There's been a lot of exciting developments in the past few years using new neural network techniques to classify images and video. Some of the new techniques address text and language processing problems as well. But, beyond LSTMs there is not as much focus in applying these new techniques to time-series modeling. This talk describes an experimental approach to modeling advertising category value using network architectures inspired by both the current wave of 2D convolutional neural network layer architectures and also the more traditional 1D convolution filters as they apply to discrete time signal processing.

We present some experimental network architectures using these inspirations that have performance at parity and better with standard ARIMA models. Practicing data scientists as well as those with some knowledge of neural networks should find this talk of interest. We also present an open source Python package using standard neural network tools which implements this approach.