Quantopian Summer Lecture: The Art of Not Following the Market
ฝัง
- เผยแพร่เมื่อ 24 ก.ย. 2024
- Algorithms that do not follow the market are very attractive to investors. We will discuss some approaches for reducing correlation to a benchmark and discuss why returns aren’t everything.
This is the first in Quantopian’s Summer Lecture Series. We are currently developing a quant finance curriculum and will be releasing clone-able notebooks and algorithms to go along with this lecture.
Speaker Details
Delaney Granizo-Mackenzie will be presenting. Delaney is an engineer at Quantopian whose focus is on how Quantopian can be used as a teaching tool. After studying computer science at Princeton, Delaney joined Quantopian in 2014. Since then he has led successful course integrations at MIT Sloan and Stanford, and is planning on expanding to many more schools this fall. Delaney’s background includes 7 years of academic research at a bioinformatics lab, and a strong focus on statistics and machine learning.
Want to learn more about Quantopian? Visit us at www.quantopian....
Disclaimer
Quantopian provides this presentation to help people write trading algorithms - it is not intended to provide investment advice.
More specifically, the material is provided for informational purposes only and does not constitute an offer to sell, a solicitation to buy, or a recommendation or endorsement for any security or strategy, nor does it constitute an offer to provide investment advisory or other services by Quantopian.
In addition, the content neither constitutes investment advice nor offers any opinion with respect to the suitability of any security or any specific investment. Quantopian makes no guarantees as to accuracy or completeness of the views expressed in the website. The views are subject to change, and may have become unreliable for various reasons, including changes in market conditions or economic circumstances.
on the continued improvement model for strategies I think we have to keep in mind continuous improvement only goes so far and you have to remember the electric light was not made from continuous improvement on the candle.
The answer to Jason's question was actually wrong. As the return of Apple and SPY are not independent, so It would definitely affect the beta on SPY.
exactly
Thanks for the great lecture and sharing it.
hi Dan,
quick qn: how do we go from our Beta hedging to adjusting the weighting on the assets in the portfolio?
49:40 I am pretty sure adding a factor will change previous beta's (unless it's uncorrelated with previous factors).
example : if your first factor is SPY prices, and you add another one fully correlated with SPY prices (let's take SPY prices !) the previous factor is divided by 2.
+Delaney Granizo-Mackenzie Thanks for the detailed info. I appreciate.
In case it helps anyone else, I ran the test which Delaney Granizo-Mackenzie suggested in the video:
When run independently TSLA had a βSPY of ~1.93 as shown in the video, and a βAAPL of 0.5268. When ran as a factor analysis, βSPY is ~1.84 and βAAPL of ~0.11 over the example time period.
> generally in regression and factor analysis you want to regress over uncorrelated independent variables
My formal stats knowledge is middling at best, (thanks for this lecture series, btw!) but I think this sounds like a job for PCA? I'd think that would let you transform your set of benchmark assets into a set of independent variables (the eigenspace) upon which you could calculate your regression, then you could reverse the transformation on the estimated independent β-values to calculate your adjusted weights for β Hedging.
I agree though that even with perfect retrospective β-hedging you're still exposed to the risk of those β-values changing quickly.
Delaney Granizo-Mackenzie I certainly will. scikit-learn makes it all too easy. I think however that PCA is likely more useful in the context of correlation reduction, as described in one of your other lectures.
"Your algorithm will continue to make money if Turkey has a revolution"
29:30
Why does the notebook delineate between regression and correlation by saying regression is limited to linear dependencies but correlation isn't? Correlation only captures linear dependencies. x1 = x, x2 = x^2, these two are related, but the correlation coefficient will not capture this. I also don't understand why in the analysis of the regression model for TSLA and SPY, the alpha was completely ignored. The notebook concluded that the model just shows that SPY is more volatile, but doesn't a positive alpha indicate that there is some return component that isn't explained by the higher volatility of TSLA?
Thank you. You have helped clarify many things.
This is awesome! Thank you!
The 'when Chins tanks' seems so true in retrospect!
A correlation between the returns of one asset and another asset at the same period has no predictive power. It would be predictive if, for example, there exists a correlation between the return of one asset today and the return of another asset tomorrow.
When should beta hedging be done (generally)? Because though it seems ideal, we can't do it always to eliminate beta (as covered above- with historic datasets).
I wish Quantopian had an R api..... linear regression and classical statistics is so good in R
Good tutorial. Even a layman understands
Hi Dan,
By any chance would you be able to tell me if the code for OLS is available or not? If it is, can you perhaps give me the link to it?
Thanks and good work! :)
inspiring and awesome!!!
Brilliant presentation - learned a lot. BTW, just one year later Turkey actually DID have a revolution - how did you know? ;-)
Also been using statsmodels for a single ols as pd.ols depreciated but haven't yet found a substitute for rolling regressions (using window method in pd.ols) - has there been a quick and easy substitute made since?
Delaney Granizo-Mackenzie Thanks but I believe this is depreciated also?
Delaney Granizo-Mackenzie thanks for all your help. Enjoyed the chat with traders series 👍🏼
I was just going to write something similar to Thomas D's note.
why did you ignore the r squared value of 0.2 thats a terrible value, this means your model can only acount for 0.2 of the varation of your data this is terrible. Plus that isn't an F-value this would be a lot higher and you would preform a F test against your critical value which is relative to the degree's of freedom(which is on the table) if you were using that method to show significance, this p-value may have came from an f-test im not sure how python does it but it isn't the f value itself.
Great explanation!
Thank you. You have helped clarify many things.