234 points by cryptocoder 7 months ago flag hide 13 comments
cryptodev 7 months ago next
I've been experimenting with Deep Learning to predict Cryptocurrency prices for the past few months. I’m amazed with the results, it’s incredibly promising. I used LSTM networks, a type of Recurrent Neural Network, to handle sequential data and they proved to be very efficient.
mlengineer 7 months ago next
That's fascinating! I've heard good things about using LSTMs for financial data analysis. What was your testing process like? Backtesting on historical data or doing any real-time evaluations?
cryptodev 7 months ago next
I backtested my LSTM models on several assets to evaluate how they would have performed historically. They showed significant improvement compared to traditional prediction models. Recently i've also setup monitoring dashboard for real-time testing, allowing me to evaluate performance in real-time too.
ai_enthusiast 7 months ago next
Incredible! Have you looked into combining any Convolutional Neural Networks (CNN) to improve your model's accuracy? I've heard of some researchers working on hybrid models for financial predictions with positive results.
cryptodev 7 months ago next
No, I haven’t but that sounds interesting! I’ve solidified my model structure with LSTMs but would be open to trying a hybrid model with CNNs if I had a more robust reason to believe it'd improve performance. Thanks for sharing this idea!
quant_expert 7 months ago prev next
Using LSTM to predict crypto prices is truly an innovative idea! One challenge could be dealing with high market volatility and making sure the model remains robust through these situations.
cryptodev 7 months ago next
Yes, market volatility has been an issue, but I implemented some regularization techniques and did extensive hyperparameter tuning on the model to minimize overfitting and help with overall robustness.
quant_analyst 7 months ago next
Great job ensuring your model accounts for volatility! Alternatively, some have considered incorporating fundamental analysis or alt data to improve their modeling outcome. Have you tried any of those?
cryptodev 7 months ago next
Not yet, but I think incorporating fundamental analysis or alt data could add interesting insights to the model. I'll definitely consider that as a future implementation!
random_hacker 7 months ago prev next
DL to predict crypto prices, cool concept! What would your advice be for fellow hackers interested in replicating this experiment?
cryptodev 7 months ago next
Thanks! I'd say start by mastering the basics of DL and financial time series modeling. Once you've grasped those, read on the latest research papers in combining them for crypto predictions. Good luck and don't forget to write a HN post about your findings as well!
neural_networks 7 months ago prev next
I agree, DL has many applications in crypto markets. In case anyone is planning to try building their own crypto price predictive model, is there a specific dataset or libraries you'd recommend using?
cryptodev 7 months ago next
For a solid starting library, I'd recommend using TensorFlow, Keras or PyTorch for DL. To get data I suggest collecting from multiple reputable sources to have robust info for your models.