Ask our experts

There are many valuable applications for data science in supply chain management, ranging from supporting management decisions on strategic issues to finding ways to optimize your daily operations. The possibilities are endless, though it is crucial to have access to sufficient useful data.

Many of OMP’s solver implementations deploy a strategy of optimizing the sequence of production campaigns or jobs. Our experts have acquired in-depth experience on this complex topic over the years. OMP implementations often involve configuring product wheel constraints, with a focus on the areas where they bring the most value to the overall planning process. Examples of value brought by product wheel implementations include minimizing setup times between jobs in the scheduling process, optimizing campaign sequences in operational supply planning, and grouping similar requirements for sourcing in S&OP.

In general, data science is about deriving actionable insights from data. While data analytics are carried out in a very focused way, answering specific questions about your current way of working, data science has a broader scope and can expose new insights that can lead to opportunities you probably did not consider before.

OMP has an in-house team of highly skilled PhD-level data scientists. They stay tuned to the latest research and apply their expertise as much to customer projects as R&D. Hence, we don’t just supply the technology, we also supply the data-driven insights you need to make the right decisions.

By nature, different business functions have their own purposes, and hence would require their own specific optimization solutions. However, OMP makes sure that data and decisions flow seamlessly between these planning functions, taking care of the mutual constraints this brings with it. For example, the S&OP solver could make decisions about sourcing from alternative plants, the quantity of raw material to be purchased, or the safety stock to be built up ready for future demand. In OMP, these decisions are translated into constraints to be applied to the solvers on the operational side.

Our team of data scientists can carry out a forecastability analysis on your data. They will evaluate the accuracy of the statistical forecast you will be able to achieve in a real-world situation. The results are displayed per product-market segment, along with the expected added value.

An expert system is an algorithm that emulates human decision behavior. It contains a set of rules used in making a decision. For example, OMP contains an expert system to decide which time series forecasting model would be the most suitable for the given time series being modeled, for example exponential smoothing, ETS, or linear regression.

We either use real costs (sales price, raw material price, production and delivery costs, etc.) or standard penalties (linked to safety stock deviations, capacity overloads, etc.), or even a mix of both, depending on the business at hand. The way we tune these costs depends on the business strategy (for example whether a new batch would require the stock to be kept below the safety level or above it) and the product categories (for example high-runners versus low-runners). OMP offers various ways to visualize the KPIs and drill down the related conflicts, while providing appropriate reporting to your organization.

We have a simple rule of thumb to assure robustness: we always model the cause, not the effect. This ensures that, should some master data change, the solver will automatically adapt appropriately. With respect to plan stability, OMP’s incremental techniques minimize changes in the plan while solving local conflicts. Even in our holistic mathematical programming solvers, we have embedded logic that can attempt to create a newly optimized solution from the current plan.

Machine-learning algorithms also make decisions, just like an expert system, but the decision rules are derived from training data and hence more adaptive.

One way to test for overfitting is to split your dataset into a training set and a validation set. You can train your model on the training set and subsequently validate it on the validation set. If the model is significantly less accurate using the validation set than the training set, your model is likely to be overfitted and you should develop a simpler model.

OMP provides standardized solutions for cutting, based on many years of field experience in the paper, film and packaging industries. Based on OMP’s proprietary optimization engine, they select the best set of cutting patterns to efficiently produce the required mix of orders of various sizes while minimizing scraps and trims.

The first step is to identify a problem where human decision-making could hypothetically be enhanced or even replaced by computer algorithms. The next step is to assess whether there is sufficient data from which the algorithms can ‘learn’, and then to select and train a suitable machine-learning model. Finally, it is important to validate your model, compare it with previous human decision-making, and quantify the benefits before putting it in practice.

We know this is a challenge. It is hard to make fundamental changes to the way people have been working for years. This requires change management, where key success factors include communicating the vision very clearly and focusing on the value to the company. Our data scientists can assist you to quantify this value. Machine learning cannot replace all human decision-making. It can, however, take over the more mundane tasks and free up time for people to focus on more complicated planning issues.

In OMP, we use a variety of functionalities to globally optimize the supply chain. For example, we could use an MRP-like solver to pull requirements upstream, a mathematical programming solver to optimize the production plan, and an allocation solver to deploy the upstream stock downstream. Alternatively, we could use a holistic all-in-one solver, perhaps in combination with a neighborhood search solver to incrementally repair any conflicts arising during the day.

We use all types of technologies without prejudice, including greedy constructive heuristics, meta-heuristics, branch & cut, etc. The mix of technologies used depends on the project at hand, while our proprietary hybrid solver platform allows technologies to be combined. For example, we can easily build an automated chain from a mathematical programming solver (to devise an optimized bucketized plan in volume planning), and a neighborhood search solver (to optimally sequence part of the Gantt plan) running on the same planning role.

OMP offers several techniques and tools for this. For example, it includes product segmentation functionality to demonstrate products that can be handled with greater accuracy by machines than with human input. It is best to leave these products undisturbed. The tool also highlights the products requiring human attention because of the machine’s uncertainty. In addition, automatic alerts are available to guide you whenever a machine output diverges too far from reality.

In the absence of capacity constraints, the OMP solvers will automatically generate a plan at an ideal production frequency, for example with inventory levels oscillating between safety stock and maximum stock. Where there are capacity constraints, our solvers propose a weighted compromise between this ideal production frequency (minimizing incurred setup costs related to production changes), and the ideal allocation of capacity to requirements (keeping inventory levels on safety stock all the time). Users can define the weighting factors to be used in these exercises.

A wide variety of external data has been proven to boost predictive performance in different industries. Examples include economic indicators, weather data, and calendar data. However, success is not guaranteed. It ultimately depends on whether significant correlations are present in the data. Our data scientists can quantify the benefit to be gained from using external factors to help you decide whether to implement the technology.

Shelf-life constraints are common, especially in the consumer goods and life sciences industries. OMP takes them into account in various ways at different stages in the planning cycle, depending on the business at hand. This functionality is integrated into a range of solvers, including Material Requirements Planning (MRP), mathematical programming and allocation solvers.

Overfitting is a very common problem in data science. It occurs when a model is too complex, i.e. when it contains too many parameters relative to the number of observations in your data set. An overfitted model describes random variations or ‘noise’ in the data rather than the real underlying patterns. Overfitting leads to low predictive performance.

Point-of-sale (POS) data can be hugely beneficial to sense what is currently going on in the market, allowing you to anticipate the effects on the upstream supply chain more quickly and accurately. You can use this to your advantage to reduce the well-known bullwhip effect in long supply chains. A clear overview of POS data also allows you to manage new product introductions better and respond more appropriately to events or disruptions in the supply chain.

One potentially very powerful way to improve short-term forecast accuracy is to implement demand sensing technology that looks at the order book or other internal or external causal factors. If there are significant predictive patterns in these data, a carefully implemented machine-learning model can pick this up and generate more accurate forecasts than traditional models can achieve.

Data science is not just analytics. A data science exercise encourages you to take a fresh view of your business and ask questions that you may not have thought of before. This opens the door to opportunities that you could take along in the project design.