Hello all, In my research to do a anomaly detection in multivariate time series data, the approach that i am choosing involves the prediction of next step from the target variable and simultaneously flagging the anomaly point or sequence within the data. What i am thinking of currently, involves the modeling of features with Vector autoregression (VAR) model to predict next target feature value and simultaneously fitting a multivariate Gaussian distribution for detect anomaly.
As the data is unsupervised, the scope of training a classifier requires labels. Can we modify DBSCAN for multivariate time series data? or what are some of the ongoing research topic tto improve VAR performance? Or which model do you think is good in my case wrt to time series data?
Hi. You should consider tools for imposing linear restrictions. These tools permit targeted restrictions on parameters, allowing incorporating information about the parameters of the model which does not fit into the VAR lag framework. Maybe you should try with the structural VAR (SVAR) model which uses additional identifying restrictions and estimations of structural matrices to transform VAR errors into uncorrelated structural shocks. In addition you could make dynamic forecasts directly from the VAR object. Hope it helps. Good luck!
See:
Barbaglia et al, 2019. Volatility spillovers in commodity markets: A large t-vector autoregressive approach. Energy Economics, 85.
The authors used an innovative approach by using a lasso estimator through the multivariate t distribution. This approach addresses some of the issues in the standard estimation procedure of the VAR model.
Hi, some time ago I faced a problem that may be considered similar to the one you describe. There, I had to modify multivariate observations that showed discrepancies with the predictions produced by a VAR model.
A collegaue and I wrote the following paper about that situation: Guerrero, V. and Gaspar, B. (2010) Edition and Imputation of Multiple Time Series Data Generated by Repetitive Surveys. Journal of Data Science, Vol. 8, pp. 555-577.
I hope this reference may be useful to you. Regards.
- Badges
- Science topic
- Similar topics
- Organizational Psychology
- Training