(Bachelor/Master/Praktikum) Analyzing Benchmark Datasets in Time Series Forecasting

Motivation

The choice of benchmark datasets in time series forecasting is crucial because it directly influences the validity, comparability, and generalizability of research findings. Well-curated and diverse benchmarks ensure that models are evaluated under realistic and varied conditions, reflecting real-world challenges such as seasonality, noise, and missing data. Without standardized, representative benchmarks, model performance claims may be misleading or limited in scope. Therefore, careful selection and use of benchmark datasets serve as a foundation for meaningful progress and fair comparison in the field of time series forecasting.

Goals

The abstract goal of this work is analyzing the validity, quality and representativeness of benchmark datasets for time series forecasting (and optionally imputation, i.e., the inference of missing values). Typically, datasets are (seemingly) arbitrarily chosen and never questioned. Once established, subsequent researchers reuse them. While consistency is a crucial aspect of benchmarking, using adequate data for evaluation is as well. In this work, determining this will be your job. Potential more specific research questions include but are not limited to:

• What (combinations of) datasets are used in evaluating forecasting models?
• Are these individually valid or do they contain many artifacts?
• Which combinations are representative?
• Which other datasets are better suited for the task?
• What are the correlations between channels?

Wir bieten

• Arbeit an innovativen und neuen Forschungsfeldern
• Einbringung eigener Ideen bei der Lösungsgestaltung
• Gute und intensive Betreuung

Dauer
   3 - 6 Monate

Kontakt

Michael Stenger, M.Sc.

michael.stenger@uni-wuerzburg.de

go.uniwue.de/stenger