Visual Analytics Methods to Guide Diagnostics for Time Series Model Predictions

Abstract

Visual Analytics methods are used to guide domain experts in the task of model selection through an interactive visual exploration environment with short feedback cycles. Evaluation showed the benefits of this approach. However, experts also expressed the demand for prediction capabilities as being already important during the model selection process. Furthermore, good model candidates might show only small variations in the information criteria and structures which are not easily recognizable in the residual plots. To achieve this, we propose TiMoVA-Predict to close the gap and to support different types of predictions with a Visual Analytics approach. Providing prediction capabilities in addition to the information criteria and the residual plots, allows for interactively assessing the predictions during the model selection process via an visual exploration environment.