78 points by deeplearner 6 months ago flag hide 10 comments
deeplearning_fan 6 months ago next
Fascinating article! I've been experimenting with similar DL techniques for stock market prediction. I see you're using LSTM in your model. Have you tried adding attention layers? It can really help the model focus on the right input features at the right time.
data_scientist 6 months ago next
That's an interesting suggestion! I haven't tried attention layers. I'm curious, can you please share some good resources for implementing attention in LSTMs for stock prediction? I'd appreciate any guidance. :)
deeplearning_fan 6 months ago next
Sure, I'd recommend checking out this tutorial on TensorFlow: <https://www.tensorflow.org/tutorials/text/nlp_with_attention>. It provides an extensive overview of attention implementation for NLP tasks, and you can easily adapt it to your model. Good luck!
deeplearning_fan 6 months ago next
You can also try using pretrained sentence encoders like BERT to extract rich feature representations from text sources like news articles. It could potentially boost model performance!
machinelearningmaster 6 months ago prev next
Great story! I do have a concern though. Stock prices are influenced by many factors, including global market trends, news articles, and investor sentiments. How does your model handle data from these external sources and factors?
dataengineer 6 months ago next
To incorporate external information, you can use techniques like web scraping or API requests to gather data feeds from news sources, social media, and market data. Then, preprocess this data along with your stock price data and feed it to your model.
mlguy 6 months ago next
Regarding web scraping, I always recommend gentle and responsible scraping. Stick to a reasonable number of requests to avoid overwhelming the target website and be sure to obey terms of use and copyright.
quant_modeler 6 months ago next
Totally agree! Responsible web scraping showcases the importance of the community, respecting others’ resources, and ensuring your code stays within legal boundaries. Well done!
hacker_algo 6 months ago prev next
Nice one! Instead of using LSTMs, I prefer GRU networks because they have fewer parameters and are thus faster. Have you compared LSTM vs GRU performance for stock prediction tasks? A/B test results?
data_analysis 6 months ago next
GRUs can indeed be faster, but they may lack behind LSTMs when working with sequential data of a longer length. My experience has been mixed. LSTMs tend to more accurately capture long term patterns, while GRUs are efficient and work better with recent context.