A0692
Title: A convolutional approach to forecast reconciliation
Authors: Andrea Marcocchia - Sapienza University of Rome (Italy) [presenting]
Serena Arima - University of Salento (Italy)
Pierpaolo Brutti - University of Rome - Sapienza (Italy)
Abstract: The goal of forecast reconciliation in a hierarchy is that observed responses/demands at each level will always add up to the observed responses/demands at higher levels. In the literature, there are numerous approaches that try to make predictions coherent: the challenge is to exploit the predictive power of the most aggregated data to benefit from it on the most granular data. The idea is to exploit the convolutions of Neural Networks to aggregate information at different levels of a hierarchy to obtain consistent and accurate predictions. In particular, we propose to generalize the convolutional approach to work in cases where more than one hierarchy is available, in order to make all the hierarchies simultaneously coherent. Another approach, in which convolutions are exploited on a graph data structure instead of a matrix, has been tested and it is currently under development. The advantage of this approach is that it is not needed to use a rigid data structure such as a matrix, but it is possible to exploit the flexibility of graphs. Several benchmark datasets are being tested, both simulated and real.
