Transforming Supply Chains Through Advanced Predictive and Prescriptive Analytics
We all know that the Supply Chain industry is undergoing a huge transformation right now. Over the next few years, Predictive Analytics is expected to reach an adoption rate of 87% and Artificial Intelligence, which was only at 13% in 2019, is expected to reach an adoption rate of 62% [1].
Legacy systems are based on archaic business rules and need manual inputs, estimates and experience. For a while these systems worked well, but as the world becomes more dynamic, we need better solutions. We need systems that can crunch years worth of data in minutes and augment Supply Chain experts’ experiences.
As the world moves away from black box solutions to open technologies, why should the Supply Chain remain trapped inside a black box? Best innovations in the advanced analytics space are happening on the open source side right now. Can you see where I’m going with this? Yes, I’m talking about bringing the power of open source data science, machine learning, and decision optimization to the field of Supply Chain.
In this blog we’ll tackle demand forecasting and optimized distribution. To put things into perspective, let’s look at these concepts with the help of an example.
Meet Sandra from BlueCoMart.
Sandra is a Warehouse Inventory Manager for BlueCoMart’s West Coast Region. She wants to understand the demand for each product over the next few months. Then she must verify if her current inventory levels will meet the expected demand, and place orders if they won’t.
She relies on legacy systems, manual interventions and her judgement to make decisions for thousands of products each day. She shared her pain points with BlueCoMart’s new team of Data Scientists, and they are now building a solution to help her out.
To get the most out of data science, machine learning and decision optimization, we’ll break down her challenges into the following categories:
📝 Inventory & Demand — she needs custom data visualizations to get insights at her fingertips.
📝 Demand Forecasting — she needs more accurate forecasts from advanced predictive methods.
📝 Supply & Logistics — she needs optimal shipment recommendations quickly.
Part 1: Inventory & Demand
This is an example of how her inventory and demand data could be visualized in a dashboard. This would give her a quick overview of where inventory can meet demand and where it cannot.
The green bubbles are product categories where inventory can meet demand. The red bubbles are product categories where it cannot. The sizes of the bubbles are the number of products for which inventory will fall short of demand.
There can also be an option for her to drill down and assess inventory and demand at the product level.
The data scientists share this draft with Sandra and she likes what she sees. Do you want to know the best part about this? This is built using R Shiny, a powerful open source tool that makes it easy to build interactive web apps straight from R. So, if data scientists realize this doesn’t help solve Sandra’s pain points, they are able to update the interface. They can remove information, add information, change color scheme, and more. It’s completely flexible.
Part 2: Forecasting Demand
The next step is to provide Sandra the demand forecasts, which might look something like this:
This can also be displayed in the form of a graph. This table looks simple but it can be backed by powerful machine learning and deep learning methodologies. Let’s see how.
Recently Sandra has had trouble with the accuracy of the forecasts from their black box system which uses traditional statistical forecasting methods. The data scientists use this opportunity to look at more advanced predictive methods. They know that independent product demand prediction is fundamental to their business. If they don’t get this right, then none of the numbers propagated downstream will be right, potentially leading to widespread disruptions. They begin brainstorming.
One of my favorite solutions to address this type of challenge is the Long Short Term Memory Network AutoEncoder (LSTM-AE). An LSTM is a neural network that can capture past interactions within your time series data and make accurate predictions. This is what an LSTM architecture might look like:
The LSTM can also capture external features. At some point, however, there may end up being too much noise to identify the signals. We want all the data, but none of the noise. What can we do? We can take the data blocks from above and stack them against other similar blocks from external features. But instead of using just an LSTM on this stacked data, we combine it with an AutoEncoder.
This is what that LSTM-AE architecture might look like. We combine product history with external features and their histories, then pass all of that through a hidden layer to get our forecasts. Sounds simple, doesn’t it? Under the hood it’s a bit more complex, but the major takeaway is the secret sauce of this architecture: the hidden layer. That layer compresses the data into a smaller dimension, which helps preserve patterns that are relevant, thereby eliminating some of the noise.
The Data Scientists at BlueCoMart try out different machine learning models and finalize one which gives them best predictions on their validation set (I’m guessing it might have been an LSTM or an LSTM-AE). They use it to obtain real-time forecasts for Sandra.
Part 3: Supply & Logistics
Now Sandra can see the gaps between inventory and demand and knows future demand. But she still has to make decisions about product routing optimized for time and costs. Doing this for a few products is manageable but doing this for multiple warehouses over thousands of products is not an easy exercise. Let’s see if data and math can help Sandra out a bit.
Sandra needs a solution that includes details on BlueCoMart’s production facilities (suppliers, manufacturers) and warehouses along with demand. BlueCoMart’s Data Scientists regroup to discuss what conditions and constraints they may need to keep in mind before they build a solution.
As they discuss these factors, they start to empathize with Sandra. BlueCoMart has an out-of-the-box solution already, but it doesn’t address all of Sandra’s challenges, so they decide to build their own models. They build an optimization solution that models production activities (number of products per production line for each warehouse) and backlog (the demand that cannot be met). The model is built to minimize production cost, travel cost and backlog cost, constrained to meeting demand and staying within suppliers’ capacity limits. The outputs are added to Sandra’s dashboard.
This plan will help Sandra place optimal orders to meet her product demand. Sandra can utilize the final routing plan at various levels of granularity.
She can see the flow of product volume from 4 suppliers to 3 manufacturers to her 4 West Coast warehouses at a high level.
And of course, she can dig into detailed product-level routing plans. Similar to the demand forecasting output table in the dashboard, this table looks simple, but showcases an in-depth output of a Mixed-Integer Programming Optimization Model.
The optimization model’s complexity could be further increased by incorporating demand uncertainty into the model mix. This solution is built from scratch by leveraging open source technologies, so it can be completely adapted for Sandra’s needs.
As we saw, the data scientists at BlueCoMart leveraged only open technologies to build a solution that addressed some major pain points for their Warehouse Inventory Manager, Sandra.
TensorFlow, Keras, PyTorch, docplex, R Shiny and Plotly are all open source options to choose from with thousands of contributors adding new features and functionalities to them every day. BlueCoMart’s Data Scientists build their solution using Keras with a TensorFlow backend, docplex and R Shiny.
This solution was built to address a specific set of requirements for Sandra, but we could expand the same concepts to cover a variety of other challenges in BlueCoMart’s Supply Chain. It is open, transparent, trustworthy, modular, and completely customizable — it represents the future.
Research shows that with the right methodologies, forecasting errors can be reduced by up to 50% and lost sales due to product unavailability can be reduced by up to 65% [2]. Enterprises know that even the smallest improvements can save them millions of dollars. This awareness is driving large-scale infusion of machine learning and artificial intelligence into next-generation Supply Chain systems and I hope that reading this had made you excited to try and leverage some of these techniques within your own Supply Chain solutions!
Are you interested in implementing something similar for your own Supply Chains, but don’t know where to begin? Request a test run of the Supply Chain Forecasting and Optimization Industry Accelerator to get started.
The IBM Data Science and AI Elite plans, co-creates and proves the project with you based on our proven Agile AI methodology. Visit ibm.biz/datascienceelite to connect with us, explore our resources and learn more.
References:
[1] 2019 MHI Annual Industry Report: Read Here
[2] Smartening up with Artificial Intelligence (AI), McKinsey: Read Here
