1314 points by mlstockgenius 6 months ago flag hide 15 comments
finance-ai 6 months ago next
This is amazing! I've been working on this project to use machine learning and predict stock market prices with high accuracy. Any questions about the methodology, let me know!
deep-learning-enthusiast 6 months ago next
Wow, that's truly impressive. Can you tell us more about how you get to 95% accuracy? What kind of data did you use to train your model?
finance-ai 6 months ago next
Sure! We used historical stock prices along with various features such as financial news sentiment, volumes, and more. The machine learning model was trained with recurrent neural networks (RNNs) and LSTMs.
another-user 6 months ago prev next
This is incredible. I have a few follow-up questions: 1. How do you address overfitting? 2. Have you tested your model with different stocks and assets? 3. Do you have any advice for someone who wants to replicate your work?
finance-ai 6 months ago next
Great questions: 1. We used dropout and early stopping to address overfitting. 2. Yes, we've tested a variety of stocks in multiple sectors and markets, with similar results. 3. My advice would be to learn about the different types of machine learning algorithms and apply a combination of approaches to your own dataset, if possible, making sure to validate the effectiveness of each model before moving forward.
just-curious 6 months ago prev next
What about the ethical dilemmas of using AI to predict stock market prices? Couldn't it make existing market inequalities worse?
finance-ai 6 months ago next
This raises an excellent point. Our primary motivation for this research is to better understand financial markets and to create a tool for investment guidance. It is not our intention to exacerbate market inequalities or enable rapid, profitable trading by institutions. Nevertheless, the topic of fairness in AI is critical and should be considered in all machine learning applications, including this one.
keep-learning 6 months ago prev next
Congratulations on your work. Could you touch on the types of machine learning algorithms used? Was it traditional ML or did you use deep learning?
finance-ai 6 months ago next
Our models primarily used deep learning, specifically long short-term memory (LSTM) networks. LSTMs are great at handling sequential data, like time series, and preserving long-term relationships within that data. It was important for us to use a deep learning model since ML is not enough to tackle the complex nature and dependencies that exist in the stock market.
fire-away 6 months ago prev next
Wow, that's incredible. Did you consider other models, like XGBoost, LightGBM or Random Forest?
finance-ai 6 months ago next
Yes, we have explored multiple machine learning algorithms. While traditional ML models can deliver reasonable results on smaller datasets, deep learning models often generalized better and provided more accurate predictions for our use case. However, I believe a combination of ML and deep learning could still provide valuable insights and complementary results.
questioner 6 months ago prev next
Did you try using reinforcement learning to predict stock markets? How do you think it would perform?
finance-ai 6 months ago next
Thanks for asking! We've experimented with reinforcement learning (RL) as well. While RL methods can capture long-term dependencies within markets and potentially better manage risk, the convergence of RL models is significantly affected by the credit assignment problem which is a challenge in financial domains. Therefore, we found that RNNs still held the best computational cost and interpretability.
skeptical-mind 6 months ago prev next
95% accuracy seems too high to be plausible. Can you clarify what you mean by stock market prices and share any details about your evaluation metrics?
finance-ai 6 months ago next
We apologize for any confusion, and we appreciate your vigilance. By stock market prices, we actually meant to say returns or percentage changes. A more realistic accuracy for returns can span from 58% to 64%, still delivering alpha generation for portfolios. We calculated the prediction accuracy based on both in-sample and out-of-sample testing.