AI-based demand forecasts for logistics, trade and production

The expected customer demand is an essential factor for companies when it comes to the correct planning of e.g. production processes or transports. And the better these forecasts, the better one can manage one's own resources - which in turn is not only interesting from a financial and organisational point of view, but also for sustainability reasons.

Demand forecasts of various kinds are already used today in many corporate areas, for example in inventory management, in staff scheduling or in personnel demand planning. Data-driven methods for forecasting demand have become established here in the last two decades. In this context, statistical, univariate methods such as exponential smoothing or ARIMA models were superior to all other methods in terms of forecasting accuracy for a long time. Now, however, the development of software, hardware and algorithms has advanced so far that forecasting methods based on machine learning (ML) have recently significantly surpassed the solutions of statistical benchmarks.

There is still a lot of research and adaptation work to be done in this field, especially when it comes to actually transferring the theoretical models - as desired in our research field - into concrete applications. That is why an interdisciplinary team from the fields of mathematics, statistics and engineering is working on ML-based demand forecasts for logistics, trade and production in order to be able to provide appropriate software solutions for the concrete requirements there.

For this, on the one hand, the evaluation and benchmarking of the developed methods is important; the central question is: Do complex methods actually also provide a higher forecast accuracy in the present application case? For this purpose, the forecast must be compared with the actual development on the basis of historical data and the forecast error calculated from this. This can then be used to evaluate different forecast models. On the other hand, knowledge about the associated processes and levels of a forecast can also help to improve the forecast. For this purpose, our researchers look for structures in concrete demand forecasting questions that translate the real world into mathematics, so that they can be described in mathematical models and then implemented in algorithms. This enables coordinated and accurate forecasts.

From forecasting to decision support

Furthermore, the aim is not only to make point forecasts, but to quantify forecast uncertainties. For this purpose, different neural network architectures are used for time series forecasting and cross learning is applied, i.e. learning in a model on the basis of many time series at the same time. With the help of stochastic optimization, different predicted scenarios can be optimized simultaneously so that the best decision is found in relation to the expected value. This means that sales forecasts can also be linked to the resulting decision, so that concrete, operative decision support systems are created with this method, which companies can in turn integrate into their process infrastructures for decision-making - either automatically or by expert decision.

With this understanding in mind, the scientists in the Center for Applied Research on Supply Chain Services are researching to bring AI-based demand forecasts into application in logistics, trade and production.