Data Science Application in Supply Chain and Logistics

Originally written as coursework for EAS 504LEC — Applications of Data Science: Industry at the University at Buffalo. Created: August 2022.

1. Introduction

Supply Chain Management has become an integral part of the world economy. Here logistics and supply chain are two separate problems. Logistics is the phase that comes after goods are made into finished forms. Supply chain is rather comprehensive, it involves procurement of raw materials, collaborating with various parties involved for designing, manufacturing, transporting and selling. With the increase in demand of products, it becomes necessary for companies to have ample supply to fulfil the demand. Companies have to look for unexplored ways to improve efficiencies across various sectors of the entire workflow. Traditionally supply chain companies have performance based metrics like defining targets, schedule planning, monitoring and reporting in place to identify inefficiencies. This paper will go through how data science will help in identifying inefficient patterns and determining optimal strategies among various sectors of the workflow.

1.1 Promise Of Data Science

As Supply Chain is an evolving field, it has incorporated many tools that help in tracking during various processes like inventory management, warehouse management, customer requirement procurement, logistics, returns management and sourcing and supplier management. SCM has to manage and synchronise tasks between various entities organised in serial workflow. Each team uses a specialised tool designed for their use cases. Hence a lot of data is generated and is spread across many tools. With the help of data science, data from various sources can be aggregated, cleaned and transformed to extract interesting insights. This new aggregated data can be used to build predictive models which would help in activities like demand planning, supply network planning, production planning, detailed scheduling, logistics and analytics.

1.1.1 Robotic Process Automation

RPA comes under business process automation which uses robots to perform a specific set of tasks. These robots with computer vision capabilities can be equipped with advanced tasks. During any manufacturing steps of things like car, there are steps in the process which are repetitive. Here RPA could be used so that humans can focus on other important tasks. It improves efficiency and also helps in maintaining strict quality controls. RPA can also be used for making intelligent decisions. For example an RPA can rely on computer vision and historical data to better identify an end product as defective. With historical data we could build build models that would minimise errors in these processes.

1.1.2 Internet of Things

Internet of things or IoT is a group of devices which are connected to the internet with wifi and are capable of gathering and transmitting data. IoT devices are capable of measuring data with high accuracy like temperature, movement, amount of light, movement , GPS. These devices can be attached to containers of raw materials so that we can track the movements and predict its shipment time. Additionally at warehouses using these devices storage conditions can be monitored. Continuous streams of data from IoT devices can be used to determine the loss of quality of raw materials due to factors like seasonal changes etc.

1.2 Problems In the Past

In a Supply chain management workflow, the coordination between different entities is really challenging. Some disruption at a point in workflow will have ripple effect in the entire system. These disruptions can be sudden and are almost unpredictable. The SCM managers have to assess carefully study the processes and try to think of scenarios which could disrupt the entire workflow. We will see a few such problems that leads to disruptions.

1.2.1 Global Chip Shortage

Covid-19 pandemic has led to massive disruptions in global supply chain. As workplaces were temporarily closed because of pandemic, large companies ranging from car manufacturers to pc makers have cancelled their orders. What they had not anticipated was the increase in demand of digital devices. With most of the workforce forced to work from home, people because more reliant on digital devices. Hence there was a sudden rise in demand of equipments. But the semiconductor manufacturers were unable to keep up with the demand. Hence it triggered the longest semiconductor shortage. It has also spilled to other industries where smart and intelligent devices are needed such as car manufacturing, industrial sensors, automation robotics. This was a serious problem as manufacturing companies relied on these small chips to control critical equipments and because of unavailability of the chips the entire assembly had to be stopped.Data Science will help in capacity and demand planning. Another interesting way Artificial intelligence could help is with Silicon Mastering. Chips are designed faster using this technique hence reducing years of work to just weeks. This would save hundreds of millions of dollars.

1.2.2 Shipping Delays

Shipping is an important process in a SCM. The workflow begins right after shipping of raw materials. Traditionally companies have put up parallel shipping processes where raw materials are delivered as their needs arrive to optimise storage costs. Hence it’s critical to have estimated shipping details. With a supplier’s historical data, manufacturers can choose who has a better track record of delivering goods without delays.

As per Deloitte CFO signals survey of 2021, shipping delays have increased cost to company by 5% and 60% of companies have reduced sales. As these companies have adopted several tools for tracking shipments, they also have enough data to be able to predict the demand. An item like silverware will have peak high demand during the festive season. By applying Time series analysis and prediction one can determine seasonality and adjust order of raw materials as per the demand.

But what about demand that arises suddenly. As customers can be unpredictable sometimes. It would become extremely hard to determine unexpected surges purely based on historical data. For this data mining needs to be applied on social media websites where likes, share, comments on posts can be mined to predict an unexpected surge.

1.2.3 Tesla Supply Chain Issues

The pandemic has lead to one of the worst shipping issues in history. Tesla’s factories at Austin and Berlin were loosing billions of dollars. As critical battery equipment was stuck in shipping from china. Because of this these factories were working in full force but there was no car output. It also lead to a 140% increase in price of raw materials hence shooting up the price of EV vehicles.

In Large scale Software System design, we always plan for contingencies. For example, In case of Netflix, it has hundreds of small services which collaboratively work to produce the best streaming experience. Netflix applied a special tool which would randomly shut down some services in realtime to check its impact on overall system stability. So when these services actually shuts down, their engineers have protocols placed to overcome it. The same principal could be applied to supply chain managements. With historical data, we can design a simulation where reducing the profit would be the AI goal. The AI would become creative in finding ways that would lead to profit reduction. By observing which factors led to such profit reduction, supply chain companies could use that information to put up cushions against any such situations.

2. Background

To better understand how data science would help a vast topic like Supply chain management, we have to understand some key domain concepts. I would be discussing some key topics where data science would be vastly helpful.

2.1 Supply Chain Workflow

A typical supply chain workflow contains three major flows. Material Flow, Information Flow and Money Flow. Material flow is the flow of items from producers to consumers. It is done through various entities like distributors, dealers and retailers. It is important to ensure continuous supply of requirements. Information Flow as the name suggests contains of quotations, schedules, tracking information, purchase orders. Money Flow is the step when the supplier sends the invoice to consumer and the consumer validates the receipt. The entire process starts at selecting a vendor for procuring raw material. Next is the sourcing stage where the best supplier is chosen based on efficiency and cost. Once a source has been finalised, the raw materials needs to be transported. It could also be a parallel process where raw materials would be delivered as and when it is required to avoid storage costs. Once the raw materials are procured, the entire process of creating a finished goods takes place. The finished goods are then stored in temporary storages or are sent to distribution centres. Distribution centres are then responsible for making the product available in retails stores from where it reaches the consumers.

2.2 Logistics/Delivery

Productivity in logistics is projected to be improved by 40% by 2035. Companies adopting AI in logistics are already having improved profit margins of 5%. Logistics is the network of transportation, warehouses and inventory. It enables the transportation of goods through most efficient routes. A full fledge logistics companies typically owns a wide range of infrastructure from trucks to ships to jet planes. Other important part is warehouse and inventory. A logistics company needs warehouses for temporary storage or halt points during transportation of packages.

Logistics companies are now relying on IoT devices. These devices can give shipping companies wide variety of details about vehicles like GPS, fuel efficiency, vehicle performance. Realtime tracking of fleet has enables operators to optimise routes, dynamic fleet scheduling along with using data from weather and google maps to accurately predict turn around times of travelling.

Amazons has an interesting anticipatory shipping strategy. With the enormous customer data, amazon looks the buying pattern of people and predicts items that they are going to buy next .This helps in accurately predicting the strength of delivery fleet, determining the most efficient way to organise the warehouse goods. It ensures a sustainable last mile delivery.

2.3 Supplier Selection

It is a process which helps SCM companies to identify suppliers based on some evaluations and metrics. This step has a large amount of financial capital deployment. In exchange of that companies expect high offering benefits from suppliers. Supplier selection involves various steps like identifying suppliers, getting historical data from suppliers, setting contract terms , negotiations from suppliers and evaluating suppliers. To avoid any unforeseen outcomes where suppliers become non-performing, the sourcing companies take additional steps in verifying the data before the contract is awarded to them. The supplier has to also have the capacity to handle unexpected surges. Additionally, the supplier should be able to meet a wide variety of requirements like references from previous customers, financials strength, quality of products, specification compliances. Apart from that, the consumer must have a constant monitoring of the quality of supplied materials. In case of defective products, the supplier should be able to detect it before it reaches assembly lines otherwise it would lead to wastage of time and sometimes even the finished products might have to be thrown off because of an inconsistent quality raw materials. In most of these steps machine learning could reduce selection errors and help in selecting the best suppliers as per the requirements of consumer.

2.4 Risk Analysis

Supply chain is bound to be shocked by unforeseen exposures. It could cause disruptions leading to huge loss in terms of time and money. It has two kinds of risks, because of uncertainty the consumer could be left with depleted supply where as overestimating risk would mean that consumer has over supply in inventory. Hence forecasting risk is very challenging. Findings presented from electronic industry shows that equipment manufacturers could not predict demand after a four week timeline. In SCM, there are sudden demand amplification or Bullwhip effect where small demand changes in consumer buying would result in surges in demand for raw material for suppliers.As a ML model detects a change in buying patterns it would suggest an increase in raw materials. Factories would be overwhelmed by this sudden demand. And many a times these demands are short lived, leading to an over sized inventory. Risks are classified into two types Internal supply chain risk and External supply chain risk. Internal factors are like operation disruptions, changes in management, inadequate contingency strategy, cybersecurity, environmental compliances. Some external factors are misjudged consumer demand, delays in raw material deliveries, natural disasters and pandemics.

2.5 Distribution

Distribution is an integral part of SCM. It is a connecting link between many parts of supply chain. For example raw materials are transported from supplier to manufacturer and finished goods are transported from manufacturers to consumers. As per reports distribution helps in driving profitability for companies as it is directly involved with supply chains and customer. Companies like Walmart and amazon have had their success entirely because of excellent distribution operations. Vehicle routing is an interesting challenge in distribution which takes advantage of machine learning. The distribution network typically has the option of choosing from a wide variety of transportation vehicles. Using ML and historical data, efficient routing can be generated based on historical data. In a real world scenario research, a neural network yielded better accuracy than traditional heuristics by 48%. Machine learning algorithms prove to be efficient in finding optimised routes tailored for the company requirements.

2.6 Tracking and Reporting

Efficiency of a supply chain can be improved when the managers have sufficient data to understand where the problems arises. Tracking provides an excellent way to finding and improving inefficiencies. A supply chain uses something called as key performance indicators (KPI) which is a quantitative way to evaluate the performance of supply chain. KPI is set in all the junctions where some kind of transaction involves.It includes entities like manufacturers, vendors, transportation, distribution and warehouses. Different KPIs are used to measure different entities like Financial SCM Kpis are used for determining financial performances like gross margin return in investments, freight cost per unit, cost of SCM as percentage of sale, cost of SCM per item sold. Others KPIs include Inventory management, Transportation management, Customer management. Combining KPIs with AI/ML would prove to be highly powerful in improving efficiency in supply chain. By taking the example of last mile delivery which is tied to KPIs, AI models can help in determining the best route which would lead to better KPIs. As these metrics extend from manufactures to suppliers AI models would be able to find solutions that would lead to higher KPIs.

3. Methods

In the above discussion it was evident that Data science and machine learning have interesting applications in all facets of supply chain management. Some of the methods that are used here are:

3.1 Time Series Forecasting

In Time series modelling, time series data is analysed using statistical modelling techniques and predictions are made based on historical trends. The prediction is not always accurate and it sometimes fluctuates highly depending on external factors. It is performed to gain understanding of data to know the underlying causes. Analysis would tell us the reason for certain outcome. Then follows forecasting to use the knowledge in making prediction and extrapolations of outcomes of future. Man made processes like stock market, medical diagnosis can generated time stamped data. In case of Supply Chain Management, time series is used for demand forecasting. With SCM tracking tools which record demand data over time accurately, it creates a perfect training data for supplying to ARIMA models. The time series methods are combines with other features enhances demand forecasting. Another method was proposed called demand based trend-mining algorithm. Prediction accuracy in this model is reduced when there is a high level of uncertainty in parameters. A time series model generally examines the historical events such historical averages, trend lines, data seasonality, cyclicality and outliers. A time series is observed when there are N observations equally distributed on timeline. .The main use of time series modelling is to find the model that can accurately represent the future and historical patterns. . Here is the observation of demand at time t. This pattern can come in all combinations like random, trend, cyclic, seasonal. Therefore the which is the time series, a linear function of actual historical values and the shocks which occur randomly. The ARIMA model has become quite popular due to its flexibility and power. But it requires prior experiences to fully make use of the technique. An ARIMA model has 3 parameters . The label is count of autoregressive terms. The label is the number of differences.The label is the number of moving averages The identification of model starts with preprocessing the data to make it stationary. In this we choose value of label which can be adjusted as the data is fitted into model. The stationary nature of data can be identified by performing the unit root test (Dickey-Fuller test). The accuracy of the model can be evaluated by checking the overlap between experimental and actual/simulated data over the same period of time. Demand forecasting in a crucial step in a supply chain workflow. It is integrated heavily with other processes which makes demand planning a very important.

3.2 Clustering

Clustering is a data science technique which partitions similar kind of data based on their similarities. A few application used in supply chain management is business analytics and pattern recognition. Using clustering customers can be grouped into similar groups to better cater them as per the groups needs. Clustering is yet another technique that would help in demand forecasting. Demand can be estimated based on the cluster size. Another interesting application of clustering is in warehouse planning. For example if a customer orders contains multiple items, then shipping time would be increased. Historical data is used to determine clusters of items that are bought together. With this knowledge warehouses can plan placing items from grouping items from same cluster together in same warehouse. The impact of this approach also has a direct impact on environment as it reduces the amount of packaging.

3.3 Neural Networks

It is an emerging technology which is a result of advancement in modern biology of how human brain works. A breakthrough in neural networks is Back Propagation. Back Propagation network was developed in 1986 by a group of scientists led by Rumelhart and McCelland. It is an algorithm for training multilayer feedforward network by back propagating error. The method for steepest descent is used to adjust the weights and threshold values in the network. A back propagation neutral network consists of one input layer, with one or more hidden layers and one output layer.

The learning happens due to forward and backward propagation. In forward process information from the first layers is sent through the middle layers and then the output layer. Once the output is received, by using the error function, the error is proposed back to the network. At this time, the weights at each neuron are adjusted until the output layers return response in acceptable degree of error.

As in a process like supply chain management, which is highly dynamic, neural network could be greatly efficient in solving such non-linear problems. A few areas where neural networks could help SCM are Optimisation, Forecasting, Decision Support. Optimisation problems can be efficiently solved by Neural networks. Various studies are going on where neural networks are being tested on optimisation problems like scheduling , managing warehouses, best transportation routes etc. Forecasting as we know is a non linear problem. Predicting demand accurately can be challenging and it sometime zooms into a bigger problem because of slight changes in customer buying pattern. Traditionally forecasting is done with statistical methods, domain experts and time series. But a neural network being a non linear system would be able to better predict the demand. Decision support is another interesting challenge in SCM where managers have difficulty in making decisions because of the amount of data and sometimes because of incomplete data. A neural network has interesting ways to classify important information to make decision making faster.

3.4 Mixed Approaches

There have been some interesting studies where combination of machine learning algorithms are used in a workflow for example, output of one algorithm is the feed to next step. Each step extracts some kind of useful information.

Supply chain can become very complex with suppliers, warehouses , products and customers, forecasting demand could become a higher dimensional problem. As per a paper, retail demand forecasting model based on data mining techniques, patterns of sales across different products are analysed using a technique called bipartite graph clustering. In the next step the accuracy for forecasting demand of each clusters was improved by using a bayesian belief network and a moving average model.

According to another paper called used time series, regression, support vector machine and deep learning in sequence to develop an intelligent demand planning model. According to the model, first for dimensionality reduction, principal component analysis was performed to extract meaningful features. Data clustering was followed. Using a novel decision integration strategy called boosting ensemble, demand forecasting for each cluster was performed. Finally it was concluded that by using a deep neural network with boosting resulted in higher accuracy model.

3.5 Genetic Algorithm

These are evolutionary algorithms which try to simulate biological evolution.These are used for solving both constrained and non constrained optimisation problems mainly by using natural selection which is inspired from biological evolutions. The algorithm uses a fitness calculation method which is used to find the perfect fit as the simulated population evolves. Genetic Algorithms has lot of applications of SCM. As supply chain has a lot of complexity optimised solution can be generated with help of biological methodology. A few active research areas are Inventory analysis, Vehicle routing problem and production cost minimisation.

Word Count: 3586

bsybin