[DSC DACH 24] Reduce waste by state-of-the art Demand Forecasting using AI and Machine Learning - Armin Fanzott

The talk explores how advanced AI and machine learning techniques can significantly enhance demand forecasting accuracy across industries. By leveraging sophisticated algorithms businesses can optimize inventory management, reduce excess production, and minimize waste. The session will cover the latest innovations in predictive analytics, practical implementation strategies, and a case study demonstrating how AI-driven demand forecasting leads to sustainable operations and improved profitability.

1. Reduce waste by state-of-the art Demand Forecasting using AI and Machine Learning DATA & AI

2. 2 Introduction Why accurate and automated demand forecasting is important Out-of-the box solutions often fail to generate accurate forecasts Case-Study: Demand Forecasting @Schrankerl Where to start: Predictability Check Agenda

3. Introduction Speaker Introduction Armin Fanzott ​ ​ Head of AI & Data Science at Ascent.io located in Graz, Austria MSc in Mathematics (Applied Statistics) ​ Extensive experience in analytics and data science in various projects and industries (semiconductors, energy, financial services, pharmaceuticals) ​ Projects always focus on the seamless integration of the implemented solutions into the existing IT system landscape Data & AI Strategy. - Use case generation in data thinking workshops - Development of a roadmap and prioritization of the cases - Requirements engineering DO SOMETHING NEW. DO SOMETHING BETTER. Topics at Ascent Data Science. - Implementation of models based on structured data - Generation of data insights and improved decision-making - Exploratory data analysis AI. - Focus on unstructured data such as text, images or videos - Modern generative AI, NLP or computer vision approaches - Fine-tuning and extension of Transformer - based models

4. Demand Forecasting Why accurate & automated demand forecasting is important Automation. Advanced algorithms can create forecasts for thousands of locations and goods in a matter of seconds. No human with expert knowledge needed. Improved planning. All business processes who rely on forecasts will be enhanced by a higher precision. Reduce waste. Ensure a lower amount of waste by not producing above the demand. Set dynamic prices.  Set prices based on demand prediction to increase revenue or improve waste-rate. Increase revenue. Through precise forecasts we can ensure to meet a customer’s demand. DO SOMETHING NEW. DO SOMETHING BETTER.

5. Demand Forecasting Out-of-the box solutions often fail to generate accurate forecasts 5 Standard solutions don’t differentiate between the goods that are predicted . Therefore, they don’t ”think” about the business 01 They need custom implementations to include external factors (which are in general highly important) 02 They assume easy characteristics of the time series to predict (see image on the right). This is often not the case in the real world 03

6. Demand Forecasting How time series actually look in reality 6

7. Demand Forecasting Or like this 7

8. Case-Study: Demand Forecasting @ Schrankerl Business Challenge Schrankerl is a start-up in the food industry They deliver different kinds of food to fridges at office locations. Their business model is highly dependent on precise forecasts Standard solutions are not suitable because: each fridge has a different consumption behaviour , delivery days and products to deliver. Planning was a manual task of experienced employees to get an appropriate delivery amount and menu selection per fridge and customer. On top some kind of optimization algorithm is needed to combine demand forecasts with food waste goals. Solution Custom solution developed together with Ascent B ased on Machine-Learning based Forecasts combined with mathematical optimization in Python Delivery plans can be derived per fridge in a matter of seconds

9. 9 Predictability Check Is my time-series data predictable?

10. 10 External influences can reduce predictability by introducing noise and variability. Data with consistent historical patterns generally offer better predictability. Data must be relevant for the behaviour of the time-series. Key-factors: Predictability Clean, well-structured and relevant data is crucial for any predictive analysis. Richer data sets and more amount of data can provide more comprehensive insights and patterns and improve predictability. Extensive Data Data Quality Data Consistency External Factors Predictability Future states of a variable can be derived from current and historical data. High predictability means that the results can be predicted with a considerable degree of accuracy. Use of statistical or machine learning models.

11. DO SOMETHING NEW. DO SOMETHING BETTER. Predictability Check: Find out if your data is predictable (fast) Use-Case definition und data acquisition Data transformation and preprocessing Implementation of a simple model and validation Review meeting to discuss results and how to move on

12. Demand Forecasting Conclusion We need to give the knowledge of the human to the model. Creating appropriate external features is extremely important. Predictability checks are a good way to start fast with low risk. Even if the data quality is not good enough the predictability check will show what needs to be improved. Standard solutions are often not good enough for complex processes. Advanced algorithms tailored on the specific data and use-case are needed. Accuracy is not everything.  Getting slightly worst forecasts than human forecasts can still be good enough due to automation. Forecasting is a key technology in almost any business. Deriving accurate forecasts can substantially improve multiple business processes. DO SOMETHING NEW. DO SOMETHING BETTER.