Weiter geht's mit vier parallelen Hands-on-Session. Wir vom Live-Blogging-Team entscheiden uns für die Session von Umer Khan (Data Scientist, Artificial Intelligence and Robotics Lab, Continental Automotive) unter dem Titel "Recurrent Neural Networks for Customer Demand Forecast."

Da die Session auf Englisch ist, werden auch wir im Blog hier für die nächste halbe Stunde die Sprache wechseln.

Paul Knecht

Umer Khan talks about forecasting the demand of the customers for automotive suppliers.
These suppliers are big companies who themselves have huge supply chains and they already are very well advanced in this area, as it’s their main business – this is what makes money.

But there are numerous reasons to further improve the forecast.
Depending on the customer demand, the production is scheduled, raw materials are sourced from suppliers and the budget is planned.

If you overestimate the demand, companies are faced with storage or scrap costs.
Underestimation leads to premium freight costs or even contract penalties.

Paul Knecht

Khan explains what kind of data pre-processing they went through, when realizing a forecast project at Continental Automotive.

First they went for a log transformation plus some deseasonalizations:
„You can do this, so your model does not get trapped in some noise or seasonal changes.“

Later on, the team of Umer Khan tried to exponential smooth the curves/data. This has lead to an unexpected well improvement as it also flattens out some unusual peaks.

Paul Knecht