business – and sometime the difference between the two is substantial. Achieving an optimal balance between accuracy and costs is what will ultimately lead to greater overall success in inventory management. 

The risk-based approach relies on two tenets, both of which must be strong to ensure success. On the one hand, a business must have a comprehensive understanding of the possible costs it faces, which requires deep business and domain expertise. In addition, while we have established that the most accurate inventory forecast is not necessarily the ideal one, achieving near-optimal accuracy levels is nonetheless crucial, as a poor demand prediction model will impact all items alike. An overview of the risk-based approach is outlined in Figure 2 and consists of three main stages: a thorough calculation of overstock and understock cost based on industry expertise, calculation of inventory level probabilities with the aid of AI models, and the combining of those two factors to calculate the costs per possible inventory level option and select of the optimal inventory level according to the minimum aggregate cost. This methodology is applied per each SKU in the business’s inventory database. 

Figure 3: Overview of the risk-based approach to inventory management


Step 1: Detailed calculation of inventory costs 

The risk-based approach to inventory management utilizes over-ordering costs and under-ordering costs in the inventory level decision-making process. Therefore, in order to best leverage this approach, it is crucial that the business has a comprehensive understanding of all possible costs it may face. All major factors pertaining to the industry and company which have potential to impact costs, rent costs, penalties due to client commitments, urgent shipment costs, expiration costs, must be transparent to all parties involved in model implementation.  

Using this knowledge, the business can then calculate the under-ordering costs and over-ordering costs per managed item. For instance, if an aircraft is sidelined due to engine failure and the airline does not have the exact engine model on hand, the airline will have no choice but to make an urgent shipment, often at a cost which may be 10-15% higher than the typical, non-urgent shipment; this is a prime example of an understock cost. Conversely, if the supplier orders too many engines which end up unused, they may be ultimately sold at a loss or written off entirely, in addition to the cost of capital incurred from ordering unnecessary items; both are considered over-ordering costs. 


Figure 4: Common types of costs in inventory management 


For traditional inventory forecasting models which try to maximize accuracy, it does not matter in which direction (over- or under-ordering) the prediction misses the mark. However, common business sense tells us that if one type of cost greatly exceeds the other, our focus should be on the factor which is causing the most harm to the business; a numerical example of this will be shown later in this chapter. For example, if an item is cheap and takes up little inventory space, the cost of over-ordering it would most likely be smaller than the cost of not having the item on hand for a customer; making a conscious decision to overorder the item (which will unequivocally harm model accuracy!) is thus perfectly reasonable from a business perspective. 

Step 2: Calculating demand probability 

Alongside the calculation of inventory costs presented above, the development of an accurate statistical model is necessary as a starting point for this risk-based model. The ability to properly estimate the distribution of demand scenarios, with the combination of an accurate representation of the inventory costs involved, can guarantee an optimal selection of the inventory level. While the factoring of inventory costs will offset the effect of reduced accuracy in a risk-based approach, achieving a reasonable level of accuracy in the calculation of demand probability is nonetheless required for this approach to have a tangible impact in cost reduction.

For an instance of slow-moving inventory, such as an aircraft part ordered sporadically, a trend-based statistical method is ineffective, and might require a more basic, Bayesian probability method to be used instead: the probability of a certain demand scenario to occur is the probability of the occurrences of that demand scenario in the prior specified time period. For instance, if a 12-month period is specified and the item was ordered exactly once in May and October, the model with give a probability of 2/12 that the item will be ordered once in the following month, and a probability of 10/12 that no order will occur. If a longer time period (e.g., three years) is specified, it is highly advisable to use a weighted average which gives recent months a higher impact on the calculated result.

For fast-moving inventory, which is likely to be ordered almost every month, a time-series model is the sensible approach, and many existing inventory management solutions utilize this method, predicting the mean future demand based on the observable past (see Figure 4). In the risk-based approach, however, the mean/median is just one of many quantiles calculated for possible demand scenarios, as the goal is not to anticipate demand with maximum accuracy, but rather to set an inventory level which minimizes the expected value of possible inventory costs. The desired quantiles can be customized based on the needs of the business, but using the 5%, 10%, 25%, 50%, 75%, 90%, and 95% quantiles as a bare minimum is prudent to ensure adequate model performance.

Figure 5: The quantile approach to time-series prediction  

Where 𝑦^ and 𝑝(𝑦ˆ) represent the predicted demand and respective probability out of possible predictions 𝑌ˆ, 𝑂𝐶 and 𝑈𝐶 are the cost of overordering or underordering by a single unit. 

An intuitive explanation of the formula is provided is Figure 5, which calculates the total expected cost per suggested inventory level by aggregating the under-ordering costs (UC) and the over-ordering costs (OC) with the respective probability that the possible demand scenario should come to fruition. For instance, consider a scenario where there is a 50% chance that there will be no demand for an item in the upcoming month, 30% chance that one item will be ordered, and a 20% chance that two items will be ordered. Therefore, for a suggested inventory of 1 item there is a 50% chance that we are overordering one item (as actual demand is zero) and must therefore pay the over-order cost, and in addition there is a 20% chance we must pay the under-ordering cost since two items were ordered.  
Figure 6: Assessment of total expected cost per inventory level suggestion


Of the numerous inventory level options calculated, the model will simply select the one which minimizes the total expected value (i.e., the risk) of inventory costs. As mentioned above, it is possible that the chosen inventory level is not the one which is most likely to occur; for instance, in the above example the most likely possibility is that no items will be ordered in the upcoming month, but if the underordering cost of not having one or two items is exorbitant compared to the underordering cost, the model will not recommend ordering no items, since the total expected cost for the suggested inventory level of zero will be higher than that of a suggestion one or two items. Conversely, if the overordering cost is far greater than that of underordering, the model will suggest ordering no items, even though there is a 50% chance that the customer’s demand will not be met. 

When considering fast-moving inventory the exact same rationale applies, except that the suggested inventory levels are determined by the quantiles previously calculated using a machine learning regression model; the quantile with the lowest total expected cost is that which will be selected. When considering the total overordering and underordering costs, a business can decide to order 95% of maximum possible predicted demand if underordering costs are high or choose a lower quantile if overordering the item will cost the business more. By understanding and applying the costs, the business is thus making a smart, informed decision which is aligned with their deep domain expertise, making this approach easily understandable for all levels of management and staff within the company.

Figure 7: Example of ideal inventory level selection for fast-moving inventory, based on quantiles. 


To summarize, this novel approach to inventory management incorporates overstock costs and under-order costs into the decision-making process alongside machine-learning time-series modeling. By combining statistical methods and the industry’s domain expertise regarding inventory costs, this system predicts the stock level which minimizes expected inventory costs. By systematically choosing this option, this model ensures a responsible strategy in setting a good inventory level for the items with high inventory costs, rather than the common approach which attempts to maximize accuracy across the board. The willingness to sacrifice accuracy in favor of maximal business impact is what defines this risk-based approach to inventory forecasting and sets itself apart from other solutions.

This model is particularly appropriate for the airline case study presented above: although adequate accuracy for slow-moving items is unachievable, the risk-based approach still mitigates inventory costs, which is of paramount importance for any industry where slow-moving inventory is a prime factor. This makes the model desirable and explainable for procurement management and staff – leading to better adoption rates and increasing the likelihood that some of the procurement process can be automated. In addition, the cost-based approach is very flexible, allowing the business to alter the overordering and underordering costs to better reflect both business objectives as well as adapting the actual costs in a dynamic business environment. However, caution must be exercised when customizing the model, and the business must not lose focus of correct business and data practices, including setting rigid metrics and KPIs.


In summary, we have presented an innovative, risk-based approach to inventory management which prioritizes impact, such as loss minimization and service optimization; this represents a departure from traditional models that are focused solely on achieving maximum accuracy, which have limited use in cases of slow-moving demand and may not be aligned with the goals of the business. By bringing business considerations to the forefront, we hope to provide businesses with an approach which can make a difference for even the most complex of warehouses, and which supply chain managers and staff can confidently feel will make the best decisions on their behalf. 

While conceptually this approach is simple to understand, implementation of the above may present challenges which require flexibility and the ability to adapt to the company’s business needs. A comprehensive understanding of the characteristics of the industry, financial and commercial objectives, and operating practices is necessary in order to offer a customized solution which maximizes impact. Being willing to iterate with the business to integrate business rules, to provide the necessary training and support, and to delegate item groups to man or machine is what will make the vision of an effective inventory management system a reality.