Machine Learning in Manufacturing supply chains

Pain areas in supply chain management

Demand forecasting is one of the most critical functions of a supply chain management (SCM) team, and one where the team struggles the most. Current tools often present inaccurate results, causing errors in the procurement process. This usually because these tools are too narrowly configured to be able to adapt to changing market dynamics, thus it is unable to provide the right sort of results required for it to be effective.

This leads to strain between the planning team and the leadership, with questions raised on the quality of data, and skill of the team to configure the forecast tool and identify the right control points. Fundamentally however, the problem is that data collected from customer demands is becoming more complex with newer data generation methods coming into play. Thus traditional formula driven forecasting tools find it nearly impossible to consider all these new datasets into a cohesive and accurate prediction. Historical demand based forward prediction methods are all but useless in the modern data landscape, as well as customer consumption patterns. (Pal)


What Machine Learning brings to the table


Andrew Ng, the global leader in gauging the future impact of Artifical Intelligence, said that AI is the new electricity (Andrew Ng Interview).  Electricity revolutionized the way our world works around us, and AI is billed to bring about a similar level of transformation in our lives, even if it wont be as visible. In the supply chain the impact of AI is discussed in detail below

The Revolution has begun

AI is already transforming supply chain planning through:

  • Collapsing traditional roles – Cutting edge tools allow the supply chain planner to independently collect all relevant information and arrive at informed and accurate decisions, based on and end-to-end awareness of Customer demand, supplier availability & production capacity. Thus this allows the planner to move from a functional silo to a ‘network planner’ who has a view on the entire enterprise. She is able to run what-if scenarios, using concurrent optimization, and provide immediate corrective action to the entire supply chain.
  • Faster decision cycles – Just in time operations move to real-time optimization, leveraging both structured and unstructured data. The decision cycle is thus accelerated, with corrections being made in immediate response to a signal received from the feedback loop.
  • Clarity of information without human bias – The biggest value of AI is to evaluate data without bias, and to make decisions without missing pieces of the puzzle. As long the data is available, the system will take cognizance of the message in the data and adapt accordingly. This changes the conversation from vague opinions such as “poor freight planning of supplier” to “65% of shipments from the supplier were delayed”.


Use cases


  1. Machine Learning in Forecasting Demand– forecasting demand for the future, forecasting the declining and end of life of a product on a sale channel and the growth of a new product introduction
  2. Machine Learning in Supply Forecasting – based on supplier commitments and lead time – Bills of material and PO data can be structured and accurate predictions can be made in supply forecast.
  3. Machine Learning in Text Analytics – This mainly is due to data cleansing to drive better master data. Text analytics can be implemented with supply data, partner data, shipment data to derive better insights from the supply chain.
  4. Machine Learning in Price Planning – increase / decrease the price based on demand trends, product lifecycle and stacking product with competition.
  5. Machine Learning in Inventory Planning– automatically raise POs with suppliers based on shortages or future demand shortages.
  6. Machine Learning recommendations in drop shipment business – recommend products that are in excess and automatically reduce price to clear inventory. Based on past buying patters by customers in the past recommend products based on inventory position.
  7. Machine Learning in Stock Analytics– based on multiple structured and unstructured data the machine can now predict why we ran out of stock or when we will run out of stock accurately.
  8. Machine Learning in Exception Analytics– Stock outs at every level in the supply chain. Retail customers particularly feel the need of the machine to study root cause of stock outs and predict accurate demand trends with better lead times from suppliers to reduce stock outs.
  9. Machine Learning in Component Level Analytics– planning supply on a component level with dynamic replenishment based on raw material planning has become a reality.
  10. Machine Learning in Production Planning– Using sensors and production automation mechanics to increase / decrease products + increase quality based on realtime customer feedback.


Methods to build a Machine Learning Solution


Machine Learning is typically of two types. A simple summary of what they mean is explained below:

  • Unsupervised learning – this is about finding hidden patterns in data. It is about identifying common threads that link different examples in the dataset and then segment the dataset using those threads. It is most effective in grouping a diverse dataset. It has been effectively used in marketing to arrive at distinct customer segments which then can be targeted more effectively. This method can possibly be used to group datasets into clusters which correlate with a particular forecasting method. Thus the system can adapt its methodology in real-time, in response to changing data dynamics. Another example is to group SKUs which have stable demand, but error margins fluctuate
  • Supervised learning – this is about using the data to predict a given number, or the category of a dataset. This is the most common methodology deployed in enterprise applications, especially in projecting the next number (ie demand forecasting) or help categorise datasets using far more parameters that would be feasible for a human (supplier categorization)

A third method which is gaining popularity is called Deep learning. This is nothing but a technique to implement supervised or unsupervised machine learning problems. Modeled on the basic architecture of the human brain, deep learning methods work by combining a large number of simple mathematical transforms in different structures, which best fit the requirement at hand. Deep learning is especially powerful in image recognition, an area where machines could not perform well, until this technique came along. In the enterprise, Deep learning will come into its own when diverse datasets of both structured and unstructured formats will need to be crunched together to arrive at the right decision.


Feature Engineering


Feature Engineering is probably the most important step in building a machine learning algorithm, other than selecting the actual algorithm ie. But what does it mean?

A feature is information which may be useful to building a machine learning algorithm. The more features the algorithm gets, the better it performs. For eg: in order to build a demand forecasting algorithm, elements such as the timing of purchases, quantity ordered, price at which purchases were made would be the most basic features required. On top of these, features which help to predict customer demand can be included to provide more information to the algorithm

Feature sets make or break the machine learning algorithm. Basically it is about transferring human intelligence to the machine. The true skill is to identify features which contribute to human intuition, and then incorporating them in a manner which allows the algorithm to extract information from these features to improve its performance. Importance of the feature is secondary, the algorithm is adept at learning that on its own. Our job is to give it as many relevant features as possible.


Incorporating Machine Learning into existing Enterprise Architecture

Enterprises typically have an ERP set up to capture their operational information. This provides a strong platform on which to build Machine Learning applications. SAP HANA for example has a structure to absorb machine learning algorithms built on its data structure. And for all platforms, a custom layer can be built on top of the core data sitting in ERP to deploy machine learning use cases. The typical methodology of software development does not fit in here. And these solutions can typically not be configured using off-the-shelf solutions. The most effective implementations have to be done by the customized route. The engineer building the machine learning algorithm needs to dive deep into the data and its relevance to business to build a truly effective machine learning algorithm.


AI is already playing a role in changing the competitive landscape in supply chain operations. Of course, there are a few companies like Amazon who seem to be out-innovating everyone when it comes to AI, but for the most part, those leading the pack have taken bold bets to experiment with AI in a few distinct areas. Those most at risk are the ones not yet taking the leap, planning to quickly follow. It’s those late adopters who will likely find themselves made irrelevant by the Silicon Valley startups. You don’t have to boil the ocean, pick a few areas to start experimenting to learn how AI can transform your supply chain planning organization – driving top and bottom line impact across your enterprise. (Ramakrishna)




Banker, S. (n.d.). Retrieved from The Arms Race To Leverage Machine Learning In Supply Chain Planning:

Daihes, K. (n.d.). Practical Applications of Machine Learning to Supply Chain Planning: A Few Suggestions. Retrieved from

Pal, K. (n.d.). How Machine Learning can improve Supply Chain efficiency. Retrieved from

Ramakrishna, K. D. (n.d.). AI is already transforming Supply Chain Planning – Are you on Board. Retrieved from


Leave a Reply

Your email address will not be published. Required fields are marked *