Visualizing Forecasts of Neural Network Ensembles

Abstract

Advanced neural network architectures like, e.g., Historically Consistent Neural Networks (HCNN) offer a host of information. HCNN produce distributions of multi step, multi asset forecasts. Exploiting the entire informational content of these forecasts is difficult for users because of the sheer amount of numbers. To alleviate this problem often some kind of aggregation, e.g., the ensemble mean is used. With a prototypical visualization environment we show that this might lead to loss of important information. It is common to simply plot every possible path. However, this approach does not scale well. It becomes unwieldy when the ensemble includes several hundred members.We use heat map style visualization to grasp distributional features and are able to visually extract forecast features. Heatmap style visualization shows clearly when ensembles split into different paths. This can make the forecast mean a bad representative of these multi modal forecast distributions. Our approach also allows to visualize forecast uncertainty. The results indicate that forecast uncertainty does not necessarily increase significantly for future time steps.