Amazon Supply Chain and Logistics

Improve supply planning accuracy with machine learning-based lead time insights

Managing a supply chain in the modern, dynamic marketplace presents numerous challenges because businesses must contend with fast-changing consumer demands, technological advancements, economic volatility, and increasing competition. These factors disrupt supply and demand balance and increase operational complexity. Supply planning, or the process of estimating the product quantities (or raw material or components) needed to meet customer demand is one of the most critical areas of the supply chain. Traditionally, supply planners make these calculations using demand forecasts and supplier lead time data using average or static lead times for supply quantities. Planners then add a percentage of excess inventory to provide a safety stock. This approach excludes lead time fluctuations and logistics delays and introduces a margin of error that forces excess inventory (over stocking) or inventory shortages (under stocking). According to McKinsey & Company, in 2022 retailers continued to overbuy inventory to mitigate shortages which increased unsold inventories by 12 percent to $740 billion. Adopting a supply planning process that uses machine learning (ML) enables the use of data-driven predictive supply planning models that provide a dynamic response to match dynamically changing conditions and trends. Gartner predicts that by 2026 more than 75 percent of commercial supply chain solutions will incorporate advanced analytics (AA), artificial intelligence (AI), and data science or machine learning (ML)-based capabilities to standard processes.

This blog post describes how AWS Supply Chain, a purpose-built cloud-based application, uses ML-based lead time variability detection to improve planning accuracy. This empowers businesses to implement better cost management strategies without sacrificing customer satisfaction. You’ll learn about the differences in using the two supply planning approaches—that is, traditional and ML-based—and how ML capabilities increase accuracy.

How ML-based lead time computations compare against traditional methods

To understand the differences between the two supply planning approaches, consider the following numerical example inspired by a typical consumer goods company scenario. The following graph represents inventory plans using traditional and ML-based supply planning methods. The black line represents monthly demand, which comes from outbound order quantities. The orange line represents the inventory levels that are calculated with the traditional method using static or average lead times. Both the demand and inventory lines fluctuate month-over-month. The demand line fluctuations are expected as seasonal demands can cause significant variances. The orange line fluctuations represent calculation errors that cause overstocking in some months and understocking in others. The reliance on static data contributes to this type of variance, so supply planners manually adjust supply plans by overestimating the required inventory to ensure demand isn’t missed. This approach works but the consequences include additional time spent by planners to adjust the models and increased costs because of excess inventory.

The picture shows demand evolution over time and how ML-based lead times follow it much closer than average lead times as a traditional method.

The blue line represents the monthly inventory level computed using the ML-based method. This approach uses both historical and current transactions (such as open orders and shipments) to determine probabilistic lead time projections based on low (P10), median (P50), and high (P90) confidence intervals. The ML algorithm analyzes and trains the projection model using key product features like day of the week (for delivery timeline), quantities shipped (historic volume), and transportation lane definitions (includes characteristics such as ship from site, ship to site, and Incoterms). These features are data points that can affect order lead times and influence the projections but are neglected in average lead time estimates. The ML algorithm is constantly learning and adjusting the model based on these variables, resulting in the computation of probabilistic lead time projections. While the median value provides an estimate of delivery dates, the low and high values help assess best- and worst-case product delivery scenarios. The preceding graph uses the median value to compute supply quantities.

The differences between the two stocking level computation approaches can also be compared using the demand-supply dispersion measure. Demand-supply dispersion is the difference between projected demand and the planned inventory level and is usually tracked monthly. The graph below calculates the demand-supply dispersion of the current example and shows the dispersion values for each supply planning approach. The x-axis of the graph represents the months of the year from January to December. The y-axis represents the difference between projected demand and planned inventory and lists numbers from negative 15 to positive 25.

In an ideal scenario when supply perfectly matches demand, the dispersion value will be 0. This is denoted with a hashed black line and is the baseline for the comparison. The orange line represents the dispersion value using the traditional planning method and using average lead times. The blue line represents the dispersion value using ML.

The picture shows lower deviations of the ML-based lead times over the baseline vs. traditional average lead time method.

The dispersion of the traditional approach (orange line) is more positively and negatively significant. A positive dispersion value means the supply level is more than the demand and represents an overstock situation. Overstock requires an excess level of working capital and introduces additional risks if the inventory is perishable or seasonal. A negative dispersion value means the supply level is less than the demand and represents a stockout situation. Stockouts can result in lost customers, missed revenues, reduced customer satisfaction, or a combination of these. The ML-based approach (blue line) provides a more accurate supply plan with lower dispersion. ML models are also adaptive, so they become more accurate as more information is collected.

Using ML-based lead time variability detection with AWS Supply Chain

AWS Supply Chain is a purpose-built business application that uses ML-based algorithms to help with lead time variability detection. The application considers historical transactions and open transactions such as open orders and shipments. The ML algorithm also incorporates additional variables such as seasonality, product characteristics, vendor characteristics, and destination-origin sites to train the model. You define the operating parameters and conditions, and the application monitors incoming transactions to determine the projected lead times and estimated delivery dates (with confidence levels). AWS Supply Chain computes any variance to the expected lead times and if the variance exceeds the user defined tolerance range (standard deviation), it recommends a new lead time estimate. This ML-guided lead time variability detection enables supply planners to confidently maintain the proper inventory levels to meet demand.

Tracking lead time deviations insights with watchlists

AWS Supply Chain monitors your supply chain network and learns from transactional data such as orders, shipments, and inventory movements. Users can define specific parameters such as products, locations, or a combination of both along with conditions such as: standard deviations and a historical time period for tracking lead time deviations using watchlists.

You create a watchlist by providing the specific criteria and entities to monitor. In the AWS Supply Chain application, you select Create an Insight Watchlist. This takes you to the following screen where you enter the key parameters you want to track. You first select products, locations, or a combination of the two. You then enter the lead time standard deviation (tolerance range of lead time variance) and time interval to track lead time data. You can share this watchlist with other users by adding their names in the Watchers box.

The illustration shows how watchlist is creates and what parameters (triggers) it needs and how to set up watchers.

The following screenshot shows the Insights dashboard screen that will alert you of products that are violating the criteria you specified. This dashboard is also shared with users you added to the watchlist. The dashboard shows the specific product with a red box to alert of a criteria violation. For example, the dashboard in the screenshot that follows shows the Deluxe Styler having a lead time deviation at the Dallas DC.

The illustration shows lead time deviations for Dallas DC.

You can then choose the deviation box and to bring up the following screen, which provides more details on the deviation. The screen shows the expected lead time based on default information and shows the recommended lead time, which is calculated using ML and is based on historical performance.

The illustration shows lead time deviation for historic purchase orders.

ML models the historical and expected performance and recommends an expected lead time of 19 days instead of the expected or published 5-day target. The application also reports that the vendor has a miss frequency of 100 percent at this location based on historical performance data. Insights also allow you to view the expected performance of a purchase order in progress. The following screenshot shows an open purchase order and lists the supplier, product, and quantity. The screen shows the expected receive date based on the adjusted lead time along with a low-confidence and high-confidence date. These dates take statistical modeling into account based on historical performance to give you a delivery range.

The picture shows the expected receive date based on the adjusted lead time along with a low-confidence and high-confidence date.

This provides a comprehensive view based on actual performance so you can adjust the supply plan.

Collaboration with other team members

Collaboration is a vital function for supply chains. The supply plan change explained in the previous section requires coordination and collaboration with other teams. Traditionally, communication regarding supply plan adjustments occurs through email or phone conversations. This approach works but adds latency to the resolution, and can introduce contextual errors during conversations. AWS Supply Chain has built-in collaboration tools so team members can chat to discuss and resolve the issue without having to leave the application. Users have the same view and the same information on their chat screen so communication and resolution can occur more quickly.


The dynamic nature of the modern marketplace demands an evolution in supply planning strategies. Traditional methods that rely on static data and average lead times can cause calculation errors, overstocking, understocking, and unnecessary expenses. Implementing machine learning into supply planning improves accuracy and reduces Demand-Supply dispersion. ML detects supplier lead time anomalies to improve planning and enact effective risk mitigation actions. Detection and adaptation are also occurring in real-time, ensuring an optimized, efficient, and agile supply chain. Continuous and improved supply chain visibility helps you identify bottlenecks and potential disruptions, while accelerated decision-making helps you to quickly mitigate potential issues without impacting operations.

These improvements increase your supply chain resiliency by allowing you to effectively address customer demands, adapt to market and environmental changes, and reduce operational expenses. To use the-ML based lead time deviation insights, visit AWS Supply Chain to learn more and get started. Also visit the AWS Workshop Studio for a self-paced overview of creating an instance, ingesting data, navigating the user interface, creating insights, mitigating inventory risks, collaborating with other users, and generating demand plans.

Vikram Balasubramanian

Vikram Balasubramanian

Vikram Balasubramanian is a Senior Solutions Architect for Supply Chain. In his role, Vikram works closely with supply chain executives to understand their goals and problem areas and align them with best practices in terms of solution. He has over 17 years of experience working with several Fortune 500 companies across different Industry verticals in the supply chain space. Vikram holds an MS in Industrial Engineering from Purdue University. Vikram is based out of North Dallas area.