OBER – Optimal inventory planning under uncertainty

Managing inventory levels is a key challenge for SMEs in the wholesale sector. In some cases, forecasting models are already used to predict possible sales. However, these are often not state of the art and also only provide point forecasts: i.e. only a single value is predicted. The uncertainty factor, i.e. the question of how likely it is that the predicted sales volume will be exceeded or fallen short of by a certain value, is either not known or is not taken into account in decisions due to a lack of specialist knowledge. The disadvantage here is that users such as dispatchers must decide for themselves how much they trust the forecast and plan inventory accordingly. This leads to companies holding unnecessarily large safety stocks because they do not correctly assess the risk of out-of-stock costs. AI can help remedy this situation. Together with its project partners, the Center for Applied Research on Supply Chain Services at Fraunhofer IIS is therefore developing and implementing a method that quantifies the uncertainty of a forecast and, based on this, automatically determines the optimal inventory level.

Together with Trevisto, the researchers of the Center for Applied Research on Supply Chain Services at Fraunhofer IIS are exploring two basic approaches for determining the uncertainty of a forecast. The first approach is so-called ensemble forecasts. Here, neural networks, other machine learning methods or modern statistical methods such as ARIMAX are used to generate a large number of different forecasts for the same value. Each individual forecast is referred to as a scenario. These scenarios are combined into an ensemble and their dispersion is used to model the forecast uncertainty. The second approach is to use Bayesian models, where the uncertainty of the forecast is implicitly modeled. Various scenarios can also be generated from this.

The researchers in the Center for Applied Research on Supply Chain Services at Fraunhofer IIS then develop a stochastic optimization model that determines the order quantity that is optimal across all the scenarios considered. For a perfectly fitting solution, both the forecast and the optimization model incorporate the expertise of the wholesale experts involved. To explore the transferability of the method, the developed models are then evaluated on the data of other wholesale companies.

For a further exploitation of the results, especially for small and medium-sized enterprises, the embedding of the process in easily usable software is of crucial importance. For this reason, Trevisto and FIS are developing prototypes of both an independent software solution and the embedding in the SAP landscape in the form of demonstrators and piloting them at the project partner from the wholesale sector, Eisen-Fischer.

• Trevisto AG, Nuremberg
• Eisen-Fischer GmbH, Noerdlingen
  
• FIS Information Systems and Consulting GmbH (FIS), Grafenrheinfeld