be sure to consider trading costs (commisions, bid-ask spread, financing, dividends...). Many strategies perform bad after costs. Also, the result should be compared with the index return as benchmark.
One idea that may not have been covered yet -- but since you're already using a neural network it wouldn't be such a huge leap -- might be to use multiple (maybe even *all* of ) the ticker histories at once to predict multiple (or again, maybe all of) the returns at once. This would allow the neural network to utilize cross-industry information and correlations etc. to make a more accurate overall prediction. So, it might be able to pick out a 'portfolio' of stocks to make the overall investment less prone to risk.
Great explanation, good luck with this experiment. I wounder if transformers is somehow applicable here as more computationally effective model for series data as ltsm?
You need a language model and news feature training to even have a chance at good averages. Regression off rsi, index, and volume alone will lose you money since trends have nothing to do with them; might as well be using standard deviation(drip) or curve equations(day trading); which work and are easy, by the way LSTM with regression is the right model, though. With good ML even a Perception network could work(simplest model) 100% of regression with stocks is done with news correlation. Anything else is hard tolerances and index curves; unless you're sitting on cheap shares waiting for volume trends to cash out which just needs a simple bot
Thanks for the video as always, I'm curious about the backtesting part. What if you just randomize the training and test set, meaning that you don't need to train the model on the first half and test on the latter half, but rather just randomly pick 50% for training and 50% for testing?
@@cornagojar that’s a very good point! Thanks for pointing that out. However, I still think it might work, depending on the nature of the time series of course. I might be wrong, but here’s what I think, it’s true that only the past can determine the future. However, if a helpful pattern is found even in the future, then basically, it’s our assumption that the pattern is generalisable, meaning that if we found that pattern whether it’s a month ago or a year later, the outcome should be relatively similar, isn’t that what every time series prediction do? The prediction should be the same whether it’s in the future or it’s in the past as long as the input is the same. Let me put this in another way, say you have a pattern that is helpful in predicting a time series, if the pattern holds true in the future, then it should holds true for the past as well, shouldn’t it? If not, then I think we have no reason to predict anything really. The reason we use the training set to be the past time points is simply because we can’t have time points from the future. But in the training set itself, it seems there is no problem for me to rearrange it as I would like it, since we are just finding that general pattern, that hopefully will be useful in predicting the real test set. Sorry for the lengthy reply, but I’m genuinely curious about this, and I’m by no means an expert (as you might have notice already from my questions lmao). I only recently started learning time series analysis. Looking forward to your response ;)
Im not watching all that, but always remember to look at commisions and sharpe ratio! If everything was that easy, hedges wouldn't need 50 math phds, 50 data scientists and propably hundret of statisitcians
☢☢☢☢☢ For forecasting it is necessary to involve the knowledge and laws of MetapPhysics. Without MetaPhysics it is impossible to predict the future. Based on the theory of probability, on the theory of statistical analysis, the future is very, very difficult to predict. MetaPhysics is that window in the Universe that will allow you to look into the future!!
Really nice explanations of the concepts here. Probably the most difficult aspect to understand. I also really like the pen and paper style diagrams.
Glad it was helpful! Appreciate the feedback
Thanks for the video! I enjoy it much. The way you structure information is brilliant! Great job!
thanks a bunch!
be sure to consider trading costs (commisions, bid-ask spread, financing, dividends...). Many strategies perform bad after costs. Also, the result should be compared with the index return as benchmark.
good tip! and yes we'll definitely compare to the index return in the results video next week
great vids as always, keep it up👍🏻
Thanks, will do!
One idea that may not have been covered yet -- but since you're already using a neural network it wouldn't be such a huge leap -- might be to use multiple (maybe even *all* of ) the ticker histories at once to predict multiple (or again, maybe all of) the returns at once. This would allow the neural network to utilize cross-industry information and correlations etc. to make a more accurate overall prediction. So, it might be able to pick out a 'portfolio' of stocks to make the overall investment less prone to risk.
Wow love this idea 💡
got my exam tomorrow, thanks
Best of luck!
Please do a video on stacked LSTM to generate single model for multiple stocks.
thanks for the suggestion!
I can't wait.
Me too!
Hi Ritvik. Thanks for another excellent video. My only question is shouldn't one normalize the volume cause it's on a different scale to returns?
Hey great comment, we did normalize the variables before passing them into the networks
Great explanation, good luck with this experiment. I wounder if transformers is somehow applicable here as more computationally effective model for series data as ltsm?
Great suggestion!!
You need a language model and news feature training to even have a chance at good averages. Regression off rsi, index, and volume alone will lose you money since trends have nothing to do with them; might as well be using standard deviation(drip) or curve equations(day trading); which work and are easy, by the way
LSTM with regression is the right model, though. With good ML even a Perception network could work(simplest model)
100% of regression with stocks is done with news correlation. Anything else is hard tolerances and index curves; unless you're sitting on cheap shares waiting for volume trends to cash out which just needs a simple bot
Thanks for the video as always, I'm curious about the backtesting part. What if you just randomize the training and test set, meaning that you don't need to train the model on the first half and test on the latter half, but rather just randomly pick 50% for training and 50% for testing?
that's an idea worth trying! will keep it in mind for the next trading video. seems like it would add some non-determinism to the process.
@@ritvikmath Yup, or applying k-fold dan give us some idea how stable/unstable the model is. Cheers! waiting how much money you’ll make out of this.
you can't with time series, as you would introduce lookahead bias
very true that we'd need to be more careful in doing this with time series data
@@cornagojar that’s a very good point! Thanks for pointing that out. However, I still think it might work, depending on the nature of the time series of course. I might be wrong, but here’s what I think, it’s true that only the past can determine the future. However, if a helpful pattern is found even in the future, then basically, it’s our assumption that the pattern is generalisable, meaning that if we found that pattern whether it’s a month ago or a year later, the outcome should be relatively similar, isn’t that what every time series prediction do? The prediction should be the same whether it’s in the future or it’s in the past as long as the input is the same. Let me put this in another way, say you have a pattern that is helpful in predicting a time series, if the pattern holds true in the future, then it should holds true for the past as well, shouldn’t it? If not, then I think we have no reason to predict anything really. The reason we use the training set to be the past time points is simply because we can’t have time points from the future. But in the training set itself, it seems there is no problem for me to rearrange it as I would like it, since we are just finding that general pattern, that hopefully will be useful in predicting the real test set.
Sorry for the lengthy reply, but I’m genuinely curious about this, and I’m by no means an expert (as you might have notice already from my questions lmao). I only recently started learning time series analysis. Looking forward to your response ;)
your point about 90 days and generalization is kind of subtle. is there a way that you can explain again in another video possibly.
It is quite subtle, I’ll try to go over it again in the next video!
Im not watching all that, but always remember to look at commisions and sharpe ratio! If everything was that easy, hedges wouldn't need 50 math phds, 50 data scientists and propably hundret of statisitcians
Are you building this LSTM model per ticker?
Any chance you might share your code? I'd love to play around with the model!
thanks for the video, I'm also working on trading bots. can i get your help and code for reference as well .
☢☢☢☢☢
For forecasting it is necessary to involve the knowledge and laws of MetapPhysics.
Without MetaPhysics it is impossible to predict the future.
Based on the theory of probability, on the theory of statistical analysis, the future is very, very difficult to predict.
MetaPhysics is that window in the Universe that will allow you to look into the future!!