Learning Discrete Temporal Patterns for Time Series Forecasting

Research topic/area
Time Series Forecasting with Deep Learning
Type of thesis
Bachelor / Master
Start time
05.06.2025
Application deadline
31.07.2025
Duration of the thesis
6 months

Description

Background

Traditional deep learning models for time series (e.g., LSTM, Transformer) often struggle with noisy, redundant, or high-dimensional input signals. Inspired by advances in sequence modeling, this project explores a novel intermediate representation to improve forecasting performance and interpretability.

Core Idea

The thesis investigates a two-stage approach where time series data are first discretized into a learned symbolic form, followed by a sequence model trained on this compact representation. This abstraction allows the model to focus on recurring temporal motifs rather than raw data.

Why It’s Exciting
  • * New representation: Extract and operate on high-level temporal units.
  • * Modular & extensible: Encourages transfer learning and hybrid architectures.
  • * Real-world impact: Applicable to scenarios with noise, missing data, or limited labels.
  • * Evaluation: Compare against existing state-of-the-art on standard forecasting benchmarks.

Learning Outcomes
  • * Implement unsupervised sequence compression techniques for time series.
  • * Apply sequence models on symbolic or latent representations.
  • * Conduct rigorous benchmarking and performance analysis.
  • * Investigate interpretability and robustness in challenging environments.

Stretch Goals (Optional)
  • * Study latent attention patterns and temporal abstraction.
  • * Experiment with self-supervised objectives for time series.
  • * Apply the model in domains such as energy, finance, or scientific sensor data.


Requirement

Requirements for students
  • Solid programming skills in Python
  • Basic knowledge of machine learning
  • Initial experience with deep learning (e.g., PyTorch or TensorFlow)
  • Interest in time series analysis and modeling
  • Willingness to engage with current research literature
  • Good understanding of mathematics (especially linear algebra and statistics)
  • Beneficial: Experience with autoencoders or transformer models

Faculty departments
  • Engineering sciences
    Electrical engineering & information technologies
    Informatics
    Energy Engineering and Management
    Financial Engineering
    Information System Engineering and Management


Supervision

Title, first name, last name
Dr-Ing., Nicholas, Tan Jerome
Organizational unit
Institute for Data Processing and Electronics
Email address
nicholas.tanjerome@kit.edu
Link to personal homepage/personal page
Website

Application via email

Application documents
  • Cover letter
  • Curriculum vitae
  • Grade transcript
  • Certificate of enrollment

E-Mail Address for application
Senden Sie die oben genannten Bewerbungsunterlagen bitte per Mail an nicholas.tanjerome@kit.edu


Back