Hello, thank you for sharing this work. Can I find the PyTorch codes of your attention layer anywhere?

Thanks for your good work.

But from parameter file which could reproduce your work,

Adv-ALSTM:
    python pred_lstm.py -l 5 -u 4 -l2 1 -v 1 -rl 1 -q ./saved_model/acl18_alstm/exp -la 0.01 -le 0.05

Adv-ALSTM:
    python pred_lstm.py -p ./data/kdd17/ourpped/ -l 15 -u 16 -l2 0.001 -v 1 -rl 1 -q ./saved_model/kdd17_alstm/model -la 0.05 -le 0.001 -f 1

seems attention("-a") and hinge loss("-hi") is not used, which is a key improvement points of your paper,
could you please tell me the reason ? Thanks !

Hello, thank you for your sharing!
I can reproduce your results successfully. But I notice that when you conduct your 'Adv-ALSTM' method, you load a pretrained model, which is not used for other methods. I wonder how can I get the pretrained model because when I train the model with random initialization, I can't get a good result.

Firstly, thank you for sharing your code and data.
However, when loading data, the code just skips samples with label 0, whose next-day return is between -0.5% and 0.55%. It's not good especially for test set. For unknown data in test set, It actually have 3 labels, and the code destroys the distribution of test data.
I just wonder whether my understand is correct.

I followed the feature definition in the paper to build the dataset, many of the data are very small numbers, but in your dataset there are many numbers greater than 1, according to the formula for defining features, these data should be very small, so it is not very clear how the data in ourpred is built

比如一只股票s，它的输入数据不应该是按时间来排列的吗？看代码中的实现是：按时间点开始，把所有当前时间点的股票数据排列在一起，然后在训练的时候，取batch个数据，这个时候这个输入序列并不是时间序列了呀，望解惑！！！

First of all, Thank you for sharing this work.
We have several questions. May I ask if you don't mind?

1. We've researched your work, but we are not able to find your code to generate 'ourpped' data.
   Could you let us know where the code exist or share the code?

2. In raw data in kdd, there are the number, -123321. What dose this number mean? Plus, I wonder meanings/information of each column. (There are 13 columns in AAPL. However, there are only 9 columns (OLHC, volumes) in kdd/price_long_50)