Whenever you evaluate a model I suggest you to do a T test on the profit values collected over time (positive and negative values) to see whether it is coming from random chance and there is any model power (you may consider 1 sigma to be weak power, 2 sigma normal and 3 sigma strong power). What T test does is tell you whether the mean value of your long term profit is far enough from the zero line considering its relative deviation. If there are more values than 30, you may just divide the profit mean by the profit standard error to get the sigma value. That simple. Also you might want to consider the commission fees. Great video!
I think the reason the Decision Tree was the most popular model was because the stocks you chose have a high trade volume, and therefore have less noise in their pricing. I would be interested to see how the model selection and accuracy would change with low trade volume stocks. Great video !
Excellent video! What features did you use to predict? Only the past price (past 99 days)? What about using technical indicators besides the past price?
Great video. As a suggestion, you may want to consider transaction costs, including (1) commission fees and (2) the bid-ask spread, since these lower the return of frequent trading relative to buying and holding.
Your analysis is excellent and well-explained. I was looking forward to your studies with longer lag periods. Is this something we can expect in a future video?
Talking about the label. Your models are built based on binary target? Where 1 is positive return more than 0.5 percent and 0 is less than treshold or the price goes opposite direction.
Excellent video, I like using decision trees, haven’t used on stock trading. I would be interested in looking at correlation of the snp to another data source. For example sentiment in the top 10 news articles over the last 90 days compared to performance, or something unconnected like sentiment of twitter posts over the same period. Determining sentiment on a given day would be an interesting subject to explore anyway without correlation
I haven’t for this one because there were so many other variables to already think about. In general, though, it’s an important part of the model selection process and a good idea!
Hi! Really enjoy your insight and the way you explain the models. Just one question: If I understood correctly, your largest lag was 3 weeks (~15 trading days), leaving only 85 days to train each tree? Would 85 observations be enough to train a tree?
Why not back test the smart selection methodology? Also why specify lagged returns as independent variables when returns are generally not autocorrelated?
I’m developing a stock trading model using a random forest and was wondering if I could run a few ideas past you. Curious if you are open to a consulting opportunity to assist with my model development? Thanks AR
Great video! It might have been good to go through the actual computation of one stock and tree computation just to see it done. But great nonetheless less!
Whenever you evaluate a model I suggest you to do a T test on the profit values collected over time (positive and negative values) to see whether it is coming from random chance and there is any model power (you may consider 1 sigma to be weak power, 2 sigma normal and 3 sigma strong power). What T test does is tell you whether the mean value of your long term profit is far enough from the zero line considering its relative deviation. If there are more values than 30, you may just divide the profit mean by the profit standard error to get the sigma value. That simple. Also you might want to consider the commission fees. Great video!
I think the reason the Decision Tree was the most popular model was because the stocks you chose have a high trade volume, and therefore have less noise in their pricing. I would be interested to see how the model selection and accuracy would change with low trade volume stocks. Great video !
(Hint) it’s 1000% more effective at analysis
Oh yes, I was waiting for this video actually! I really like tree based methods :D
Woo hope you enjoy!
Interesting stuff! Would like to see more of such content around financial markets.
More to come!
Excellent video! What features did you use to predict? Only the past price (past 99 days)? What about using technical indicators besides the past price?
Love your videos, thanks for such easy and simple explanations on ACF,PACF,AR,MA,ARIMA
Of course! Thanks for watching
Great video. As a suggestion, you may want to consider transaction costs, including (1) commission fees and (2) the bid-ask spread, since these lower the return of frequent trading relative to buying and holding.
Would love to see the code preferably in form of jupyter notebook. Thanks
Your analysis is excellent and well-explained.
I was looking forward to your studies with longer lag periods. Is this something we can expect in a future video?
Can u share the features u used and the Jupyter notebook as well
Talking about the label.
Your models are built based on binary target? Where 1 is positive return more than 0.5 percent and 0 is less than treshold or the price goes opposite direction.
Excellent video, I like using decision trees, haven’t used on stock trading. I would be interested in looking at correlation of the snp to another data source. For example sentiment in the top 10 news articles over the last 90 days compared to performance, or something unconnected like sentiment of twitter posts over the same period. Determining sentiment on a given day would be an interesting subject to explore anyway without correlation
Do you do hyperparameter tuning when you create a model?
I haven’t for this one because there were so many other variables to already think about. In general, though, it’s an important part of the model selection process and a good idea!
@@ritvikmath Your idea gives me great insight! Thank you for nice video
Hi! Really enjoy your insight and the way you explain the models. Just one question: If I understood correctly, your largest lag was 3 weeks (~15 trading days), leaving only 85 days to train each tree? Would 85 observations be enough to train a tree?
Why not back test the smart selection methodology? Also why specify lagged returns as independent variables when returns are generally not autocorrelated?
good note to backtest the smart selection strategy and using richer features in future videos
I’m developing a stock trading model using a random forest and was wondering if I could run a few ideas past you. Curious if you are open to a consulting opportunity to assist with my model development?
Thanks
AR
Random forest should have the highest bias no?
Do you share your code?
crazyy video!
🤯
Great video! It might have been good to go through the actual computation of one stock and tree computation just to see it done. But great nonetheless less!
a