search

nayash / stock_cnn_blog_pub

This project is a loose implementation of paper "Algorithmic Financial Trading with Deep Convolutional Neural Networks: Time Series to Image Conversion Approach"
Apache License 2.0
164 stars 95 forks source link

Differing results in article and ipynb #6

Open via986 opened 3 years ago

via986 commented 3 years ago

Hello!

Thanks for you article and repo. I'm really interested in it! Could you explain something?

At the and of your article https://towardsdatascience.com/stock-market-action-prediction-with-convnet-8689238feae3 you wrote results: https://miro.medium.com/max/403/1*YLKDhlcx6TtzIoVSdJ2xNw.png For example, for class 0: TP 49 FP 195.

But, in https://www.kaggle.com/darkknight91/predicting-stock-buy-sell-signal-using-cnn/ in the end: [[ 53 0 11] [ 0 50 11] [ 38 53 784]] TP 53 and only 38 FP !!! It's fantastic result!

I tried all params from your example and I can easily get first result with FP = 4*TP, but didn't get even close to TP =~ FP Val_loss is too big and f1_metric poor (( How did you manage to get second results? This results is really promising!

Thank you!

nayash commented 3 years ago

Hi,

I fixed some bugs and updated the article and GitHub code, can't recall if Kaggle notebook is updated. I would suggest to follow the article. You will also find the latest fixes at the bottom of article.

via986 commented 3 years ago

I fixed some bugs and updated the article and GitHub code, can't recall if Kaggle notebook is updated. I would suggest to follow the article. You will also find the latest fixes at the bottom of article.

Ok, but it is pity ( I have a huge experience in making trading algo systems on stocks and futures. Kaggle results would be "holy grail' in trading, really )) If you interested in such kind of activity, we could try to improve results of this system together. Or try another ML algorithm.