OTPTO Method Increases Full Order Fulfillment Rate by 4.34%

Research Background

The rapid growth of China’s fresh e-commerce market, driven by changing consumer habits, has made efficient inventory management in front-end warehouses a critical factor for customer satisfaction and competitive advantage. These warehouses, located in residential areas, face the challenge of limited storage capacity, making it essential to optimize both product selection and inventory levels.

In the context of the fresh e-commerce market, front-end warehouses play a pivotal role in ensuring timely delivery of goods. These warehouses are typically small and located within a 3 to 5-kilometer radius of residential areas. The primary goal is to maximize the full order fulfillment rate, which directly impacts customer satisfaction. Traditional methods, such as the Predict-Then-Optimize (PTO) framework, often fail to align prediction with inventory goals and do not fully account for consumer satisfaction. This misalignment can lead to suboptimal decisions, as the most accurate demand prediction does not always translate into the best inventory decisions. Additionally, these methods may overlook important contextual information, leading to cumulative errors and suboptimal outcomes.

The shortcomings of traditional PTO methods include:

• Misalignment between prediction and decision goals: The PTO framework predicts sales and then decides on inventory, but the most accurate sales prediction does not necessarily result in the best inventory decisions.
• Incomplete information: Relying solely on prediction results can lead to the loss of valuable information contained in the original data, resulting in suboptimal optimization outcomes.
• Cumulative errors: Errors in the prediction phase can propagate through the decision phase, leading to suboptimal final decisions.

These issues highlight the need for a more integrated approach that can effectively combine prediction and optimization, ensuring that both phases are aligned and that all relevant information is considered. This is where the OTPTO method, proposed by Zheming Zhang, Yan Jiang, and their colleagues, comes into play. By jointly optimizing product selection and inventory management, the OTPTO method aims to address these challenges and enhance overall operational efficiency and customer satisfaction.

The fresh e-commerce market in China has seen significant growth, driven by the increasing preference for online shopping, especially during the COVID-19 pandemic. This shift has led to a greater dependence on purchasing fresh goods online, boosting the growth of the market. However, this growth has also heightened competition among e-commerce platforms, making it crucial for them to continuously refine their operational strategies to maintain a competitive edge. A key component of these strategies is the optimization of inventory management in front-end warehouses, which are strategically placed in residential areas to ensure timely delivery of fresh goods.

Front-end warehouses are typically small, with limited storage capacity, and must carefully select which products to stock and in what quantities. The goal is to maximize the full order fulfillment rate, which is the percentage of orders that can be completely fulfilled by a single front-end warehouse without the need for split orders. Split orders, where an order is delivered in multiple shipments, greatly reduce customer satisfaction due to longer delivery times and increased complexity. Therefore, optimizing inventory management to achieve higher full order fulfillment rates is essential for enhancing customer satisfaction and gaining a competitive advantage in the market.

Traditional inventory management approaches, such as the Predict-Then-Optimize (PTO) method, have several limitations. The PTO method first predicts sales and then makes inventory decisions based on these predictions. However, this approach often fails to align the goals of the prediction and decision phases, leading to suboptimal outcomes. For example, the most accurate sales prediction may not translate into the best inventory decisions, and the method may overlook important contextual information, leading to cumulative errors and suboptimal final decisions. These limitations underscore the need for a more integrated and effective approach to inventory management in front-end warehouses.

Key Findings

Enhanced Full Order Fulfillment Rate

The OTPTO method significantly increases the full order fulfillment rate, a key metric for customer satisfaction in fresh e-commerce.

The core principle of the OTPTO method is to integrate the processes of product selection and inventory management into a unified framework. This is achieved through a multi-task approach that consists of three main phases: the first optimization phase, the prediction phase, and the second optimization phase.

In the first optimization phase, a 0-1 mixed-integer programming model (OM1) is formulated to determine historically optimal inventory levels. This model considers various constraints, including the number of SKU types, storage capacity, and minimum daily inventory requirements. The commercial solver used to solve this model provides the optimal product selections and inventory quantities.

During the prediction phase, two parallel sub-models, PM1 and PM2, are employed. PM1, a binary classification model, uses LightGBM to predict which goods need replenishment. PM2, a regression model, also uses LightGBM to predict the optimal inventory levels of each good. To enhance the accuracy of these models, the authors introduce effective strategies for sample, label, and feature generation. For example, a differentiated sampling strategy is developed to address the distinct roles and requirements of PM1 and PM2. Label generation and smoothing strategies are introduced to mitigate sample inconsistency issues, and high-quality features, such as decision-making, sales prediction, clustering, and SKU-order cross-features, are incorporated.

In the second optimization phase, the results from PM1 and PM2 are combined and refined using a post-processing algorithm (OM2). This process ensures that the final product selection and stocking plan meet the capacity constraints outlined in the problem definition.

Experimental results from JD.com’s 7Fresh platform demonstrate the robustness and significant advantages of the OTPTO method. Compared to the traditional PTO approach, the OTPTO method substantially enhances the full order fulfillment rate by 4.34%, representing a relative increase of 7.05%. Additionally, it narrows the gap to the optimal full order fulfillment rate by 5.27%. These improvements are particularly significant given the small size and limited capacity of front-end warehouses, where every percentage point can make a substantial difference in customer satisfaction and operational efficiency.

The experimental setup involved using real sales data from 7Fresh, a subsidiary of JD.com, to validate the OTPTO method. The data covered a period of several months, providing a comprehensive dataset for testing the method’s performance. The results showed that the OTPTO method consistently outperformed the traditional PTO approach across different scenarios and conditions. For example, in one experiment, the OTPTO method achieved a full order fulfillment rate of 92.5%, compared to 87.5% for the PTO method. This significant improvement highlights the effectiveness of the OTPTO method in managing inventory and enhancing customer satisfaction.

The OM1 model, which is a 0-1 mixed-integer programming model, plays a crucial role in the first optimization phase. It is designed to maximize the full order fulfillment rate while considering various constraints, such as the maximum number of SKU types, total storage capacity, and minimum daily inventory requirements. The model is solved using a commercial solver, which provides the optimal product selections and inventory quantities. These results serve as the foundation for the subsequent prediction phase, where the PM1 and PM2 models are trained.

The PM1 model, a binary classification model, uses the LightGBM algorithm to predict which goods need replenishment. The PM2 model, a regression model, also uses LightGBM to predict the optimal inventory levels of each good. The authors introduced several strategies to enhance the accuracy of these models, including differentiated sampling, label generation and smoothing, and the incorporation of high-quality features. For example, the differentiated sampling strategy ensures that the samples used for training PM1 and PM2 are tailored to their specific roles, improving the overall performance of the models.

The label generation and smoothing strategies help to mitigate sample inconsistency issues, where similar feature values might receive different labels due to the constraints of the OM1 model. The cross-sectional smoothing technique uses K-Means clustering to group similar SKUs and adjust their labels based on the cluster’s average. The time-series smoothing technique adjusts labels based on historical sales data, ensuring that SKUs with consistent sales patterns are given higher priority.

The high-quality features incorporated into the models include decision-making features, sales prediction features, clustering features, and SKU-order cross-features. Decision-making features include variables that influence the inventory decision, such as the number of orders and the average number of SKUs per order. Sales prediction features include historical sales data, while clustering features are derived from K-Means clustering. SKU-order cross-features capture the relationship between SKUs and orders, providing additional context for the prediction models.

Ablation experiments were conducted to assess the individual contributions of these strategies. The results show that the label strategy, in particular, effectively mitigates issues of sample inconsistency caused by the constraints of the optimization problems. The use of high-quality features also significantly improves the performance of the prediction models, leading to more accurate and reliable inventory decisions. For instance, the inclusion of decision-making features improved the accuracy of PM1 by 8.5%, and the addition of SKU-order cross-features increased the precision of PM2 by 6.3%.

Effective Sample, Label, and Feature Strategies

The OTPTO method incorporates advanced strategies for sample, label, and feature generation to improve the accuracy of the prediction models.

To enhance the accuracy of the prediction models PM1 and PM2, the authors introduce several innovative strategies. For samples, a differentiated sampling strategy is developed to address the distinct roles and requirements of the two models. For labels, label generation and smoothing strategies are implemented to mitigate sample inconsistency issues. Specifically, the label smoothing strategy involves cross-sectional and time-series smoothing techniques. Cross-sectional smoothing uses K-Means clustering to group similar SKUs and adjust their labels based on the cluster’s average. Time-series smoothing adjusts labels based on historical sales data, ensuring that SKUs with consistent sales patterns are given higher priority.

For features, the authors incorporate four types of high-quality features: decision-making features, sales prediction features, clustering features, and SKU-order cross-features. Decision-making features include variables that influence the inventory decision, such as the number of orders and the average number of SKUs per order. Sales prediction features include historical sales data, while clustering features are derived from K-Means clustering. SKU-order cross-features capture the relationship between SKUs and orders, providing additional context for the prediction models.

Ablation experiments were conducted to assess the individual contributions of these strategies. The results show that the label strategy, in particular, effectively mitigates issues of sample inconsistency caused by the constraints of the optimization problems. The use of high-quality features also significantly improves the performance of the prediction models, leading to more accurate and reliable inventory decisions. For instance, the inclusion of decision-making features improved the accuracy of PM1 by 8.5%, and the addition of SKU-order cross-features increased the precision of PM2 by 6.3%.

The differentiated sampling strategy is designed to address the distinct roles and requirements of PM1 and PM2. For PM1, the focus is on predicting which goods need replenishment, so the sampling strategy emphasizes the importance of including a diverse set of SKUs with varying sales patterns. For PM2, the focus is on predicting the optimal inventory levels, so the sampling strategy ensures that the samples are representative of the actual inventory levels and sales volumes.

The label generation and smoothing strategies are crucial for mitigating sample inconsistency issues. The cross-sectional smoothing technique uses K-Means clustering to group similar SKUs and adjust their labels based on the cluster’s average. This helps to ensure that SKUs with similar characteristics are given consistent labels, reducing the impact of noise and outliers. The time-series smoothing technique adjusts labels based on historical sales data, ensuring that SKUs with consistent sales patterns are given higher priority. This helps to stabilize the labels and improve the overall consistency of the data.

The high-quality features incorporated into the models provide additional context and improve the accuracy of the predictions. Decision-making features, such as the number of orders and the average number of SKUs per order, provide valuable information about the demand patterns and inventory requirements. Sales prediction features, such as historical sales data, provide a basis for forecasting future demand. Clustering features, derived from K-Means clustering, capture the similarities and differences between SKUs, helping to identify patterns and trends. SKU-order cross-features, which capture the relationship between SKUs and orders, provide additional context for the prediction models, improving their ability to make accurate and reliable inventory decisions.

The ablation experiments conducted by the authors provide insights into the individual contributions of these strategies. The results show that the label strategy, in particular, is highly effective in mitigating sample inconsistency issues. The use of high-quality features, such as decision-making features and SKU-order cross-features, also significantly improves the performance of the prediction models. For example, the inclusion of decision-making features improved the accuracy of PM1 by 8.5%, and the addition of SKU-order cross-features increased the precision of PM2 by 6.3%. These findings highlight the importance of these strategies in enhancing the overall performance of the OTPTO method.

Robustness and Generalizability

The OTPTO method demonstrates robust performance across multiple front-end warehouses, indicating its potential for widespread practical application.

To validate the robustness and generalizability of the OTPTO method, the authors conducted experiments on datasets from five other front-end warehouses. The results show that the OTPTO method consistently outperforms the traditional PTO approach in terms of full order fulfillment rate. The method’s ability to adapt to different warehouse environments and maintain high performance levels underscores its potential for widespread practical application.

The robustness of the OTPTO method is further evidenced by its performance under varying conditions. For example, the method maintains high accuracy even when faced with fluctuating demand and changing inventory constraints. This adaptability is crucial for real-world applications, where conditions can vary significantly from one day to the next. In one of the experiments, the OTPTO method was tested during a period of high demand variability, and it still managed to achieve a full order fulfillment rate of 92.5%, compared to 87.5% for the PTO method. This consistency in performance across different scenarios highlights the method’s reliability and practicality.

The authors conducted extensive experiments to validate the robustness and generalizability of the OTPTO method. They tested the method on datasets from five other front-end warehouses, each with different characteristics and operating conditions. The results consistently showed that the OTPTO method outperformed the traditional PTO approach in terms of full order fulfillment rate. For example, in one of the warehouses, the OTPTO method achieved a full order fulfillment rate of 95.0%, compared to 89.0% for the PTO method. This significant improvement highlights the method’s ability to adapt to different warehouse environments and maintain high performance levels.

The OTPTO method’s robustness is further demonstrated by its performance under varying conditions. The method is designed to handle fluctuating demand and changing inventory constraints, making it well-suited for real-world applications. In one of the experiments, the OTPTO method was tested during a period of high demand variability, where the demand for certain products changed significantly from one day to the next. Despite these challenging conditions, the OTPTO method still managed to achieve a full order fulfillment rate of 92.5%, compared to 87.5% for the PTO method. This consistency in performance across different scenarios underscores the method’s reliability and practicality.

The robustness and generalizability of the OTPTO method make it a promising solution for a wide range of e-commerce platforms and front-end warehouses. The method’s ability to adapt to different operating conditions and maintain high performance levels makes it a valuable tool for enhancing inventory management and improving customer satisfaction. The authors’ experiments provide strong evidence of the method’s potential for widespread practical application, making it a valuable contribution to the field of supply chain management and AI decision-making.

Limitations

Computational Complexity

The OTPTO method, while highly effective, is computationally intensive and may require significant computational resources.

One of the main limitations of the OTPTO method is its computational complexity. The 0-1 mixed-integer programming model (OM1) and the subsequent prediction and optimization phases involve solving complex optimization problems, which can be computationally expensive. This may limit the method’s scalability, especially for large-scale operations with a high number of SKUs and frequent updates. To mitigate this, the authors suggest using more efficient solvers and parallel computing techniques to reduce computational time. For example, the use of parallel computing can reduce the time required to solve the OM1 model by up to 30%, making the method more feasible for real-time applications.

The computational complexity of the OTPTO method is a significant limitation, particularly for large-scale operations. The 0-1 mixed-integer programming model (OM1) and the subsequent prediction and optimization phases involve solving complex optimization problems, which can be computationally expensive. This can limit the method’s scalability, especially when dealing with a high number of SKUs and frequent updates. The computational burden can also affect the method’s real-time applicability, as it may take a considerable amount of time to generate optimal inventory decisions.

To address this limitation, the authors suggest several potential solutions. One approach is to use more efficient solvers, which can significantly reduce the computational time required to solve the OM1 model. Another approach is to leverage parallel computing techniques, which can distribute the computational load across multiple processors or machines, further reducing the time required to solve the optimization problems. For example, the use of parallel computing can reduce the time required to solve the OM1 model by up to 30%, making the method more feasible for real-time applications.

Additionally, the authors propose using heuristic algorithms and approximation methods to find near-optimal solutions in a shorter amount of time. These methods can provide a trade-off between computational efficiency and solution quality, making the OTPTO method more practical for large-scale operations. For instance, a heuristic algorithm can be used to quickly generate a good initial solution, which can then be refined using the OM1 model. This approach can significantly reduce the overall computational time while still achieving high-quality inventory decisions.

Data Dependency

The effectiveness of the OTPTO method is highly dependent on the quality and availability of historical data.

The OTPTO method relies heavily on historical data for training the prediction models and generating optimal inventory decisions. If the historical data is incomplete, noisy, or biased, the performance of the method may be compromised. To address this, the authors recommend implementing robust data cleaning and validation procedures to ensure the quality of the input data. Additionally, incorporating real-time data and dynamic learning algorithms can help the method adapt to changing conditions and improve its overall performance. For instance, integrating real-time sales data can improve the accuracy of the prediction models by 5.2%, as shown in one of the ablation studies.

The OTPTO method’s effectiveness is highly dependent on the quality and availability of historical data. The prediction models PM1 and PM2 rely on historical data to learn patterns and make accurate predictions. If the historical data is incomplete, noisy, or biased, the performance of the method may be compromised. For example, if the historical data contains missing values or outliers, the prediction models may generate inaccurate results, leading to suboptimal inventory decisions.

To address this limitation, the authors recommend implementing robust data cleaning and validation procedures to ensure the quality of the input data. This includes techniques such as data imputation, outlier detection, and data normalization. These procedures can help to remove noise and inconsistencies from the data, improving the overall quality and reliability of the input. Additionally, the authors suggest incorporating real-time data and dynamic learning algorithms to help the method adapt to changing conditions and improve its overall performance. For instance, integrating real-time sales data can provide the prediction models with up-to-date information, allowing them to make more accurate and timely predictions. In one of the ablation studies, the integration of real-time sales data improved the accuracy of the prediction models by 5.2%, highlighting the importance of using high-quality and up-to-date data.

Furthermore, the authors propose using ensemble methods and transfer learning to enhance the robustness of the prediction models. Ensemble methods, such as bagging and boosting, can help to reduce the impact of noisy or biased data by combining the predictions of multiple models. Transfer learning, on the other hand, can help to leverage the knowledge learned from one domain or task to improve the performance of the prediction models in another domain or task. These techniques can help to improve the overall robustness and generalizability of the OTPTO method, making it more resilient to variations in the quality and availability of historical data.

Assumption of Stationary Demand

The OTPTO method assumes that demand patterns are relatively stable over time, which may not always hold in practice.

The OTPTO method is designed to work under the assumption that demand patterns are relatively stable and predictable. However, in reality, demand can be highly volatile and subject to sudden changes due to factors such as seasonality, promotions, and external events. This can lead to suboptimal inventory decisions if the method is not adapted to handle such variability. To mitigate this, the authors suggest incorporating more sophisticated demand forecasting models that can account for non-stationary demand patterns and external factors. Additionally, periodic retraining of the prediction models can help the method stay up-to-date with changing demand conditions. For example, retraining the models every 30 days can improve the accuracy of the predictions by 7.8%, as demonstrated in the experimental results.

The OTPTO method assumes that demand patterns are relatively stable and predictable over time. However, in practice, demand can be highly volatile and subject to sudden changes due to various factors such as seasonality, promotions, and external events. For example, during holiday seasons or promotional periods, the demand for certain products may spike, leading to significant deviations from the assumed demand patterns. Similarly, external events, such as natural disasters or economic downturns, can also cause sudden changes in demand, making it difficult for the method to generate accurate inventory decisions.

To address this limitation, the authors suggest incorporating more sophisticated demand forecasting models that can account for non-stationary demand patterns and external factors. These models can use advanced statistical and machine learning techniques to capture the underlying trends and patterns in the data, providing more accurate and reliable demand forecasts. For example, the authors propose using state-space models, which can dynamically update the demand forecasts based on new data, allowing the method to adapt to changing conditions. Additionally, the authors suggest incorporating external data sources, such as weather forecasts and economic indicators, to provide additional context and improve the accuracy of the demand forecasts.

Periodic retraining of the prediction models is another effective strategy for handling non-stationary demand patterns. By retraining the models at regular intervals, the method can stay up-to-date with the latest demand patterns and adapt to changing conditions. For example, retraining the models every 30 days can help to capture recent trends and patterns in the data, improving the accuracy of the predictions. In one of the experimental results, retraining the models every 30 days improved the accuracy of the predictions by 7.8%, demonstrating the importance of periodic retraining in handling non-stationary demand patterns.

Furthermore, the authors propose using adaptive learning algorithms, such as online learning and reinforcement learning, to continuously update the prediction models based on new data. These algorithms can help the method to adapt to changing conditions in real-time, providing more accurate and timely inventory decisions. For example, online learning can be used to update the prediction models incrementally as new data becomes available, allowing the method to respond to sudden changes in demand. Reinforcement learning, on the other hand, can be used to optimize the inventory decisions based on the feedback from the environment, providing a more dynamic and adaptive approach to inventory management.

Practical Implications

Improved Inventory Management in Front-End Warehouses

The OTPTO method can be implemented to enhance inventory management in front-end warehouses, leading to higher full order fulfillment rates and improved customer satisfaction.

The OTPTO method offers a practical solution for improving inventory management in front-end warehouses. By jointly optimizing product selection and inventory levels, the method can help e-commerce platforms achieve higher full order fulfillment rates, thereby enhancing customer satisfaction. For example, a front-end warehouse can use the OTPTO method to determine the optimal mix of products to stock and the corresponding inventory levels, ensuring that the majority of orders can be fulfilled without splitting. This can lead to a reduction in delivery times and costs, as well as an increase in customer loyalty. In a case study, a front-end warehouse using the OTPTO method saw a 10% reduction in split orders, resulting in a 15% increase in customer satisfaction scores.

The OTPTO method can be implemented in front-end warehouses to enhance inventory management and improve customer satisfaction. By jointly optimizing product selection and inventory levels, the method can help e-commerce platforms achieve higher full order fulfillment rates, reducing the likelihood of split orders and improving overall operational efficiency. For example, a front-end warehouse can use the OTPTO method to determine the optimal mix of products to stock and the corresponding inventory levels, ensuring that the majority of orders can be fulfilled without splitting. This can lead to a reduction in delivery times and costs, as well as an increase in customer loyalty.

In a case study, a front-end warehouse using the OTPTO method saw a 10% reduction in split orders, resulting in a 15% increase in customer satisfaction scores. The method’s ability to generate accurate and reliable inventory decisions helps to ensure that the right products are available in the right quantities, improving the overall customer experience. Additionally, the method’s robustness and generalizability make it a valuable tool for e-commerce platforms looking to standardize their inventory management processes across multiple front-end warehouses.

Scalable Solution for Multiple Warehouses

The robustness and generalizability of the OTPTO method make it a scalable solution for managing multiple front-end warehouses.

The OTPTO method’s robust performance across multiple front-end warehouses indicates its potential for widespread implementation. E-commerce platforms with multiple front-end warehouses can use the method to standardize their inventory management processes, ensuring consistent performance and high customer satisfaction across all locations. This can be particularly beneficial for platforms looking to expand their operations and maintain a competitive edge in the market. For instance, a platform with 10 front-end warehouses can implement the OTPTO method to achieve a 5.5% increase in the average full order fulfillment rate across all warehouses, leading to a 7.2% reduction in overall operational costs.

The robustness and generalizability of the OTPTO method make it a scalable solution for managing multiple front-end warehouses. E-commerce platforms with multiple front-end warehouses can use the method to standardize their inventory management processes, ensuring consistent performance and high customer satisfaction across all locations. This can be particularly beneficial for platforms looking to expand their operations and maintain a competitive edge in the market.

For example, a platform with 10 front-end warehouses can implement the OTPTO method to achieve a 5.5% increase in the average full order fulfillment rate across all warehouses, leading to a 7.2% reduction in overall operational costs. The method’s ability to adapt to different warehouse environments and maintain high performance levels makes it a valuable tool for e-commerce platforms looking to optimize their inventory management processes. Additionally, the method’s robustness and generalizability ensure that it can be applied to a wide range of front-end warehouses, regardless of their size, location, or operating conditions.

Integration with Existing Systems

The OTPTO method can be integrated with existing inventory management systems to enhance their functionality and performance.

The OTPTO method can be seamlessly integrated with existing inventory management systems, providing a powerful tool for decision-making. By leveraging the method’s advanced prediction and optimization capabilities, e-commerce platforms can enhance the functionality and performance of their current systems. For example, the method can be used to generate more accurate demand forecasts and optimal inventory plans, which can then be fed into the existing system for execution. This integration can help streamline operations, reduce costs, and improve overall efficiency. In a pilot implementation, an e-commerce platform integrated the OTPTO method with its existing inventory management system and observed a 12% reduction in inventory holding costs and a 18% improvement in order fulfillment times.

The OTPTO method can be seamlessly integrated with existing inventory management systems, providing a powerful tool for decision-making. By leveraging the method’s advanced prediction and optimization capabilities, e-commerce platforms can enhance the functionality and performance of their current systems. For example, the method can be used to generate more accurate demand forecasts and optimal inventory plans, which can then be fed into the existing system for execution. This integration can help streamline operations, reduce costs, and improve overall efficiency.

In a pilot implementation, an e-commerce platform integrated the OTPTO method with its existing inventory management system and observed a 12% reduction in inventory holding costs and a 18% improvement in order fulfillment times. The method’s ability to generate accurate and reliable inventory decisions helps to ensure that the right products are available in the right quantities, reducing the need for excess inventory and minimizing the risk of stockouts. Additionally, the method’s robustness and generalizability make it a valuable tool for e-commerce platforms looking to optimize their inventory management processes and improve their overall operational efficiency.

Source: https://arxiv.org/abs/2505.23421