Trading bots are mostly scams... but there are exceptions. You just need a lot a patience to backtest and run many bots, discard the ones that are scams and keep playing with the ones that are good. EDIT: since so many of you ask, yes I use Galileo FX.. .
Good stuff mate. You missed out the impact of transaction costs, spreads, and slippage. This changes the analysis from a cross-sectional supervised problem like the one you've explored into a Markov Decision Process. The fees are not negligible and must be accounted for in any real trading algorithm. Additionally, back-testing the strategy on the same data you built the PCA components on is meaningless, since you are essentially overfitting. You need to split your data into train/validation and then backtest the validation portion.
Thank you for your insight, and thanks for watching! I did mention in the video, maybe I should have been more clear that a backtest is definitely the next step if you were to develop this into a trading algorithm; my previous couple videos I’ve built an environment to do that where trading fees are accounted for. And you’re exactly right, the validation should be on a time period that was not included in the fit of the PCA model. The point of this video was more so to explore the signal in some of these indicators, and to give viewers ideas as to what data is available to develop other strategies. To your first point, I agree, even a backtest isn’t a perfect representation of live trading. I hope to address some of the things you mentioned, like slippage, later on in the series. A couple remaining assumptions to test would be to see if a signal like this one holds up in a validation data set, and if so is it on average better than a 0.075% trading fee. I’ve explored some of this personally, but the video was already a long one so decided to end it there for now :)
He wasn't making predictions of future labels. He was describing the relationship between some components of his features and the labels. Instead, I saw this as more of a descriptive exercise. But, even as a descriptive exercise, I'm a little suspicious of the data. When look at the correlations between the coins, they vary pretty widely. So I suspect that the correlations between the indicators would vary as well. For example, I don't think the "momentum rsi" values for ADA are correlated to the "momentum rsi" values for BAT. But the way you've structured the data assumes that they share a distribution function. So, I think you've added unnecessary noise to the data. Anywayyyyyy, solid exercise and example. Subbed!
Hi lee, I checked your github code very last line (I think, It's an error). shade_plot(pcas,'pca1','pca2', 'label', 50) I think, It should be: shade_plot(pcas,'pca1','pca2', 'profit / loss',50) Thanks,
Mate your content is top tier, I hope the algorithm realizes this soon.
Trading bots are mostly scams... but there are exceptions. You just need a lot a patience to backtest and run many bots, discard the ones that are scams and keep playing with the ones that are good. EDIT: since so many of you ask, yes I use Galileo FX.. .
Wow man! Congratulations, really amazing work! Thanks for share!!
Good stuff mate. You missed out the impact of transaction costs, spreads, and slippage. This changes the analysis from a cross-sectional supervised problem like the one you've explored into a Markov Decision Process. The fees are not negligible and must be accounted for in any real trading algorithm. Additionally, back-testing the strategy on the same data you built the PCA components on is meaningless, since you are essentially overfitting. You need to split your data into train/validation and then backtest the validation portion.
Thank you for your insight, and thanks for watching! I did mention in the video, maybe I should have been more clear that a backtest is definitely the next step if you were to develop this into a trading algorithm; my previous couple videos I’ve built an environment to do that where trading fees are accounted for. And you’re exactly right, the validation should be on a time period that was not included in the fit of the PCA model. The point of this video was more so to explore the signal in some of these indicators, and to give viewers ideas as to what data is available to develop other strategies. To your first point, I agree, even a backtest isn’t a perfect representation of live trading. I hope to address some of the things you mentioned, like slippage, later on in the series.
A couple remaining assumptions to test would be to see if a signal like this one holds up in a validation data set, and if so is it on average better than a 0.075% trading fee. I’ve explored some of this personally, but the video was already a long one so decided to end it there for now :)
He wasn't making predictions of future labels. He was describing the relationship between some components of his features and the labels.
Instead, I saw this as more of a descriptive exercise.
But, even as a descriptive exercise, I'm a little suspicious of the data. When look at the correlations between the coins, they vary pretty widely. So I suspect that the correlations between the indicators would vary as well.
For example, I don't think the "momentum rsi" values for ADA are correlated to the "momentum rsi" values for BAT. But the way you've structured the data assumes that they share a distribution function. So, I think you've added unnecessary noise to the data.
Anywayyyyyy, solid exercise and example. Subbed!
Congrats on 1000 subs mate :)
(I'm one of them, ofc)
😄
Looking forward to the next video. Thanks for sharing 🙏.
How mention at the end that you backtested this strategy and it did pretty well. Mind sharing the results of your backtest?
Could you apply multiple soft SVMs on the output of the PCA clusters?
Hi lee,
I checked your github code very last line (I think, It's an error).
shade_plot(pcas,'pca1','pca2', 'label', 50)
I think, It should be:
shade_plot(pcas,'pca1','pca2', 'profit / loss',50)
Thanks,
Fixed :)
I am happy to be here.
excellent video, thanks
ill be your friend
Wow man! Congratulations, really amazing work! Thanks for share!!