Abstract
Stock price prediction is a challenging task due to its complexity and the dynamics associated with stock prices. In this paper, we propose a deep-learning based end-to-end framework with a novel architecture, for multi-step ahead stock closing price forecasts. The architecture exploits an encoder-decoder framework with variants of convolutions and recurrent neurons, in order to perform representation learning for the past behavior of the stock as well as associated exogenous factors. We incorporate an attention mechanism to capture the long term dependencies between inputs and outputs, and deploy Monte Carlo dropout layers in the architecture design to provide a stochastic setting for uncertainty estimation. We validate our model on two real datasets for AMZN and AAPL stocks.