Although machine learning is now mainstream, the technology still puzzles many supply chain professionals. According to Gartner, “While machine learning promises transformative benefits for the supply chain, current expectations of near-term readiness and benefits remain unrealistic.” It is therefore important to look beyond the hype and examine how machine learning can help solve business problems and Value creation can be used in the supply chain.
Companies are advised not to rush into machine learning projects blindly; To achieve the best results, you should allow enough time for thorough consideration and preparation in advance.
Every machine learning project should start with a clear description of the goal. This description should include key figures and data on the initial situation so that a clear before and after comparison is possible. Since machine learning systems continually learn, it is possible to show how the systems improve over time and the company benefits from the project. Long-term, sustainable success can only occur if the machine learning project is built on a stable foundation. To do this, it is recommended to combine machine learning technologies with an adaptive probabilistic forecasting approach.
A short excursion into probabilistic prediction
The opposite of probabilistic prediction is deterministic prediction. This method calculates a future outcome using exact numbers, often based on historical averages. In contrast, probabilistic forecasting works with probability distributions rather than exact numbers. While a deterministic forecast is generally expressed as a series of time series of exact numbers, a probabilistic forecast is expressed as a series of time series of probability distributions. Supply chain planning uses advanced algorithms to analyze multiple demand variables to identify the probabilities of a range of possible outcomes.
The benefit of the probabilistic approach is that it can distinguish between error and natural variability and between signal and noise, which is impossible with the deterministic view. This results in three main consequences:
First, it is impossible to accurately determine risks and opportunities based on deterministic plans and forecasts;
secondly, it is impossible to deterministically correctly judge how good or bad a plan or prediction is and
third, it is impossible to determine with any degree of accuracy where improvement efforts should focus based on deterministic plans or forecasts.
Die probabilistische Vorhersage hingegen liefert reichhaltige Informationen zur Ermittlung von Risiken und Chancen auf allen Detailebenen, sodass fundierte Geschäftsentscheidungen getroffen werden können. Sie ermöglicht auch eine perfekte Abgrenzung zwischen den Dingen, die sich kontrollieren und verbessern lassen, und jenen, die außerhalb unserer Kontrolle liegen.
Combining probabilistic prediction with machine learning
The combination of probabilistic forecasting and machine learning allows supply chain planners and dispatchers to create forecasts at a granular level and for different time horizons. Here, the adaptive probabilistic forecast model is first created using historical data. Once the model is established, machine learning algorithms are applied to improve the prediction probability, with both internal (such as product characteristics or other master data) and external data sets (including weather data, economic indicators or social media data) being gradually added. This step-by-step approach makes it easier to check the performance of the model and adjust it if necessary.
What should be taken into account with the data
Clearly, machine learning projects benefit from large amounts of data. The larger the amount of data, the more precise the statistical significance of a machine learning model. To get started, however, smaller, traditional data sets such as the history of a product are often sufficient. The granularity of the data sets is also important. Unlike traditional analytics approaches that aggregate the data to filter out the noise, machine learning analyzes that very noise to find correlations that train the model and make it more powerful.
As with any aspect where data plays a fundamental role, data quality is also an important aspect of machine learning projects. The tools used should therefore definitely have governance functions to maintain information quality throughout the data lifecycle.
Operationalize the results
While it's tempting to develop a machine learning model to solve a unique business challenge, it's not efficient. One-off projects create “black boxes” that only the programmer truly understands and that business users distrust. To achieve sustainable business value and get the best return from the project, the results should be operationalized. Therefore, it is important to use adaptive models that do not require constant manual adjustment, otherwise changing business conditions will make the models unreliable.
Application examples from supply chain planning
Demand processes such as demand forecasting, demand sensing and demand shaping are particularly suitable for the application of machine learning due to their complexity and fast-moving nature. The most popular application examples include:
Seasonality: Clustering and classification of multiple seasonal patterns (day-in-week, week-in-month, month-in-year)
Sales Promotion: Clustering of past promotions, classification of new promotions based on attributes and uplift calculation
New product launches: Clustering past launch profiles, classifying new items based on their attributes, and regression to create baseline forecasts
POS Demand Capture: Advanced techniques to improve sell-in forecasting using sell-out demand data
External demand conditions: weather, social media, IoT, market trends, indicators and other external data
Product lifecycle management: Algorithms weigh attributes and sales of similar items to estimate the shape and length of the product lifecycle
The limits of machine learning
Like all technologies, machine learning also has its limitations. That's why employees' business knowledge and process knowledge play an important role in coordinating machine learning models and evaluating the results. Because the algorithms relieve them of boring, repetitive tasks, supply chain planners can concentrate on new, strategic tasks.
In a guide to developing future supply chain professionals in the digital age, Gartner cites business acumen, adaptability, political savvy and the ability to collaborate as key
Improving “digital dexterity”. This highlights how important it is for digital supply chain organizations to focus on the human side of supply chain planning now that many processes are being automated through machine learning. That's why supply chain professionals who develop their skills in negotiation, business communication and simplifying complex data are becoming increasingly valuable to companies.