Max Bömer
Max Bömer

Reputation: 1

Model drift for VARIMA model when forecasting multivariate time series

I am currently trying to train a VARIMA model on multivariate time series data about 5 different kinds of sensor measurements of a cooling system. The data is of cyclical nature so the exact same pattern keeps reoccurring every 25 datapoints or so. I have a dataset of the 5 different components where there is a datapoint for every minute. I used 2 weeks of data to train the model on and let it synthesize data after. The VARIMA model I use is imported from the Darts package and i defined the model like this: model_VARIMA = VARIMA(p=12, d=0, q=0, trend="n"). The model was trained on 2 weeks of training data. When predicting 30 points in the future or more the predictions clearly starts to show model drift. Instead of producing cyclical multivariate time series data, all components slowly start to follow a straight horizontal line that doesn't change in value anymore. In other words the standard deviation of all components slowly become zero and all components slowly start to drift towards their mean value and stay there forever. I was wondering whether someone has an explanation for this and a way to solve the issue. Thanks!

Upvotes: 0

Views: 149

Answers (0)

Related Questions