Great video, I love it so much! I feel the ARMA model in this instance if combine with short selling would be somewhat similar to a weak or loose mean reversion strategy. As I found that in your experiment it is mostly buying stocks that have negative returns on the last day. Additionally, instead of training solely based on past stock return, I feel it could also potentially train on both idiosyncratic and systemic risk/factors using APT framework, dissect the holistic daily return into different factor loads/components to improve the accuracy of the model. I am not sure it would work, just some thoughts, as I just graduated from high school.
I am really curious about how an arima model would perform. Also I would suggest doing a side by side experiment with day trading like in this video and week trading to see which performs better using the same method.
@ritvikmath I am curious to know your thoughts on whether using a model that is trained on past values of the series itself only (i.e., no exogenous features) to predict stocks is a good practice in the first place or not?
imo you should've allowed both long and short so that the ARMA can actually perform by its design and we could've known whether this model is any better than mere random chance.
Hey, this video was amazing. Can I ask where can I follow along with the full series. If there isn't, it would be great if you could provide the link to the notebook to look at your implementation approach. That would be really helpful.
Hey there! I really enjoyed this video and was impressed by the model you presented. Have you considered adding wavelet transform to your analysis? I think it could add a lot of value to your model. Once you've transformed the data, you could then build ARMA models for each component of the wavelet transform. This would allow you to capture more nuanced trends in the data and make more accurate predictions. Just a thought - keep up the great work!
I'ver read that if you have a sudden large change in a stock, it is likely to be followed by another large change in the stock. However, it can go either direction. It was possible that the model was picking up this big changes, which could have resulted in greater variability to returns. Also the stock returns don't follow normal distribution and have fat tails. So confidence interval probably overestimates the probability of a stock falling within a certain range.
Thank you for running this experiment, but let me add a prior belief you should have had from the outset: it is essential that you avoid the Ludic Fallacy. Stock markets are controlled by banker guys like "Fat Tony" who understands that the rules of the game are rigged in favor of the big fish. They will take you for a ride. There are smart ways to use ML to invest, but blindly running ARIMA on any given stock and not considering all the other domain knowledge that goes into investing is absolute insanity (read: foolish stupidity). I recommend you read Nassim Taleb's books like "The Black Swan" and avoiding the intellectual hubris that this experiment showcases - the efficient market hypothesis works because of corporate insiders; retail chumps running ML do not stand a chance, you're trying to apply the models to a rigged game - markets are social constructs, this ain't physics!
It would be interesting if you can build a model that takes into account sentiment as a weighting factor. Something such as search frequency for a stock and/or positive and negative keywords tied to a company.
could it be worth trying to fit an ARMA model on the S&P to see if the market as a whole is worth investing in, then if so, running the individual tickers to see whats worthwhile?
Interesting idea! One possible corner case is when the ARMA model doesn’t think the market is worth investing in as a whole but there are some stocks in the market doing better. Of course, finding stocks to “beat the market” in this way was precisely what we tried and failed to accomplish here
Good thing you had those extra $0.80 laying around on Tuesday morning to make up for Monday's loss, otherwise you'd have to invest only %99 on Tuesday. having said that, you made $1.02 overall on a $100 investment in 5 days. that's annualized 51% return (based on 250 trading days). Many portfolio managers would give a lot to have those kinds of results. :) ps where's the video you've mentioned in the beginning?
@@stevenpham6734 elaborate please?same size of what? I didn't comment on portfolio creation, I haven't seen the video yet - I personally have no problem with an equal weighted portfolio. as for the fact that the author was investing $100 daily, that's his choice - that's why I took total weekly gain and divided by 5.
@@notasan I think what steven is refering to is that the video is only a singular experiment. To validate the method we would need multple to see if there is a significant positive return
@@VRS-hj6ep oh, absolutely, you cannot judge a strategy by its performance over 5 days. SP 500 returned about 100% annualized over the same period. However, not sure that's what he said though.
Interesting teaching application! But isn't it also true that it is basically impossible to consistently make money on the stock market regardless of any model used?
Hii ritvic, can you make a model were the main objective is to get more of the asset. Like sell 5% of 10% gains and buy with 40% of cash when price dips 10% or something. I would like to grow my btc holdings automatically. 😅
Did you buy the stocks in the morning? I guess that's not the same price as when they closed the day before right? so that deviates from training for sure. I'm guessing this is a simulation but if it was real then you would need not to buy them in the morning
Great video, I love it so much! I feel the ARMA model in this instance if combine with short selling would be somewhat similar to a weak or loose mean reversion strategy. As I found that in your experiment it is mostly buying stocks that have negative returns on the last day. Additionally, instead of training solely based on past stock return, I feel it could also potentially train on both idiosyncratic and systemic risk/factors using APT framework, dissect the holistic daily return into different factor loads/components to improve the accuracy of the model. I am not sure it would work, just some thoughts, as I just graduated from high school.
Dude if you just graduated HS, I hope you have plans in place to make the most of your savvy/knowledge. Saying as someone who blew it!
a couple of more attempts and we'll crack this whole stock market thing WIDE open, right here
Great Video! Liked it so much and had so much fun watching it
So glad!
Great video! Would be nice to include the link of the other videos in the description so it's easy to follow the serie.
I am really curious about how an arima model would perform. Also I would suggest doing a side by side experiment with day trading like in this video and week trading to see which performs better using the same method.
Great suggestion!
@ritvikmath
I am curious to know your thoughts on whether using a model that is trained on past values of the series itself only (i.e., no exogenous features) to predict stocks is a good practice in the first place or not?
Great videos. Adding an “only-blue-chip” filter for the stocks’ selection criteria would be interesting. Again, I very much enjoy your content.
imo you should've allowed both long and short so that the ARMA can actually perform by its design and we could've known whether this model is any better than mere random chance.
Thanks! Will look into it for next time
This is awesome stuff. Thanks for doing this and sharing it!!!
Hey, this video was amazing. Can I ask where can I follow along with the full series.
If there isn't, it would be great if you could provide the link to the notebook to look at your implementation approach. That would be really helpful.
Hey there! I really enjoyed this video and was impressed by the model you presented. Have you considered adding wavelet transform to your analysis? I think it could add a lot of value to your model. Once you've transformed the data, you could then build ARMA models for each component of the wavelet transform. This would allow you to capture more nuanced trends in the data and make more accurate predictions. Just a thought - keep up the great work!
I'ver read that if you have a sudden large change in a stock, it is likely to be followed by another large change in the stock. However, it can go either direction. It was possible that the model was picking up this big changes, which could have resulted in greater variability to returns.
Also the stock returns don't follow normal distribution and have fat tails. So confidence interval probably overestimates the probability of a stock falling within a certain range.
Thank you for running this experiment, but let me add a prior belief you should have had from the outset: it is essential that you avoid the Ludic Fallacy. Stock markets are controlled by banker guys like "Fat Tony" who understands that the rules of the game are rigged in favor of the big fish. They will take you for a ride.
There are smart ways to use ML to invest, but blindly running ARIMA on any given stock and not considering all the other domain knowledge that goes into investing is absolute insanity (read: foolish stupidity).
I recommend you read Nassim Taleb's books like "The Black Swan" and avoiding the intellectual hubris that this experiment showcases - the efficient market hypothesis works because of corporate insiders; retail chumps running ML do not stand a chance, you're trying to apply the models to a rigged game - markets are social constructs, this ain't physics!
It would be interesting if you can build a model that takes into account sentiment as a weighting factor. Something such as search frequency for a stock and/or positive and negative keywords tied to a company.
really interesting analysis.
Thanks!
could it be worth trying to fit an ARMA model on the S&P to see if the market as a whole is worth investing in, then if so, running the individual tickers to see whats worthwhile?
Interesting idea! One possible corner case is when the ARMA model doesn’t think the market is worth investing in as a whole but there are some stocks in the market doing better. Of course, finding stocks to “beat the market” in this way was precisely what we tried and failed to accomplish here
Good thing you had those extra $0.80 laying around on Tuesday morning to make up for Monday's loss, otherwise you'd have to invest only %99 on Tuesday. having said that, you made $1.02 overall on a $100 investment in 5 days. that's annualized 51% return (based on 250 trading days). Many portfolio managers would give a lot to have those kinds of results. :)
ps where's the video you've mentioned in the beginning?
you forget one important concept: SAMPLE SIZE!!!
@@stevenpham6734 elaborate please?same size of what?
I didn't comment on portfolio creation, I haven't seen the video yet - I personally have no problem with an equal weighted portfolio. as for the fact that the author was investing $100 daily, that's his choice - that's why I took total weekly gain and divided by 5.
@@notasan I think what steven is refering to is that the video is only a singular experiment. To validate the method we would need multple to see if there is a significant positive return
@@VRS-hj6ep oh, absolutely, you cannot judge a strategy by its performance over 5 days. SP 500 returned about 100% annualized over the same period. However, not sure that's what he said though.
Interesting teaching application! But isn't it also true that it is basically impossible to consistently make money on the stock market regardless of any model used?
Hii ritvic, can you make a model were the main objective is to get more of the asset. Like sell 5% of 10% gains and buy with 40% of cash when price dips 10% or something. I would like to grow my btc holdings automatically. 😅
Interesting idea!
can we consider looking at factor based approach for future videos.
Will look into it, thanks!
Did you buy the stocks in the morning? I guess that's not the same price as when they closed the day before right? so that deviates from training for sure. I'm guessing this is a simulation but if it was real then you would need not to buy them in the morning
0.234% a week is pretty good, no? Assuming this can be repeated, which is really not a given :)
But and hold for 1 yr plus 1 day.
Nice SpongeBob reference :D
😂 glad someone got it
@@ritvikmathI laughed every time. But I can’t tell if it was because of how goofy your impersonation sounded 😂