dual_stage_attention_rnn

In an attempt to learn Tensorflow, I have implemented the model in A Dual-Stage Attention-Based Recurrent Neural Network for Time Series Prediction using Tensorflow 1.13.

• Nasdaq data is used for testing, which is from repo da-rnn.
• Based on the discussion, i implemented both cases where current exogenous factor is included, i.e., as well as excluded, i.e. . The switch between the two modes is control by FLAGS.use_cur_exg in da_rnn/main.py.
• To avoid overfitting, a flag to shuffle the train data has been added, activated by FLAGS.shuffle_train in da_rnn/main.py.
• A ModelRunner class is added to control the pipeline of model training and evaluation.

Run

Source Data

Put the downloaded Nasdaq csv file under data/data_nasdaq.

da_rnn
|__data
    |__data_nasdaq
            |__nasdaq100_padding.csv

Run the training and prediction pipeline

Suppose we want to run 100 epochs and use Tensorboard to visualize the process

cd da_rnn
python main.py --write_summary True --max_epoch 200

To check the description of all flags

python main.py -helpful

To open tensorboard

tensorboard --logdir=path

where path can be found in the log which shows the relative dir where the model is saved, e.g. logs/ModelWrapper/lr-0.001_encoder-32_decoder-32/20190922-103703/saved_model/tfb_dir.

Test result

Results of my experiments are listed below. Running more epochs and applying larger encoder/decoder dimension could possibly achieve better results.

# Epoch	Shuffle Train	Use Current Exg	Econder/Decoder Dim	RMSE	MAE	MAPE
100	False	False	32	105.671	104.60	2.15%
100	True	False	32	29.849	29.033	0.59%
100	False	True	32	46.287	32.398	0.66%
100	True	True	32	1.491	1.172	0.024%

# To shuffle the train data and use current exogeneous factor
python main.py --write_summary True --max_epoch 100 --shuffle_train True --use_cur_exg True

After 100 epochs(with data shuffled and current exogenous factors used) the prediction is plot as

Requirement

tensorflow==1.13.1
scikit-learn==0.21.3
numpy==1.16.4

Although I have not tested, I guess it should be working under tf 1.12 and tf 1.14 as well.

Reference

• A PyTorch Example to Use RNN for Financial Prediction
• Github: da-rnn