simply because at the tail all the components are directional less mrt. Hence the conclusion.
@jan7356
Жыл бұрын
Wow. I can’t believe you shared the whole code for this for free. You should work on Wall Street in a big quant fund. And hopefully run your own fund one day. I am not a crazy expert, but I think the stuff you do is top notch quality compared to most other stuff I have watched on KZitem, or seem in books. I think you have great market intuition and great quantitative skills. Hoping for more content. 🙏
@MrBhavishya
Жыл бұрын
Please create more video we like to watch new concept .
@andreasklippinge9665
Жыл бұрын
Would be interesting to implement this model on z-score or pure volatility. Great video!
@SliverHell
Жыл бұрын
Would love to know if you use any semi-supervised techniques for a group of the same indicator, with different windows? Such as NCA?
@neurotrader888
Жыл бұрын
I haven't used any semi-supervised methods for component analysis. I would be cautious about overfitting. Finding the components with the help of the label and then fitting a model with those components may 'learn' too much noise. But I'll add that to my to-do list. It might work. Occasionally principal components can obfuscate the information that is actually predictive, because the predictive information isn't on the axis of maximum variance.
@SliverHell
Жыл бұрын
@@neurotrader888 The cost function for the algorithm, if we have NCA for instance, just needs to have an appropiate regularization term to prevent overfitting? And also, using the entire range of the indicator, as features, would be probably induce the model to more noise.
@ademolaorolu5930
4 ай бұрын
This is a great one!. Thank you. By the way, if you have an academy or tutoring class, I will love to be a memeber.
@santalaszlo6858
Жыл бұрын
Thank you ✌️
@cadebruce4401
6 ай бұрын
how do you think about doing this versus simply ridge regression with all the lookbacks? It should be fairly similar, no?
@poisonza
6 ай бұрын
imo, smoothing is quite nice! great job. pca and other stuff is quite common. * extreme values offer predictive power... true in this case(rsi). but for some indicator its just lack of sample and perhaps luck so we never know...
@henrykim8938
5 ай бұрын
Could you please give any references of this video? except for just definitions of tools you used in this video.
@matthiaswiedemann3819
7 ай бұрын
Great video! Maybe make something about Monte Carlo Tree Search as well one day. One indicator with different lags for PCA is a super idea 😀
@CS_n00b
4 ай бұрын
Isn’t it bad practice to use the test data for so many different model parameter combinations?
@cuteandfunny9154
Жыл бұрын
what major did you take? im asking cuz im kinda lost rn in my 2nd year of studies
@homealone75
Жыл бұрын
Love your content! Keep it coming!
@mozkhiyar9486
Жыл бұрын
Borther another great video❤️❤️. First thing i do when i get my paycheck is the Patreon ong 😅😂
@saeedrahman8362
Жыл бұрын
Thanks for the content. Just out of curiosity , what motivates you to share this information ?
@favian1622
5 ай бұрын
Is there a quick and dirty way of knowing how much correlation is enough correlation to use this method? Expanding on this example, based on some data I saw the RSI 250 has a 0.18 correlation with the RSI 3. Would this also be sufficient to PCA? I would guess yes, and just play with the number of components. Really appreciate the content by the way. Been going through the vids, and ordered some of the recommended books.
@neurotrader888
5 ай бұрын
I wouldn't worry about 'enough correlation' . PCA finds linear combinations of features that maximize variance. Imagine you did PCA on 3 RSIs with periods of 3,4 and 250. The 3,4 period will be highly correlated. The 250 not so much with the 3 and 4. Of the first two 2 principle components, one would be a combination strongly focused on 3,4 and the other on the 250. In other words if the features are already uncorrelated PCA isn't going to do anything.
@martinsandor707
2 ай бұрын
Great in-depth explanation for the theory behind basic machine learning models! However, checking the timeframe used to train and test the model shows how doing a simple buyhold would have yielded much more profit than the model you trained here. While using such extreme quantiles ensures that the model would only extremely rarely (if at all) be wrong, the infrequency of the trades make it far from profitable. In what ways do you think the model could be improved, if we wanted to make it more profitable?
@neurotrader888
2 ай бұрын
You're not wrong, but total return is not everything. Many (me included) prefer to measure the quality of a trading model with a risk adjusted measure. Sharpe, profit factor, etc. While the model shown trades infrequently the trades are of decent quality. Imagine having 10 models with infrequent trades of similar quality. Put them all together and then its not so infrequent anymore.
@jonahkeller8832
Жыл бұрын
How could one go about automating this system to take future trades? How long do you hold or short the position till? What indication is given by the strategy to sell? Also, I think the eigenvalue plots may be dependent on the sample size, not necessarily reflective of the actual data. I replicated your code with Ford trading data and got very similar (identical) eigenvalue plots.
@neurotrader888
Жыл бұрын
Use in realtime would be similar to the walkforward section of the code. A trained model would be used on the most recent bar (The code from lines 76 to 88 at 8:20 in the video). You would need need to gather the most recent data somehow and use whatever exchange's API to execute the trades. Not financial advice, this strategy is risky and not complete. The signal output represents a percentage allocation of the strategy for the next bar. 1 meaning a full long, 0.5 meaning a half long, 0 meaning flat, and so on. The risk management used will dictate what a "full long" means. I think the correlation/covariance structure of the RSI is probably similar across markets, its the same calculation after all. I view the eigenvectors being stable as a good thing. I also got nearly identical eigenvectors from other crypto pairs than BTCUSDT.
@jonahkeller8832
Жыл бұрын
@@neurotrader888 In the video I noticed that you said the positions are meant to be held for the lookahead window. however, when the profit factor is calculated I think it assumes positions only last 1 period. How long should positions actually be held for?
Пікірлер: 31