Porsche Holding Demand Prediction Dashboard

Jul 1, 2016
Results of LSTM and SARIMA models predicting monthly car orders.

Project Overview (July 2016 – July 2018)

At Know Center Research, we collaborated with Porsche Holding GmbH (Porsche Holding Salzburg), Europe’s largest automotive distributor, representing Volkswagen Group brands in wholesale, retail, and after-sales services. Porsche sought an accurate time-series prediction solution to optimize monthly car orders based on historical sales data.

Technical Challenges

Key challenges faced in this project included:

  • Handling large and varied time-series datasets
  • Performing thorough data cleaning and exploratory analysis
  • Conducting comprehensive hyperparameter optimization
  • Implementing accurate linear and non-linear predictive models

Technologies & Methods

The solution was built using advanced data analytics and visualization tools:

  • Machine Learning & Analysis: Python, R, ARIMA, SARIMA, LSTM neural networks (Keras)
  • Data Visualization & Dashboard: Interactive dashboards using Plotly
  • Data Exploration & Cleaning: Extensive data preprocessing, feature engineering, and exploratory data analysis (EDA)
  • Research & Publication: Scientific publication highlighting methodology and results

Results & Impact

  • Delivered an end-to-end predictive analytics dashboard integrated into Porsche Holding’s operational processes
  • Improved accuracy in predicting monthly car order requirements
  • Recognized internally with the company’s Project Excellence Silver Award
  • Successfully collaborated on a research publication titled “Gone in 30 days! Predictions for car import planning” in the journal Information Technology
  • Featured as a success story in Austria’s Trend magazine

Porsche Project Featured in Trend Magazine

Personal Contribution

As a Software Engineer, ML Engineer, and Data Scientist, my responsibilities included:

  • Developing the complete data processing and cleaning pipeline
  • Conducting extensive data exploration and feature engineering
  • Researching and implementing linear and non-linear time-series predictive models (ARIMA, SARIMA, LSTM)
  • Creating interactive visualizations and predictive analytics dashboards with Plotly
  • Co-authoring the research publication and presenting
Machine Learning Time-Series Analysis Data Visualization Dashboard Development Predictive Analytics Automotive
Tomislav Đuričić
Tomislav Đuričić
Researcher / Machine Learning Engineer / Software Engineer

My research interests include social-based recommender systems, graph neural networks and user modeling.

Publications

Gone in 30 days! Predictions for car import planning
A challenge for importers in the automobile industry is adjusting to rapidly changing market demands. In this work, we describe a practical study of car import planning based on the monthly car registrations in Austria. We model the task as a data driven forecasting problem and we implement four different prediction approaches. One utilizes a seasonal ARIMA model, while the other is based on LSTM-RNN and both compared to a linear and seasonal baselines. In our experiments, we evaluate the 33 different brands by predicting the number of registrations for the next month and for the year to come.
Emanuel Lacić, Matthias Traub, Tomislav Đuričić, Eva Haslauer, Elisabeth Lex
PDF Cite Project DOI
Gone in 30 days! Predictions for car import planning