Autoregressive Convolutional Neural Networks for Asynchronous Time\n Series

Abstract

We propose Significance-Offset Convolutional Neural Network, a deep\nconvolutional network architecture for regression of multivariate asynchronous\ntime series. The model is inspired by standard autoregressive (AR) models and\ngating mechanisms used in recurrent neural networks. It involves an AR-like\nweighting system, where the final predictor is obtained as a weighted sum of\nadjusted regressors, while the weights are datadependent functions learnt\nthrough a convolutional network. The architecture was designed for applications\non asynchronous time series and is evaluated on such datasets: a hedge fund\nproprietary dataset of over 2 million quotes for a credit derivative index, an\nartificially generated noisy autoregressive series and UCI household\nelectricity consumption dataset. The proposed architecture achieves promising\nresults as compared to convolutional and recurrent neural networks.\n