Open mg64ve opened 5 years ago
Not to speak for the developer, but you are aware that that is how you train an AI model right? You can preprocess the dataset and then use the model when you run it on a live sample.
In general I would say that if you preprocess your dataset and then you split into train/test, you train the model and you check the results in the test part, then you are making a mistake. Because you assume to have knowledge of the future in order to preprocess the whole train/test dataset. I was thinking this is the case, but I am not sure anymore, I need to check the code again and I don't have time right now.
@mg64ve I am looking into this exact issue in implementing the WSAE-LSTM model, which uses the wavelet transform to denoise data (Bao et al., 2017): https://github.com/timothyyu/wsae-lstm
My implementation is a work in progress/currently vastly incomplete, but my understanding so far is that you cannot apply the wavelet transform to the entire dataset in one pass - but you can arrange the data in a continuous fashion in a clearly defined train-validate-test split that appears to mostly sidestep this issue.
From Bao et al. (2017) defining the train-validate-test split arrangement for continuous training (Fig 7): https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0180944#pone-0180944-g007
@timothyyu absolutely right! You should apply wavelet and any kind of preprocessing separately on train and test dataset. I am also working on this topic and I recommend you the following article:
Recurrent Neural Networks for Financial Time-Series Modelling / Gavin Tsang; Jingjing Deng; Xianghua Xie
It has some interesting concepts. Cheers
Here's an example of applying the wavelet transform to the first two train-validate-test
splits of the csci300 index data
:
Here's an example of applying the wavelet transform to the first two
train-validate-test
splits of thecsci300 index data
:
ok @timothyyu , does this example come from your code?
@mg64ve yes, this is from my own code. I have an updated implementation of the above (scaling is done on the train set, and then applied to the validate and test set per period/interval, and then the wavelet transform is applied to each train-validate-test split individually): https://github.com/timothyyu/wsae-lstm/blob/master/wsae_lstm/visualize.py
@mg64ve here is an updated version of the above that clearly illustrates the train-validate-test
split, with the effect of scaling and scaling + denoising being visualized:
Implemented as of v0.1.2 / b715d88 https://github.com/timothyyu/wsae-lstm/releases/tag/v0.1.2
Hi @timothyyu thanks for your reply, let me check the code. One more question: how do you apply scaling? Also to scaling you should apply the same concept. Validate and test datasets should be scaled without knowing them in advance.
Scaling is done with RobustScaler
on the train
set, and then the same parameters used to scale the train
set are applied to the validate
and test
sets.
ddi_scaled[index_name][intervals from 1-24][1-train,2-validate,3-test]
def scale_periods(dict_dataframes):
ddi_scaled = dict()
for key, index_name in enumerate(dict_dataframes):
ddi_scaled[index_name] = copy.deepcopy(dict_dataframes[index_name])
for key, index_name in enumerate(ddi_scaled):
scaler = preprocessing.RobustScaler(with_centering=True)
for index,value in enumerate(ddi_scaled[index_name]):
X_train = ddi_scaled[index_name][value][1]
X_train_scaled = scaler.fit_transform(X_train)
X_train_scaled_df = pd.DataFrame(X_train_scaled,columns=list(X_train.columns))
X_val = ddi_scaled[index_name][value][2]
X_val_scaled = scaler.transform(X_val)
X_val_scaled_df = pd.DataFrame(X_val_scaled,columns=list(X_val.columns))
X_test = ddi_scaled[index_name][value][3]
X_test_scaled = scaler.transform(X_test)
X_test_scaled_df = pd.DataFrame(X_test_scaled,columns=list(X_test.columns))
ddi_scaled[index_name][value][1] = X_train_scaled_df
ddi_scaled[index_name][value][2] = X_val_scaled_df
ddi_scaled[index_name][value][3] = X_test_scaled_df
return ddi_scaled```
Hi @timothyyu , I had a look to autoencoder.py and model.py. You basically don't use embedding. So you basically denoise all your test dataset and then use a value from denoised test dataset to predict the next step. This can't happen in real life because we only know the past also in the test dataset. That's why I am thinking that embedding is more useful because at each instant t you process the interval [t-N, t]. What do you think about it?
Hi, thank you for sharing your work and it's interesting. I am looking at the codes, but there were always some errors in generating results for stocks(I see well in FX rate). I would like to compare my results with yours for AAPL. Could you also present a predicted log return vs historical log return for AAPL for the most three years or one years if possible? Thank you very much!
Why stock_price = np.exp(np.reshape(prediction, (1,)))*stock_data_test[i]
?
File: model.py
https://github.com/VivekPa/AIAlpha/blob/18a58ef85b789d294c2fe4d2ab9ee7bc487fdb2f/model.py#L54
@az13js I believe it is because log return it is used during preprocessing
yeah, you should first split, then preprocess
This means that you assume to know the future! This would never work. Regards.