432 points by alex_cortez 6 months ago flag hide 12 comments
fintech_fan 6 months ago next
This is really impressive, I've been following DL algorithms for financial predictions and 92% accuracy is a game changer!
deeplearner 6 months ago next
Thanks, I am part of the research team. We worked hard on this and the results are promising. AMA if you have any questions.
deeplearner 6 months ago next
@ml_hodler Great question. We used N-fold cross-validation, took frequent backtests, and incorporated regularization methods to avoid overfitting.
ml_hodler 6 months ago prev next
How is the accuracy measure defined over time, isn't there a danger of overfitting to historical data?
quant_guru 6 months ago prev next
We've seen similar results with our projects, but haven't announced specific numbers openly yet. Good to see more traction in something beyond supervised learning.
fintech_fan 6 months ago next
Imagine if someone managed to build an API or a platform out of this and created a crowdsourced prediction system to lower uncertainty for all users.
crypto_enthusiast 6 months ago prev next
I hope they can reduce the latency for real-time predictions. Predicting 0.1 seconds before the price shift can be lucrative for high- frequency trading.
deeplearner 6 months ago next
@crypto_enthusiast We accounted for latency too. We're working on bringing the average prediction time down whilst keeping accuracy high.
data_tactician 6 months ago prev next
What are the input features? Can you share a preprint on Arxiv or a blog post?
deeplearner 6 months ago next
@data_tactician Yes, we are preparing a blog post and will share on Arxiv soon! We use a combination of traditional features plus a few news and social media sentiments as input.
science_novice 6 months ago prev next
What was the biggest validation hurdle to overcome?
deeplearner 6 months ago next
@science_novice One major challenge was dealing with bias from specific data periods. We tried a variety of techniques such as transfer learning.