Evaluating Statistical Models for Time Series Analysis
In statistics, a wide range of different models for time series exist. Our focus are those that also support the cyclic structure of time granularities. However, it is not clear which models are more suitable for which kinds of data. The effect of the chosen granularity also is a central factor.
A statistical model for a time series can have many forms, but it is often based on some kind of function that transforms time into an estimated value. An important method to evaluate such models is the comparison of residues between the estimated and actual value.
Granularities are a way to describe how time can be grasped according to structures of a calendar. The finest unit of a dataset which is rastered into the discrete time domain is called chronons. These chronons can be mapped into different granularities, i.e. days can be mapped to weeks. Granularities can be mapped to other granularities, i.e. weeks to fortnights.
The overall aim is to support the selection and assessment statistical models for time series analysis, which can be divided in the following sub-goals
- Researching the state of the art in statistical models for time series analysis with focus on models that support granularity cycles.
- Implement state of the art models using the statistical computing environment R (or reuse existing R implementations).
- Evaluation of the state of the art models with example data sets.
- (optionally) Visualization of the results, i.e. the residues between a given dataset and the interpolation from the various models and their various configurations using a pre-existing external framework.
This project is organized by the Information Engineering Group and the Department of Statistics and Probability Theory, working together on the research project HypoVis. The mentoring regarding statistical models will be performed by Ao. Univ.-Prof. Dipl.-Ing. Dr.techn. Peter Filzmoser.