This repository holds the reproduction codes for A Dual-Stage Attention-Based Recurrent Neural Network for Time Series Prediction-IJCAI2017.

In ./darnn-based, we reproduce the work in this paper on the Nasdaq 100 dataset. Besides, we offer a modified version in ./darnn-improved to solve the instability of the original model during training period. The requirements are Python 3.6 and Pytorch 1.0 (or higher).

1. ./darnn-based: the origin model introduced in A Dual-Stage Attention-Based Recurrent Neural Network for Time Series Prediction-IJCAI2017.
2. ./darnn-improved: we found the attention weights used in encoder are highly unstable in the trainning period when we try to reproduce the results. To solve this issue, we added an additional attention layer in encoder. With this modification, the first attention layer in encoder is responsible for learning the highly variable dependences, as well as the second attention layer is responsible for learning the stable and long-term dependences.

Limited by data acquisition, we only tested the performance of two models on the Nasdaq100 dataset, and the results are shown in the following table:

These two file folders both contain all the code and data needed to run the model. You can use any one of them to start the trainning period.