Hey there!

I'm Kai Reffert – M.Sc. in Data Science | Applied ML Developer | Time Series Enthusiast

🎓 About Me

I'm currently pursuing my Master's in Data Science at the University of Mannheim, with a strong focus on time series forecasting, machine learning, and applied data science. My journey includes building real-time forecasting pipelines, developing interpretable ML models, and working on impactful data-driven projects in both academic and industry settings.

I’m driven by curiosity, problem-solving, and making data useful — whether it's through forecasting infectious diseases or analyzing football statistics.

📂 Selected Academic Projects

Here’s a selection of my academic data science projects (B-Bachelor and M-Master):

• 🧪 M-Master’s Thesis (30 ECTS): Proabilistic LTSF – [Github] [📄 Thesis PDF]
  Explored cutting-edge challenges in Time Series Forecasting (TSF), focusing on two key dimensions:

  1. IMS vs. DMS trade-off in probabilistic TSF: Demonstrated that Direct Multi-Step (DMS) models often produce unreliable sample paths due to their failure to model temporal dependencies between future predictions.
  2. Bridging probabilistic TSF with point long-term TSF: Combined two active TSF research streams (probabilistic TSF and point LTSF), revealing strong performance from quantile-based methods and highlighting a promising, underexplored research direction.

• 🕵️ M-Team Project (12 ECTS): Explainable Deep Fake Detection – [GitHub]
  Over a 6-month period, our team of six master's students investigated the potential of B-Cos networks for deepfake detection. Our key contributions include:

  1. Introducing B-Cos networks into the domain of deepfake detection by integrating them into the open-source DeepFake Benchmark, thereby enabling further research in this area. Moreover, we also developed the first B-Cos-enhanced version of XceptionNet by applying the B-Cos transformation.
  2. Proposing the novel Mask Pointing Game, a new method for evaluating explanation quality that advances research on inherently interpretable models within facial forensics. Our qualitative and quantitative comparisons demonstrated that ante-hoc B-Cos models outperform state-of-the-art post-hoc explainable AI techniques.

• ✈️ M-Data Mining Project (6 ECTS): Flight Price Prediction – [GitHub]
  Modeled flight prices using XGBoost, RandomForest, and SVMs.

• ⚽ M-Web Data Integration Project (3 ECTS): Football Data Integration – [GitHub] [📄 Report PDF]
  Integrated data from multiple football sources using web scraping and schema matching in Java.

• 🦠 B-Bachelor's Thesis (12 ECTS): DeepAR for Infectious Disease Forecasting – [GitHub] [📄 Thesis PDF] [📑 LNI Paper]
  Introduced the application of DeepAR for probabilistic influenza forecasting by addressing the underexplored gap of global deep learning models in epidemiology, showing that DeepAR significantly outperforms traditional and neural baselines in accuracy and uncertainty quantification — offering a scalable, data-efficient solution for public health forecasting.

• 📊 B-Bachelor Seminar (4 ECTS): Regression Feature Selection – [GitHub] [📄 Seminar Report PDF]
  Compared LASSO and Ridge for feature selection in regression tasks.

🧠 Professional Experience

• ICIS (Internship): Developed a binary classification pipeline for forecasting EU power market price equality. Designed and implemented an OTC backtesting dashboard.
• Karlsruhe Institute of Technology (Student Research Assistant): Built a DeepAR-based probabilistic forecasting pipeline for infectious disease spread.
• CompuGroup Medical – STADS Datathon (Winner): Designed a data-driven simulation to support targeted vaccination strategies using interactive dashboards.

🛠️ Tools & Skills

• Languages: Python, R, Java, SQL
• Libraries: PyTorch, pandas, NumPy, scikit-learn, XGBoost
• Other: Git, Docker, Web Scraping, Jupyter, MLflow

   

📫 Let’s Connect

You can reach me on LinkedIn or check out my repositories right here on GitHub.
I'm always happy to chat about data, research, or collaboration opportunities!