567 points by machinelearn 6 months ago flag hide 14 comments
finance_ai 6 months ago next
Fascinating case study! I appreciate the in-depth exploration of neural networks (NNs) in stock market prediction.
quant_analyst 6 months ago next
Indeed, the author demonstrates the feasibility of using NNs in predicting the stock market. However, the challenge lies in time series forecasting when new data is continuously introduced.
rnn_ftw 6 months ago next
Recurrent Neural Networks (RNNs) have been quite successful in time-series prediction. How well do they apply to short-term vs long-term predictions?
quant_analyst 6 months ago next
RNNs do tend to perform better with short-term predictions due to the effects of vanishing gradients in long sequences. However, there are variants like LSTMs or GRUs that can alleviate this issue in long-term predictions.
ma_master 6 months ago prev next
great research. I want to learn more about how to build similar models for my own needs. How do you create an architecture that addresses the need for constant data input?
finance_ai 6 months ago next
Thank you! Streaming data into NNs can be done using recurrent or convolutional models that can handle sequences/temporal aspects. Another option is using LSTMs or GRUs to cope with variable length input sequences.
learnml 6 months ago prev next
Would you care to elaborate on the process of hyperparameter tuning for your NN model in this project?
finance_ai 6 months ago next
Certainly! I mainly used grid search for hyperparameter tuning to ensure that the model did not fall into local minima. I tuned learning rate, batch size, number of layers, and number of neurons for our specific use case.
deep_learning_fan 6 months ago prev next
I've been following the research in using NNs for stock market prediction, and I've noticed a lot of inconsistency in results. What do you think the primary reason for this is?
the_auth 6 months ago next
Excellent question! Inconsistency can result from factors such as overfitting, lack of data, or erroneous data. It can also arise from the volatile and non-linear nature of the stock market.
algorithmic_rules 6 months ago prev next
Practically speaking, how do NNs compete with traditional algorithmic trading models? Obviously, this is an important factor in real-world application.
the_auth 6 months ago next
That's a critical point. Traditional algorithmic trading models focus on rule-based strategies, which may offer better explainability than NNs that are considered black boxes. In terms of returns, it can be a mix. NNs have the potential to find unconventional patterns but might not always generalize well.
frequentist 6 months ago prev next
How do NNs stand against classic statistical forecasting methods? Given the need for large and explainable models, is it worth investing resources in NNs?
t_stat 6 months ago next
Statistical forecasting typically delivers results quickly and relies on fewer resources than NNs. However, NNs can detect complex and subtle patterns that might be missed by statistical methods. It ultimately depends on the desired outcome and available resources.