Construction practice of AI driven Jingdong end-to-end replenishment Technology


Reading Guide:Automation is embedded in the gene of the whole intelligent supply chain y. one vision of our service is to reduce the cost and improve the efficiency of the whole supply chain through automation technology. This article will share how uses AI to drive end-to-end replenishment construction, including the following aspects:

  • Current situation of JD intelligent supply chain
  • End to end replenishment construction driven by AI and or
  • Future outlook


Current situation of JD intelligent supply chain

The whole supply chain business of JD group is divided into pop and self support. More than 80% of the work of Y Department of intelligent supply chain is to serve the self support supply chain. Self operated supply chain refers to self procurement and self sales. This part of work is mainly undertaken by Y’s technical team, and the whole business volume is very large.

From the recent financial report, the active users of JD platform are about 552 million, and there are more than 9 million kinds of self operated goods, covering all categories such as home appliances, computer digital, fast sales and so on. More than 200000 + suppliers have been connected to purchase different goods. The scale of the whole network is very complex, and all warehouses of logistics are basically used. In this way, the user experience can be better improved and the cost can be reduced. Through a series of supply chain optimization actions, this year’s Q3 financial report shows that the turnover period has reached 30.1 days. In terms of benchmarking industry, it is at a very leading level, with more than 93% of self operated orders reaching the target within 24 hours. On the one hand, it is because of the operation of business personnel, on the other hand, it is the blessing of technical services.


End to end replenishment construction driven by AI and or

1. Layout five capabilities

There is a very long process for a commodity to be purchased by consumers. From the initial creative design, R & D and manufacturing to the final retail manufacturer’s pricing, trading, warehousing, logistics distribution and after-sales, the whole chain is very long. How can such a long chain improve efficiency? We need to optimize all links of the whole chain. Our y Department has laid out five capabilities in the optimization process:

① C2m capability

C2m capability is a core of serving suppliers and brands. Through industry insight algorithm and based on sales data, it can identify which commodities are more favored by consumers in the future, and empower the whole brand with the characteristics, characteristics and some key attributes of commodities. Brand makers make production and design of new products based on insight and suggestions. In this way, the production efficiency of goods is very high, and they can sell faster and better.

② Category planning

Category planning in the industry covers a wide range, mainly category planning strategies. Y mainly focuses on commodity structure and multi-end selection.

What is commodity structure? For example, Jingdong has more than 9 million goods. It is difficult for consumers to find out which goods are repeated. These products should reduce some self operated, let consumers choose, and expose some better products to consumers. In addition, it also depends on which commodities are not available in self operated commodities, and whether these commodities can be introduced into JD platform to make consumers better choose.

In addition, for multi-end and multi-channel selection, is a large platform with different fields, channels and different goods of different quality. The sales efficiency of different channels is completely different. We hope to identify different commodities through technology and push them to different channels to accelerate sales and circulation.

③ Automatic pricing

A core of Y is to improve efficiency through automatic means. We hope to realize the automatic pricing of the whole JD platform in the future, provide consumers with more reasonable prices and maintain relatively low profits.

④ Intelligent inventory

Inventory capacity is a relatively mature capability of Y, which is arranged in multiple links from forecasting, replenishment, transfer to clearing the backlog.

Warehouse network performance

In the whole retail supply chain, we hope to optimize the whole cost and efficiency through the regulation of the network and the optimization of performance.

2. Carry out three types of work

In order to optimize the whole chain, most of our work can be divided into three categories:

① The first type is forecasting work, which is the basis of all kinds of decision-making and planning.

In the whole self operated supply chain, the core is how to predict the future sales. Because we know the future sales volume, we can decide what kind of time to purchase. We are very mature in terms of sales forecast. The purpose of flow forecast and financial forecast is to better plan the sales volume and make things more extreme.

② The second is the ability of layout optimization.

Our ability of layout optimization is very sufficient. Because the optimization ability can help us make better decisions. For example, in the replenishment scenario, when and how much to buy is a very clear operational research optimization problem.

If we want to optimize the whole turnover efficiency, we need to make decisions on the best time and quantity, which depends on the construction of operational research capability. The establishment of operational research capability will serve the performance, including order distribution optimization, performance route optimization, and additional output.

③ The third type is simulation platform.

For any algorithm, a round of detailed evaluation shall be conducted before landing and online. The simulation platform is the carrier for us to complete the last round of preparation before going online. It is very difficult to optimize the supply chain inventory, which is affected by many interference factors and foreground factors. How to evaluate this impact? We have established a supply chain simulation platform to simulate the influence of various factors through the simulation platform, which can make our scheme more robust. The simulation platform won the Innovation Award in retail.

3. Inventory management

Inventory is the most important and difficult part of the whole supply chain management. If you want to manage the inventory well, the process involved is very long, including prediction. If you predict more accurately, the plan will be more reasonable. Based on the plan and rhythm, we can decide how much to purchase. After mining, enter Jingdong warehouse, how to allocate to the warehouse closer to consumers, and improve the timeliness level. In the process of allocation and procurement, there are still some inventories that are not easy to sell, that is, unsalable inventory. How to identify unsalable inventory, give a reasonable price, accelerate digestion, and optimize the whole turnover efficiency.

① Core indicators of inventory management

Inventory management serves three major objectives within retail:

  • The goal of turnover, the faster the inventory turnover, means to release better cash flow at lower cost, which is very key for enterprises;
  • Spot rate: the higher the spot rate, the higher the sales opportunity, and the better to maximize the whole Gmv;
  • Satisfaction rate refers to the satisfaction rate of the front warehouse. If more goods are sent from the front end, the user’s timeliness experience will be better. For example, more than 93% of our goods can be delivered in 24 hours, which depends on the rationality of allocation to make the goods closer to consumers and reach consumers at a faster speed.

② Intelligent replenishment

Intelligent replenishment is one of the cases. Y serves the business in an automated way. The first challenge is sales forecasting. JD has more than 9 million self operated commodities and 1300 warehouses in eight RDCs. The more than 9 million commodities multiplied by the eight RDCs are nearly 100 million, which needs to be predicted. The scale is very large. After the precipitation of the whole forecast is very mature, in order to ensure the universality of the forecast, the current forecast will output the forecasts of the next 91 days by day, by SKU and by RDC to serve the downstream intelligent replenishment system.

Replenishment is more about purchasing the whole commodity from the supplier to JD warehouse. About tens of millions of purchase suggestions are output every day, and these goods are automatically ordered and warehoused from suppliers to realize the whole automatic process. When the goods enter JD warehouse, they should be quickly allocated to FDC to complete the transfer within JD. At present, the automation rate of transfer is as high as 90%, and the level is very high.

③ Intelligent clearing stagnation

When the goods are distributed in JD, one task is how to accurately identify the unsalable inventory. We have purchased a lot of goods, but not all goods are easy to sell. If the sales of some goods are less than expected after purchase, they will become unsalable, which is called unsalable inventory. We need to identify quickly, match the reasonable price and sell it, so that the turnover efficiency can be minimized. JD can release more cash flow and make the whole finance more competitive.

At present, the whole inventory management is a fully automated process, which looks simple, but it is difficult to implement. JD’s self operated business is highly complex, with a full range of commodities, different categories have different models, and different networks need to be matched. The network often changes. When it changes, the demand for each warehouse coverage will also change greatly. This is a great challenge to forecasting. For example, when the demand of a large warehouse fluctuates greatly, it will become very different from the previous characteristics, and there will be more challenges to forecasting and replenishment.

For example, small and medium-sized goods are two-tier networks, and large goods are three-tier networks. After the network changes from two layers to three layers, the difficulty of replenishment will also increase. In the two-tier network, we mostly prepare goods for RDC, but when building the three-tier network, the three-tier network can purchase directly and transfer between networks. In this case, the network coverage relationship changes more violently. It is necessary to use technology to identify its changes, match the best and reasonable prediction method in real time, and give the best and reasonable replenishment logic.

4. Forecast

Our prediction is currently supported by the intelligent prediction platform. Its bottom layer has two core capabilities:

  • The first is the ability of big data;

  • The second is mature machine learning and deep learning technology.

In order to make the intelligent prediction platform more adaptable, we will bring some business information to the intelligent prediction platform, such as different business formats, such as the plan driven business format of power grid, the futures business format of fashion, and the demand driven fast sales of small and medium-sized parts. We need to take different business formats into account in the prediction process.

In addition, we will establish different impact factors, which have a great impact on the final effect, such as promotion factors and seasonal factors. We model various factors separately to support a variety of business forms. At present, the intelligent prediction platform supports them.

There are three main directions in this area:

  • The first is financial forecast, which helps the business plan its financial plan;

  • The second is the sales volume forecast, which is very mature at present. It is the basis for realizing automatic replenishment. It outputs more than ten million lines of forecast suggestions and more than ten million lines of purchase behavior every day;

In addition, it is single quantity forecast, which is more used to serve logistics, make purchase warehousing plan and configure capacity.

At present, the sales volume forecast mainly serves two major businesses:

  • The first is the purchase business, which comes in from suppliers;

  • The second is to serve FDC allocation.

① Predict the problems faced

The previous sales volume forecast is a point forecast, such as the output of 91 days. What is the result of each day. This model is difficult to meet the demand in the current scenario, so it gradually changes from point prediction to distributed prediction. We need to tell the downstream system about the sales volume and probability of occurrence at each point, and give corresponding suggestions based on the distribution comprehensive decision. The change from point prediction to distributed prediction has brought great challenges, such as the challenges of computing power, computing difficulty and prediction difficulty, which have increased exponentially.

In reality, there are many difficulties in making sales forecast, especially in the self operated supply chain.

The first difficulty is the growing customer base, such as generation Z, tob, tog and TOC. When more people are served, there will be more demand for long tails. The long tail itself is particularly unfriendly to prediction. How to accurately predict the long tail? In the industry, it is very difficult to do it well first.

In addition, we need to consider the life cycle of different commodities. The sales characteristics of a commodity vary greatly from introduction to growth to maturity to recession. It is necessary to accurately identify different goods. What are the differences in sales behavior under different life cycles? Take it into account in the sales forecast to improve the results of the sales trend. This is a very big problem in the industry, which has brought us great challenges.

We often see that there are a lot of marketing activities on e-commerce platforms. Each marketing activity will lead to some spikes in sales. Usually, it is the long tail demand of selling one or two pieces, but when it comes to marketing and promotion activities, there will be a massive explosion of demand. The prediction accuracy of stationary series is relatively high, and the prediction of non-stationary series is very difficult, so it is also a great challenge for us to predict the peak more accurately.

② Solutions to problems

In order to solve the above problems, we have disassembled the whole problem:

  • The first is the time series model

In many cases, the training time series model can better solve the problem of sales volume prediction. Therefore, the adaptability of some categories here is relatively good. From our experience, the trend model, including periodic model, performs well, and the model can better solve such problems. Many scenes we explore can achieve better results with time series model.

  • The second type is machine learning model

When the time order model cannot be solved, it is necessary to introduce machine learning model, including deep learning model. What kind of effect is better solved by this model? For example, the information of promotion and marketing. Complex information is difficult to take into account in the time series model, because his learning ability is relatively weak. In this case, the effect will be improved by introducing the machine learning model. From the perspective of practical exploration, XGB and TFT models can better solve these problems, but the demand for computing power will increase significantly.

  • The third category is the new product model

The new product model is a difficulty in the current industry. We have also explored many ideas for new product prediction. At present, there is no particularly perfect idea. A useful idea is the logic of similar products, which will play a greater role in the prediction of new products. For example, for many products, after the introduction of the platform, some old products have very high similarity with it. After the introduction of it, there is relatively perfect data to do model training for new products, and the effect is relatively good.

  • The fourth type is routing model

Because the above three models improve the effect of differentiation for different commodity layers after algorithmic personnel analyze the commodities, but the efficiency is relatively low. Therefore, we propose a routing model, hoping that through the automatic identification of the algorithm, we can automatically match whether there is an appropriate meta model adaptation for each commodity in a certain period of time. After more than a year of exploration, we have found some selection mechanisms that can well match the relationship between commodities. From the actual effect, the routing model has played a very important role in improving the accuracy of the market. At different frequencies, the rise of pulse is more than one point, and the rise is very huge. Therefore, the routing model is a key model at present, which can improve efficiency and prediction effect.

Prediction framework

The prediction framework is similar in the whole industry. The data system of large factories is very perfect, and the data of commodities, users, orders and promotion are very complete.

On the basis of complete data, a lot of work is done on the construction of differentiated feature database. Very good matching results can be achieved with relatively simple data. For example, marketing activities, how to build the characteristics of marketing activities. At present, we can take the marketing activity information as a feature and put it into the model, but the output effect of the model is often not ideal. The whole marketing activity is divided into price time series and marketable time series, and then combined into the existing model, which will achieve better results. Therefore, feature engineering is the work of excellent algorithm engineers. Of course, it is also a very difficult work.

The construction of model library and component library is a relatively mature technology.

In order to improve the efficiency of the online algorithm, we have improved some mechanisms, including the online process evaluation system and back test system since last year, which greatly accelerated the online speed of the prediction algorithm.

5. Replenishment

Replenishment is a job with business attributes. If you want to do a good job of replenishment, you must understand the pain points of replenishment on the business side. Only by understanding its pain points can you turn the manual replenishment of the business into the automatic replenishment that is constantly promoted.

① Problems faced by replenishment

For example, the magnitude of Jingdong’s 9 million goods is very terrible. How to achieve a relatively good and recognized result under reasonable circumstances. Among the 9 million commodities, there are not many explosive products, most of which are conventional products. It is difficult for the product itself to predict and replenish. In the case of great prediction uncertainty, how to find out the bottom through the replenishment model and give good results is a very challenging problem. In the process of replenishment, the business model is very complex and the network changes a lot. When the network changes, the parameters must be adjusted adaptively, which also brings us a great problem in this process.

For us, the purpose of turning manual replenishment into automatic replenishment is that there are many kinds of goods. If we use people to replenish, the efficiency is very low, the workload is very large, and it is difficult to achieve refinement. When the system is replenishing goods, it can make differentiated prediction for each SKU and give different parameters to make the effect better. In this way, offline manual replenishment can be changed into automatic replenishment.

At present, the automatic replenishment has reached the level of more than 70%, which is very encouraging for us, which makes the efficiency of automation very high and brings more space for algorithm optimization.

② Solutions to problems

If you want to replenish well, you must disassemble the inventory more clearly and match different replenishment strategies and model algorithms for different inventory levels. At present, we mainly disassemble the existing inventory based on the safety inventory model. Disassembly based on inventory model:

  • The first is turnover inventory: in an ideal situation, turnover inventory only needs to cover the amount of replenishment interval. However, in many cases, due to the instability of supply and demand, the supply will change greatly and the demand fluctuates greatly. If only turnover inventory is prepared, it will cause the risk of out of stock.

  • The second is safety stock: in order to avoid the stock out risk caused by turnover stock, safety stock is introduced to meet the needs of consumers. For example, different safety stocks correspond to different service levels. For example, the spot rate reaches 97%, and the increased safety stock is much greater than 95%.

The last is strategic inventory: in addition to turnover and safety inventory, there are some strategic inventory. Strategic inventory is to deal with some special scenarios, such as the big promotion and preparation scenario. For example, during 618 and double 11, the whole rhythm is greatly affected by the supplier’s capacity and the capacity limit of logistics warehousing. It is difficult to collect the goods a few days before 618 and double 11. In this case, it is necessary to introduce the goods rhythmically. This part of inventory is strategic inventory. Its magnitude is often very large, which brings great challenges to optimize turnover.

To meet this challenge, we have introduced two common replenishment strategies:

  • The first category is regular replenishment

Regular replenishment serves safety stock and turnover stock more. In this process, we need to do two kinds of work: recommend a reasonable inventory level based on the uncertainty of prediction, and then match the corresponding parameters based on the reasonable inventory level. In the process of parameter recommendation, we will use the optimization model of operation research to give the parameters. For example, in order to reach a certain inventory level, different parameters and suggestions should be given for different commodities, such as the number of days to prepare goods and the service level. Through the parameter model of conventional replenishment, we have achieved a very high automation level of 70%.

  • The second category is to promote replenishment

Promoting replenishment is a very special scenario, because its volume erupts very large at a certain point in time. In this case, we will disassemble the warehousing rhythm and quantity in a long cycle based on the forecast, business purchase, sales and inventory plan, logistics capacity and supplier capacity. The difficulties here are as follows:

  • The first is the sales forecast. Our forecast team has done a lot of model optimization for the sales during the promotion period. From last year to 618 this year, after the new promotion forecast was launched, the achievement of the whole plan has been greatly improved.

  • The second is the matching of warehousing rhythm, which is more linked with logistics measurement.

In addition, in many cases, business plans will be adjusted, which will have a great impact on the output results of large promotion and preparation. At present, through two rounds of pilot, we have solved the main challenges. From the current follow-up effect, the adoption rate and automation execution efficiency are very high. The adoption rate of big promotion and preparation is more than 80%, and the automation rate is maintained at more than 60%.

③ End to end replenishment

A cutting-edge technology we have been promoting since 2020 is multi terminal replenishment. The core logic of replenishment is to make prediction first, and then combine the replenishment model, such as parameter recommendation, to place a purchase order. One disadvantage of this method is that there are many links, and there will be an error accumulation in each link, making the prediction effect worse. Therefore, we want to turn the two links of prediction and capture decision into one link, and directly give optimization suggestions to further improve the effect of replenishment in this way.

A lot of work needs to be done to realize this concept. The first step is to give each capture behavior in history. What is the best capture experience? The best replenishment accuracy of each purchase order in history is calculated back through the supervision model. As a training sample, give the large model and give the best suggestions through deep learning training.

In the whole end-to-end model, the current sales forecast, VLT forecast and replenishment decision will be combined into a large model through a large neural network, and finally give suggestions. In order to improve the interpretability of the whole model, the results of the intermediate process, such as sales forecast and VLT forecast, will be output for your reference.

From our verification results, in the actual landing scenario, we selected three specific categories for online experiments, and the experimental results are very good. The turnover and spot can reach double liters, which can achieve better results for both conventional products and best-selling products.

In the past, the business required us to predict more accurately, but now it is increasingly expected that our prediction will be more explanatory. We should not only have a stable forecast, but also explain why today’s forecast is different from yesterday’s, or there is a peak in the future. What is the reason for this peak.

We are also constantly exploring the interpretability of prediction. How to do it? At present, in the process of exploration, we have two steps to take:

  • First, the interpretability of the prediction process, that is, what actions have been taken and the number has increased;

  • Second, the interpretability of the results. We need to disassemble the predicted total amount. Which marketing volume is generated by promotion and which is generated by price. We need to disassemble the results to facilitate the tracking of badcase after business work. We have better interpretation ability.

④ Integration of mining and distribution

The integration of purchase and distribution, forecasting, replenishment, allocation and clearance of delay are carried out in series. In this process, because the decision-making chain is relatively long and there are many nodes, the technical practice of integrated decision-making is very difficult and the robustness challenge is also great. With the continuous improvement and progress of technology, how to achieve integrated decision-making in the future is a subject that needs to be overcome. The JD scenarios we need to solve include purchase and allocation. Transfer includes the transfer between large warehouses, from large warehouse to front warehouse, and from front warehouse to large warehouse. There are many factors to be considered in the whole process, such as turnover, cost, damage and so on. We are also exploring how to improve the integration level of procurement and distribution products and accelerate the improvement of efficiency.


Future outlook

Our vision of Y is to realize the super automation of the whole supply chain in the future, from single point automation to full link automation from selection, pricing, inventory to performance.

The automation process is very difficult. From the current evaluation, we are more in the second stage. Some links can be deeply managed, automatically set parameters and automatically issue orders. There is a complete set of automation process to achieve a very high level of automation. In the future, we hope that in this process, the whole process can be completely executed by the system without manual intervention. Finally, we hope to realize the automation and adaptive operation in all fields, and realize the goal of technology to better serve the business.


Wonderful Q & A

Q: After the introduction of distribution prediction and alternative point prediction, how much is the index of turnover spot loss improved? Are there quantitative analysis results?

A: We have done several rounds of evaluation before. After switching from point prediction to distributed point prediction, it does have an impact on the effect, but the specific quantitative results are difficult to give specific values because of some factors.

Q: What suggestions do you have for the sales volume forecast in the peak period? For example, how to predict the big promotion, or some special holidays, or some emergencies?

A: From the perspective of the whole industry, the promotion link needs to be predicted on the one hand and planned on the other. For example, when the business is doing trading, it will set itself a goal, which has a great impact on the prediction. If the sales target of the business side is one billion, it will move towards the target of one billion when doing business marketing means. If you can get the whole sales target and take the target into account in the prediction environment, the improvement of prediction effect is very significant. In particular, the effect will be significantly improved on the nodes with large 618, double 11 and new year goods Festival.

Q: Does the end-to-end replenishment decision model have applicable requirements for the amount of data? Will the end-to-end decision-making model replace the two-step process of prediction and replenishment optimization in the future?

A: First answer the first question. From the current point of view, the mqrnn terminal we do generally requires a large data set, and the training samples will be at the level of hundreds of thousands. If it is too small, the effect may not be particularly good.

The second problem is that the development of technology is spiraling. At present, the prediction and replenishment optimization will be very explanatory. If people in the downstream make replenishment, they may prefer the idea of prediction and optimization. If there are not many business students in the downstream, the prediction plus optimization becomes the end-to-end method, which will have less obstacles when it is actually implemented. In the long run, the end-to-end model will gradually replace the prediction plus optimization model.

Q: Strategic inventory comprehensively considers the uncertainty of big data prediction and warehousing rhythm. What is the essential difference between strategic inventory and safety inventory?

A: The biggest difference between strategic inventory and safety inventory is the targeted extra part of inventory. At present, we recognize that the reason for promoting the preparation of goods is that the demand fluctuates greatly and can not be met.

First, limited by some production capacity, for example, the quantity of goods used is too large, and the supplier cannot produce in a short time.

Second, due to the limitation of logistics bottleneck, it is impossible to have so much capacity in the short term.

In this case, it is necessary to differentiate and prepare the goods in advance according to the rhythm. The stock volume is often much larger than the current safety stock level to meet the sales, which is an explosive volume. This inventory is called strategic inventory because business strategy and some trading means make its inventory come in ahead. Strategic inventory is a kind of strategic inventory in which suppliers may have cheaper prices and I may be more willing to prepare more goods under the behavior of excessive and large purchases. These inventories can be brought in in in advance.

That’s all for today’s sharing. Thank you.

For more technical dry goods articles, please follow the wechat official account “datafuntalk”

About us:
Datafun: focus on the sharing and exchange of big data and artificial intelligence technology applications. Launched in 2017, more than 100 + offline and 100 + online salons, forums and summits have been held in Beijing, Shanghai, Shenzhen, Hangzhou and other cities, and nearly 1000 experts and scholars have been invited to participate and share. Its official account datafuntalk has accumulated 500+ original articles, million + reading, and 130000 + accurate fans.

Note: welcome to reprint, please leave a message or send a private letter.