Abstract
Support Vector Machines are a successful machine-learning algorithm used for classication, regression and prediction of time series. We optimize learning parameters and select the feature set with the smallest predictive error.
We also explore the development of errors for predictions with larger time skips. We've applied the method for the prediction of CISS (Composite Indicator of Systemic Stress), a stability indicator created by the European Central Bank.
We've chosen this indicator among other state-of-the-art indicators because of its high frequency, which indicates a quick response to distinctive changes on the market. The results show that CISS can be partially explained just by its past values up to six to eight weeks ahead, but large behaviour shifts are still surprising for the model.
We've also discovered that including data over three months of age into the prediction context won't improve the results.