Beschreibung
Learn relevant modeling and cutting-edge techniques for time series analysis and prediction and apply them in a capstone project.
Learning Goals:
- Understand "classical" models from SARIMA(X) and GARCH to state-space models and filtering.
- Analyze dependence concepts like copulas, gaussian processes, random matrix theory and graphs.
- Study extremes, anomalies and signatures of data-streams.
- Apply tree-based models like XGBoost, LightGBM and CatBoost.
- Use recurrent architectures (xLSTM, GRU and LMU), VAEs and GANs for time series problems.
- Gain experience with transformers, TFTs, NBEATS(x), NHITS and recent LLMs for time series data.
What you get:
Per session we will cover one topic or concept:
First we motivate statistical and methodological concepts. Second we discuss ML-techniques to analyze and forecast time series. Break-out rooms are regularly added to discuss group projects
We start with SARIMA(X) and GARCH models (also in a multivariate settings) as an advanced "classic" approach and move on to state-space models and filtering techniques to study time-series data.
The next step on our journey are dependence concepts like graphs, copulas, gaussian processes and random matrix theory to deepen our knowledge in general multivariate settings.
In the next sessions we focus on extremes (i.e. extrem value theory) and anomalies and highlight signature methods for time series data.
Our "parallel" stream on ML techniques will start from recent tree-based methods like XGBoost, LightGBM and CatBoost as highly valuable tools in high dimensional settings.
You will become familiar with deep learning sequence models like xLSTMs, auto-encoders, GANs, Transformers, TFTs, NBEATS, NHITS plus recent LLM-models and how to apply them.
Everything what is on the market right now - we will try to cover it ;-)
The hands-on portion of this course focuses on using best practices and testing assumptions derived from statistical learning.
This course is mainly based on various online material and references.
The course instruction will be online via Zoom.
For online participation, it is expected that you turn on your camera and have a sufficient internet connection.
All required software and online course content is free.
This course is part of the opencampus.sh "Machine Learning Degree".
These participants will receive preferred access to this course.
What you should bring:
To get the most out of this course, you should be familiar with programming in Python and have a fundamental understanding of basic statistics, EDA, solid machine learning and time-series skills.
These courses or comparable knowledge are required to succeed and get approval for the "Advanced Time Series Prediction"course:
-Introduction to data science and machine learning -Machine learning with Tensorflow -Basic time series analysis skills
So please provide these necessary information to your application
During the week you will be expected to work through the assigned online course content and the weekly exercises.
If you plan to use your own dataset, you should bring it to the introductory session to clarify whether it is suitable for the tasks throughout the course.
We expect to have about 4-6 groups with 3-5 people per group.
Potential capstone projects and working datasets are:
Financial/Economic data: YahooFinance, FRED data...
Energy Data: Electricity Consumption, Heat Load...
Environmental data: Climate forecasts, Water level prediction (Floods)...
Health Data: Mortality, Heartbeats...
Engineering data: (Audio) Signals, default rates...
You should plan to spend at least 10 hours per week working through the learning content and completing the assignments.
Additional time will be needed to prepare the final project presentation.
Allocating sufficient time is therefore critical.
To receive a certificate of achievement (Leistungszertifikat/ ECTS) for this course, active participation is expected, no more than two classes may be missed, and you have to conduct an applied capstone project in a team. At the end of the course the project has to be presented and a well-documented project source code (GitHub) has to be submitted.