Abstract
We present a method for conditional time series forecasting based on anadaptation of the recent deep convolutional WaveNet architecture. The proposednetwork contains stacks of dilated convolutions that allow it to access a broadrange of history when forecasting, a ReLU activation function and conditioningis performed by applying multiple convolutional filters in parallel to separatetime series which allows for the fast processing of data and the exploitationof the correlation structure between the multivariate time series. We test andanalyze the performance of the convolutional network both unconditionally aswell as conditionally for financial time series forecasting using the S&P500,the volatility index, the CBOE interest rate and several exchange rates andextensively compare it to the performance of the well-known autoregressivemodel and a long-short term memory network. We show that a convolutionalnetwork is well-suited for regression-type problems and is able to effectivelylearn dependencies in and between the series without the need for longhistorical time series, is a time-efficient and easy to implement alternativeto recurrent-type networks and tends to outperform linear and recurrent models.