Introduction
The recent war in Ukraine has underscored the importance of logistical planning. In a 2023 study, RAND noted that “Russia’s failures in the war with Ukraine were due to poor planning”.1 Seeing these missteps and recognizing the critical importance of logistics, US Army Pacific named it as one of its four core “Interior Lines” that must be met to achieve its 2025 posture plan.2 The Defense Logistics Agency (DLA), the Department of Defense’s (DOD) combat support supply chain organization, will be a key piece in achieving these plans. Because of this importance, DLA’s way of material planning needs to evolve. DLA supports the DOD with most of its consumable items, everything from clothing to fuel to medical supplies to weapon system repair parts. While historical-based planning to predict future DOD requirements has served DLA well, it may not be fully equipped to handle the complexities and uncertainties of the modern battlefield. Current models rely solely on historical demand, limiting DLA’s ability to account for the wide range of factors that now influence military supply chains. This limitation has become even more pronounced as enhanced data sharing from DLA’s customers has opened access to richer, timelier, and more diverse datasets. Machine learning-based planning methods are uniquely positioned to leverage this expanded data environment by incorporating multiple variables, identifying nonlinear relationships, and adapting to changing patterns in real time. Transitioning to ML-based material planning that is informed by new data represents a necessary evolution for the organization to improve accuracy, resilience, and strategic responsiveness.
DLA manages a significant portion of the DoD’s supply chain: it plans for future customer requirements, acquires inventory to meet that demand, stores the inventory, and fulfills customer orders from that inventory. Like private industry, DLA’s demand planning seeks to predict what customers will order in the future. Based on that input, DLA’s supply planning strives to ensure the right inventory is available at the right location when that demand materializes. DLA HQ Planning controls policy and maintains oversight of these planning functions and then works with DLA Acquisition to ensure the supply plan is met. Similar to private industry, when these core functions are out of sync cash flows are interrupted, orders go unfilled, and objectives are not accomplished. However, unlike the commercial sector DLA does not seek to maximize revenue; it is tasked with maximizing warfighter readiness. Because of this, the ramifications of a misaligned DLA supply chain are much greater than private sector organizations: DLA impacts the DOD’s ability to defend the homeland and defeat its adversaries.
DLA is also in an incredibly difficult business because 80% of its five million items average zero demands a year, 70% of its stocked items have vendor lead times over six months, and the number of small businesses in the defense-industrial base has declined by 40% in the last decade.3 Because of these challenges, DLA relies on five statistical methodologies that are univariate, meaning they are solely influenced by one variable (historic demand, historic vendor lead times, etc.). Across these five methods, DLA routinely achieves a steady balance of 60% of items that are adequately planned, 20% over-planned (too much inventory and/or not enough demand), and 20% under-planned (too little inventory and/or too much demand). DLA also receives formal collaborative demand forecasts from the military services that are used to influence DLA’s final demand plan. DLA’s Collaborative Horizon Metric (CHM) assesses customers’ collaborative forecasts and reveals that customer input only adds value to 25% of DLA’s statistical forecasts. This poor customer forecasting drives the same outcomes as poor DLA planning: increased inventory that ties up funding and decreased customer support. With the smallest American industrial base in recent history, DLA must increase the quality of its planning as it no longer has the luxury of having suppliers that can produce items with a moment’s notice. These are warning signs that keep the door open to the DOD’s supply chain being a limiting factor in a potential future conflict similar to what the Russians initially experienced in Ukraine. It is time to transform DLA’s material planning with machine learning algorithms operating on new and more data to meet the intent of the new DLA Strategic Plan.4
Objective
To meet the Director’s intent, DLA HQ Planning is undertaking the five-step process outlined in Nicolas Vandeput's Demand Forecasting Best Practices. Vandeput, a leading expert in supply chain optimization and demand forecasting, provides a practical framework for improving forecasting accuracy and aligning it with business objectives.5 The objective of DLA’s planning system is clear: DLA must enable the American warfighter to successfully defend the homeland and deter its adversaries.
Data
Data will inevitably be the most important factor in moving DLA’s planning process forward. Luckily, there have been massive gains in data processing, storage, and willingness to share data up and down the DOD supply chain. However, a significant challenge remains: the DOD's data landscape is often characterized by siloed systems and inconsistent data standards, hindering effective collaboration and optimization. The most important data DLA needs is downstream customer demand data that is currently not visible to DLA. Having that data will allow DLA to match downstream customer demand with supply, which will ultimately improve DLA’s demand plan. Secondly, DLA needs customer on hand inventory and inventory policy data. This will allow for global optimization of the DOD supply chain inventory, rather than its current siloed approach. This concept is called Multi-Echelon Inventory Optimization (MEIO) and has reduced inventory by as much as 35% in some business use cases while still maintaining high service levels.6 In DLA’s case, a 35% reduction in inventory would amount to almost $6B.7 In addition to order and inventory data, commercial data from DLA’s vendors, exogenous data like weather patterns and economic indicators, and internal DOD data like maintenance schedules, unit employment schedules, and other internal logistics data unique to each military service should all be curated and fed into DLA’s planning algorithms. Finally, all data must be fully integrated into DLA’s Enterprise Advanced Analytics’ (EA2) data warehouse to ensure the planning system and analysts can make the most of it. There can no longer be spreadsheet swapping or intermittent pushes of data. In addition, data quality teams must address the challenges of inconsistent data formats, disparate systems, and security concerns. A robust and secure data infrastructure is essential for enabling effective ML-driven planning. Overcoming the challenges of breaking down data silos, using that new data, and ensuring high levels of data quality are foundational steps in transforming DLA's material planning process.
Metrics
DLA can benefit from a more balanced approach to Planning metrics, giving greater attention to forecast bias and inventory-related measures. Bias is important because it tells if a forecast is systematically wrong in a particular direction. Unchecked, it can compound stockouts, lead to buying dormant inventory, and result in a loss of trust in the forecasting process. Figure 2 shows that ideal forecasts are accurate AND unbiased. DLA’s naïve forecast (replaying historical demand) has an accuracy of 50%, which beats DLA’s published forecast by three points.8 DLA has a 15% over forecast bias, which sounds like a much better alternative than having a bias toward under forecasting. But digging deeper, being biased toward over forecasting results in tying up money on inventory of which DLA already has too much.10 Since DLA is in such a variable demand environment, it is astonishing that DLA gets the results it does, but there is still room for improvement to get DLA’s demand plan centered and concentrated around the bullseye. Finally, there is not enough exposure to inventory metrics. Looking at inventory is important because it shows where DLA is putting its money. If that money is misplaced, it leads to poor customer results and bad DLA financials.
Baseline Model
DLA can also move away from its current univariate, statistical models and toward ML based planning because of the variety of factors that impact DLA planning. Factors like global operations, seasonality, maintenance schedules, and unknowable patterns to the human eye impact customer demand to which DLA is currently blind. Statistical models can only handle these factors one at a time, while ML models are trained on entire datasets. DLA should take advantage of all data available to it via a model that can consider all relevant variables. To get this ML-enabled model, HQ Planning should encourage open competition to see if internal and external DLA stakeholders can create ML models that improve DLA’s supply and demand planning. To this end, DLA should hold its own Makridakis Competition like Kaggle’s 2020 M5 Forecasting-Accuracy challenge, during which users predicted daily Walmart sales for the next 28 days.9 Five teams beat the benchmark by 20% and the winner won $25,000. Instead of money, the DLA competition’s prize could be a winner take all event where the winning model is adopted for use by DLA.
Review Processes
It is a good rule of thumb to assume that the further up the DOD food chain you get, the more influence budgets have on the planning process. It is also safe to assume the closer to the warfighter you get, the more the planning process becomes readiness focused. Competing priorities are natural in any organization, and the combination of these biases should culminate in an improved plan in the aggregate. Acknowledging this fact will allow all planning stakeholders to speak frankly and honestly with each other. To that end, DLA should work with OSD to re-examine its processes and procedures, ensuring that both efficiency and warfighter readiness are effectively balanced. For example, the Weapon System Priority Program (WSSP) that identifies mission essential critical parts must be updated to ensure DLA’s ML models maximize critical platform readiness to better align with warfighter readiness needs.
Ultimately the end state for DLA’s Planning system and community is to become more analytically adept and engaged. Planners will increasingly leverage data-driven insights and advanced analytics tools, augmenting their existing expertise and judgment. The future is centered around fine-tuning models, collecting better data, generating a better understanding of support strategies and customer requirements, and being more agile in response to shifting customer demands. HQ Planning is currently working on several projects to move DLA’s planning ability to this desired end state, which, on their own merits, serve as examples of how ML and advanced analytics capabilities can improve logistics planning.
What We Have Learned
Reducing the Bullwhip Effect Via Enhanced Army Data Sharing
Nicolas Vandeput in Demand Forecasting Best Practices defines the bullwhip effect as “a situation where the more upstream an echelon is in the supply chain, the more demand variation it faces, even if the final demand (from end customers) is steady.”10 As the most upstream organization in the DOD supply chain, DLA struggles mightily with this phenomenon. Demand variation has DLA playing Whac-a-Mole, trying to distinguish if large swings in demand are trends or one offs.11 Vandeput asserts data sharing is the only way the bullwhip effect can be counteracted. Being able to see upstream in the supply chain allows DLA to see when demand will be coming its way.
Imagine a scenario in which DLA can see that ten widgets are ordered internally in the Army every week instead of one large order of one hundred and thirty every quarter. DLA would know this widget demand is not randomly happening four times a year but rather is steadily demanded. This new insight allows DLA’s vendors to maintain steady production schedules with steady cash flows, and it also allows DLA to carry less inventory (only having to carry a few weeks of safety stock vice multiple quarters).
To this end, the Army is actively sharing internal consumption data for Bradley Fighting Vehicle items, allowing DLA to see internal Army consumption. The hypothesis was that this internal consumption data allows DLA to smooth out the demand, which would lead to a dampened DOD bullwhip. In trial simulations using this internal Army consumption data, results showed an 8% increase in forecast accuracy for fast moving items. But it also showed that almost 80% of items receive no forecast accuracy improvement. For the items that seemed to benefit from this new insight, their actual demand plan was modified to incorporate this enhanced insight in the July 2025 demand plan. The team will monitor the production results, learn from them, and incorporate the lessons learned into a larger effort that includes all Army data.
Using Machine Learning to Demand Plan
DLA HQ Planning is reviving a 2019 R&D ML based demand forecast effort that had promising results. Simulated over-forecast dollar error decreased by a simulated $102M for the sample fifty thousand items and average demand plan accuracy was increased by 3.5%. The approach used an ensemble method where multiple forecasting algorithms were used together to produce a single demand plan. A neural network was trained using item level characteristics, which allowed the business to adjust what characteristics were heavily weighted depending on business need. In theory, this can allow DLA to plan by readiness driver or weapon system platform instead of its aggregated demand profile. After six years, DLA has seen massive gains in areas that held the 2019 project back. Specifically, DLA J6 rolled out a dedicated ML environment, Cloudera Data Science Workbench, for analysts to create their own ML models.12 The workforce is also much more analytically advanced, and the data warehousing and compute power has drastically expanded. This project began in July 2025 and hopes to radically transform DLA’s planning algorithms.
Estimating Vendor Lead Times with Random Forests
Estimating when vendors will deliver material is critical to supply chains because it allows businesses to estimate how much inventory will be needed. This predictability reduces costs, minimizes delays, and improves service to customers.
In 2024, the team started a project that used a random forest machine learning model to see if it could better predict vendor lead times. This model analyzed ten years of procurement history and was able to more accurately predict when vendor deliveries would be delivered. It is estimated the model will reduce buyer workload by 5% in its first year and will reduce lead time estimation error by 32 days. Because of these promising results, DLA HQ Planning changed 165k vendor lead times according to the recommendations from the model in March 2025. As this was the first ML project to drive business decision making, it revealed many things. The first was that it took buy-in from key stakeholders to put these changes into production. Close collaboration with stakeholders ensured the successful transition of insights into production. The second major finding was how important communication was throughout the project’s life cycle. Communicating the strategic intent to the modelers, the modelers communicating the fundamentals of the model to the Planning team, and the Planning team communicating to senior leaders and other MSC stakeholders about important process differences all needed to take place before changes could be made to the lead times in the system. Finally, because AI and ML are not easily reverse engineered, the team found that buy-in is a hard, slow road when it comes to early adoption of new technology. Not every model recommendation was adopted: DLA HQ Planning did not change every lead time as the model said because there was not total buy-in. Exceptions were made and countless late night phone calls were fielded to get enough buy in so that everyone could sleep at night.
Posturing for the Future
One of the biggest revelations that has come out of the initial projects is how people are the most critical piece of advancing planning’s analytic capabilities. The team has encountered disbelievers, evangelists, early adopters, naysayers, and cultures of tradition that are difficult to break. Without the right people willing to share data or come to the table with an open mind, many efforts would have fallen flat. The team learned that putting people in positions to succeed analytically is the only way to garner effective outcomes across the DLA supply chain. This paradigm saw glimmers of enhanced agility and responsiveness. With real time data monitoring and predictive analytics, the team could quickly identify and respond to shifts in demand, supply disruptions, or logistics bottlenecks which reduced customer wait time and stock outs. Lastly, since the team was structured around analytics it was able to simulate changes in environmental conditions, supplier reliability, and customer behavior. Putting people in positions to succeed analytically is one of the key takeaways from our experience with these ML projects.
What Success Looks Like
As stated in DLA’s Strategic Plan 2025-2030, DLA is in a transformative era. Planning’s central role in shaping supply chain outcomes sets it up as a key catalyst to enable that transformation in support of the warfighter. To get there, DLA HQ Planning is pushing the Agency into a more data-centric material planning ecosystem. DLA’s material planning is becoming more integrated with its customers both by giving and receiving data. Planning is also modernizing its technology with a push toward ML based demand and supply planning over traditional univariate ones. And finally, DLA HQ Planning is creating a culture of analytical excellence internally that will encourage partnership with other like-minded organizations across the Joint Logistics Enterprise. Come join us.
1 Bradley Martin, D. Sean Barnett, and Devin McCarthy, Russian Logistics and Sustainment Failures in the Ukraine Conflict: Status as of January 1, 2023 (Santa Monica, CA: RAND Corporation, 2023), https://www.rand.org/pubs/research_reports/RRA2033-1.html
2 United States Army Pacific. Theater Army Strategy: Get in Position to Compete, Fight, and Win. Interim edition. Fort Shafter, HI: USARPAC, April 2025.
3 U.S. Department of Defense, Small Business Strategy (Washington, DC: Department of Defense, January 2023), https://www.defense.gov/News/Releases/Release/Article/3279279/dod-releases-small-business-strategy/
4 Mark Simerly. DLA Transforms: A Call to Action. https://www.dla.mil/Info/Strategic-Plan/
5 Nicolas Vandeput. Demand Forecasting Best Practices. 2023, page 5.
6 Nicolas Vandeput. Inventory Optimization, Models and Simulations. 2020, page 182.
7 DLA, Tim Morefield, “FY2024 End-of-Year Materiel Management Review”. March 14, 2025 slide 5.
8 DLA, Tim Morefield, “FY2024 End-of-Year Materiel Management Review”. March 14, 2025 slide 10.
9 Spyros Makridakis, M5 Forecasting = Accuracy (Kaggle, 2020), accessed July 2, 2025, https://www.kaggle.com/competitions/m5-forecasting-accuracy
10 Nicolas Vandeput. Demand Forecasting Best Practices. 2023 page 61.
11 https://en.wikipedia.org/wiki/Whac-A-Mole
12 https://www.cloudera.com/products/machine-learning.html