How can manufacturers improve back-order predictions using machine learning?
What is a back order?
Before we get into the details of the problem, let’s spend a minute understanding what’s a back order. Imagine a business does really well. Sales numbers are at their all their all-time high, so much so that the demand for the products supersedes the supply side. This may lead to lead to a temporary unavailability of the products. What should the organization do in such a scenario? Should it stop selling? Certainly not. Instead, the organization should accept the consumer’s order and fulfill later subject to the availability of the product. Such an order is called as back order.
Why should you be concerned about back orders?
While on the face value, it might look like a “happy problem” for a manufacturer to have but in reality this creates additional and undue tangible and intangible costs. The tangible cost arises due to additional efforts required to procure raw materials, manufacture and then supply the finished products to customer in due time. A back order may also result in low customer satisfaction levels and in worst cases, shift of customer loyalty to competition. This is an example of an intangible cost.
The cause for a back order could be many, some in control of the manufacturer (unavailability of suppliers, warehouse discrepancies, etc.) and some out of control of the manufacturer (abnormal demand, seasonality, customer’s convenience in placing order, etc.). It is important for a manufacturer to predict back orders to a reasonable degree of accuracy ahead of its occurrence so that it can proactively take measures to optimize supply chain and save on the above-mentioned costs. However, this prediction is difficult to make because of the below listed factors:
- Lack of visibility on key supply – chain indicators which affect back orders
- Presence of multiple product lines and their respective supply-chains makes it difficult to formulate strategies to optimize inventory
- Unavailability/lack of workable data
How machine learning comes to rescue?
Through machine learning classification techniques, supply chain parameters could be analyzed for different product lines. This allows the manufacturers to predict and identify products that would be unavailable in near future and then take the right measures to neutralize the shortage. Additionally, neural networks and deep learning could be used to analyze the supply chain parameters and predict product lines which would experience backorder with high accuracy. These predictions are based on models that improve with time as they get to work on more data. These models could be validated by decision trees and random forest algorithms. The degree of accuracy of these predictive models can by improved by ensemble modelling which uses diverse models and combines them to produce improved results.
What are the challenges involved?
One of the biggest challenges involved in building a machine learning model is the lack of data. For many organizations, back orders is not a frequent phenomenon. This leads to situations where the training data is imbalanced towards a specific class. Thus, techniques like SMOT (Synthetic Minority Oversampling) are used to tackle this imbalance in the data.
Also, it is necessary to note the nature of tradeoffs here: a product unit which is predicted to have a backorder in upcoming weeks, when manufactured would lead to inventory costs. Also, if back order is not predicted with accuracy, that would lead to a lost sale. Hence, profit per unit of a product needs to be plotted against its expected probability of back order to take an informed decision.
While running these predictive models, there’s always a trade-off between precision and recall, that the manufacturer must manage
- Precision determines how accurate the forecast is – predicted back orders vs. actual back orders. The presence of false positives (incorrect prediction) would add to the manufacturer’s inventory costs.
- Recall determines how exhaustive the forecast is – does the model cover all the products that will face back orders? The presence of false negatives would lead to loss of sales or opportunity costs.
Machine learning for a smart manufacturer
In nutshell, applying machine learning to solve supply chain problems in manufacturing domain is essentially a team sport. An optimized predictive model requires a combined expertise of an experienced data scientist team and supply chain experts to clean, prepare the data, train and test the model. By predicting back order requirements with a fair degree of actionable accuracy, organizations can drive better customer satisfaction levels and optimize costs.
