Suprise Loss of Multivarite Time Series (SL-MTS)
SL-MTS is a script implemented in R for studying the transitions in multivariate time series data (MTS). The script uses State Space Model (SSM) to model the dynamics of MTS and determine points in the time series where the error in the forecast of SSM model (i.e. out-sample error) is worse than its in-sample performance, where the performance is measured for a fixed loss function in a rolling window manner. The error is called Surprise Loss (SL)1. Relatively High SLs can be an indicator of transition in the system. The script is implemented on Intensive Care Unit (ICU) MTS data for studying the transition to Septic shock in the ICU setting.
1 Giacomini, R., & Rossi, B. (2009). Detecting and predicting forecast breakdowns. Review of Economic Studies, 76(2), 669–705.
How do I use SL-MTS?
The input data format is a list, where each component of the list is multivariate time series in form of a data-frame (e.g. data pertaining to a single patient). The rows and columns of this data-frame denote equally spaced time points and variables respectively. The calculation of SL can be parallelized over multiple cores as well.
| Arguments | |
|---|---|
mts_data |
Multivariate Time Series Data [List format] |
num_trends |
Number of hidden trends in SSM |
rolling_window_size |
Length of rolling window [hours] |
bin_size |
Length of intervals between time points [minutes] |
num_cores |
Number of cores for parallel computing |
Examples
result = computeSL(
mts_data = ICU_data,
num_trends = 3,
rolling_window_size = 18,
bin_size = 30,
num_cores = 10
)
OS Compatibility
SL-MTS has been tested in the MacOS.
Dependencies
The following softwares need to be pre-installed before before running this program:
- R packages:
License
SL-MTS is an open source software and is licensed under LGPL.
Getting help
For queries regarding the software write to: samal@combine.rwth-aachen.de , farhadi@combine.rwth-aachen
Citing SL-MTS