Would you have a repository of your mathematica approaches ? Especially if those are illustrated in your book on stat. conseq. of fat tails. I did few on my own but struggled a lot. Thanks
I am really happy you are making these. The topics are spot on. I am having trouble following them though - maybe I just don't understand the math enough
Wow!! Literally spent the entire Saturday afternoon re-watching and re-listening, second by second and learning something brand new. Feel great, thank you.
Man absolutely love these videos. Do any of you guys know where I could learn more about what he is talking about slope differentials and entropy, like how .1 is closer to 0 than to .2?
Thank you for the video. What about the aspect of time and space relationship? It seems that the closer two things in time, for example, increases the likelihood for a correlation. For example when troubleshooting an issues, say programming, it's a good idea to check what was last done as it's likely to point to the source of the problem. When looking at the correlation between a tornado and it's damage the closer you are in time and space to the damage is likely the path of the tornado.
You're probably not going to see this, but would you consider writing and publishing a mathematics and statistics textbook that would teach the subject properly?
For real-life variables and distributions that show a lot of skewness, can the similar approaches for analysis be taken after bifurcating models for +ve from mean and -ve for mean data points? This are the kind of stuff I try empirically
Bonjour Monsieur Taleb, est-ce que votre ouvrage "Statistical Consequences of Fat Tails: Real World Preasymptotics, Epistemology, and Applications" va sortir en langue française ? Avec mes sincères salutations. Karim
Why do you act like p-values don't exist? The null distribution of Pearson correlation has already been derived, you can just calculate how unusual a given effect size is for a given sample size without having to do simulations. I do agree on the multiple comparisons problem, but criticising correlation for not always reflecting the population value is a bit silly?
I don't understand how we can have indépendante et correlated variables. If X, Y ind. by definition for every f, g measurable functions E[f(X)g(Y)]= E[f(X)]E[g(Y)] So for g(X)=f(X)= X-E[X] corr(X, Y) proportional to E[f(X)g(Y)] = E[f(X)]E[g(Y)]=0 So independant variables cannot be correlated.
This sorta thing I covered in a first year regression unit. It can be hard to believe that researchers can be that uninformed. You might have to showcase your rogues gallery of incompetent science in a compilation at some point. Cheers!
Technically when P is large you need to do some kind of correction for multiple comparisons. FDR controlling works best when P is huge( genome for example). You accept some false discoveries in exchange for capturing most of the true ones. In science, you can't publish without indicating how you control for multiple comparisons.
There was this comment that correlation between two random sets should be 0. Seems like there two different correlations with same name - one being mathematical correlation, other is "everyday correlation". Those are not same things, and mistaking one for other is the problem. There is nothing in mathematical correlation definition that requires correlation between random sets to be 0. Similar to how "Consistency" in CAP theorem and ACID DB properties are defined differently, but are sometimes mixed up because same word is used for both.
Great video as always. Always I have no idea what this means. If you have something that seems correlated say copper/gold ratio and the 10-year us bond. How do you know this isn't just a lucky correlation and actually meaningful and that they react to one another? Please someone with a brain help me out.
You can just use p-values with Fisher's transformation for example to test correlations (I don't know why Taleb pretends like those don't exist in this video), but perhaps time series methods like Granger causality in a VAR model are better for that kind of time dependent data.
This lecture misses an opportunity to discuss rank correlation such as Kendall's Tau which is the truly fundamental measure of dependence, separate from the shape of the distributions of individual variables. Pearson's rho tangles the effect of the marginal distributions with the effect of "dependence". (Best shown by sensitivity of rho to the location of an extreme outlier, whereas tau is unaffected.) Notice that I use the word "dependence", not "correlation". "Correlation" is often used in an imprecise an confusing way as in this video. Taleb never mentions rank correlation particularly in multivariate statistics (aka copulas) so I suspect that he doesn't appreciate it's power and utility.
Hi, I didn't get the last part. "Not having enough data, but having too many variables", I can intuitively understand, but what's the solution? Partial correlation? Partial derivatives?
The sum of reflected light, is directly proportional to the non readability of the blackboard. A cover or angled blackboard would increase the usable blackboard space by +60 percent. Are you saying CB policy is not correlated to stock market performance, but media says input A, gives output B. One moves, the other also. My world has crumbled, give me a new religion.
Its incredible how he gets tidied up for the videos with his cardigan but hasn't solved the simple problem of the chalkboard reflecting the light of the window that makes it hard to read.
Nassim, how is this any different from Hume's view that all causality is merely habit or convention (or arbitrary)? I can see you are arguing against the very common mistake of spurious attribution of causality to things that are only artificially correlated, but I don't see you saying what is the correct way to attribute causality? Yes causality is the key problem.. it seems to me you're just articulating the problem, not the solution..
Can you make a video about cryptocurrencies? Why is it a good way to go bankrupt? I like the principle of crypto, decentralized, less reliance on the state. I can even imagine a near future without the state, or anarchy (covid-20) where trade exclusively uses cryptocurrencies?
I am not mathematically literate enough to discuss the correlation between random variables. I am however literate enough to know that the “soft” “sciences” are our only hope. For example the science of “BS” which is being foisted upon the least literate of society will ultimately have a much more profound impact in the near term than any of the observations of mathematicians like Dr. Taleb. Big picture, you would have a much greater impact working to manipulate the feeble minded than to explain to the choir why all hope is lost. In other words, figure it out, use your freaking voice to affect a freaking positive outcome. Make your time horizon 100 years!
This line pure gold. “The point is not that correlation is not causation but very often correlation is not correlation.”
This the only channel for which I have enabled notifications. I don't want to miss even one of these...
everyone shoud follow your behavior
The parts that use Mathematica help tremendously with visualizing the concept. Thanks for these videos!
Thank you, Professor Taleb, these lessons are proving to be really helpful. Keep up the great work.
Maestro is so enlightening
I love that we got saved by the bell at the end of the lesson - this is like real school!
I sent this to a friend who didn't understand why I hate consultants/analyst or as I like to call them astrologers in suits.
A lot of data consultants are statisticians
I find these lessons interesting, they give me some thought food. Thank you, Nikolas.
Great insights professor Taleb. Thank you very much.
Thank God for this channel.
I hated Statistics until a friend told me about Prof. Taleb. I love Statistics, thanks Prof.
This is simply brilliant
Great video as always, Professor Taleb. Thank you.
You are are most intelligent arab I know.
Shukran!!
I do NOT self identify as Arab.
no no , never say that. He is a Phoenician
6:05 Thanks a Lot for this. I love this lectures.
Would you have a repository of your mathematica approaches ? Especially if those are illustrated in your book on stat. conseq. of fat tails. I did few on my own but struggled a lot. Thanks
I am really happy you are making these. The topics are spot on. I am having trouble following them though - maybe I just don't understand the math enough
Monte Carlo demonstration is very compelling.
Thank you for these brilliant uploads.
Today is a good day.
Yes
Indeed, now that NNT is doing these tutorials, he should also suggest some books on probability and statistics for us beginners.
Thank you for the enlarged fonts on Mathematica.
Wow!! Literally spent the entire Saturday afternoon re-watching and re-listening, second by second and learning something brand new. Feel great, thank you.
Great lecture, Taleb.
Thank you.
Man absolutely love these videos. Do any of you guys know where I could learn more about what he is talking about slope differentials and entropy, like how .1 is closer to 0 than to .2?
Thank you for the video. What about the aspect of time and space relationship? It seems that the closer two things in time, for example, increases the likelihood for a correlation. For example when troubleshooting an issues, say programming, it's a good idea to check what was last done as it's likely to point to the source of the problem. When looking at the correlation between a tornado and it's damage the closer you are in time and space to the damage is likely the path of the tornado.
You're probably not going to see this, but would you consider writing and publishing a mathematics and statistics textbook that would teach the subject properly?
Thank you for this lesson, it is immensely helpful.
Balfroni upper bound? I can't understand what NNT said at 11:52 please can someone help?
Bonferoni.
@@nntalebproba thanks, you're the man NNT!
Thank you very much Professor
The p-problem reminds me of "big data". Agglomerating all the data spaghetti, throwing it against the wall, some of it will stick.
Thank You Professor
Thank you kindly ✍️
Very well done.
"Psychology... and some fields that shouldn't. exist... like political science"
For real-life variables and distributions that show a lot of skewness, can the similar approaches for analysis be taken after bifurcating models for +ve from mean and -ve for mean data points? This are the kind of stuff I try empirically
What was the method to deal with data mining? (Something) upper bound. If anyone has a link, much appreciated
What about partial correlation. Does it help to spot these kinds of false correlations? Thanks for the great content.
Bonjour Monsieur Taleb, est-ce que votre ouvrage "Statistical Consequences of Fat Tails: Real World Preasymptotics, Epistemology, and Applications" va sortir en langue française ? Avec mes sincères salutations. Karim
that is because it calculate it as group of data, not as sequences .... you need to see and calulate based on event time/sequence
what should I study to know all of this?
"I dont know, whenever I here the name Greenspan I shutdown." -N.N Taleb
Why do you act like p-values don't exist? The null distribution of Pearson correlation has already been derived, you can just calculate how unusual a given effect size is for a given sample size without having to do simulations. I do agree on the multiple comparisons problem, but criticising correlation for not always reflecting the population value is a bit silly?
Is mathematica better than R???
I don't understand how we can have indépendante et correlated variables. If X, Y ind. by definition for every f, g measurable functions E[f(X)g(Y)]= E[f(X)]E[g(Y)]
So for g(X)=f(X)= X-E[X]
corr(X, Y) proportional to E[f(X)g(Y)] = E[f(X)]E[g(Y)]=0
So independant variables cannot be correlated.
You can have DEPENDANT and UNCORRELATED, which is the point.
It's Taleb's sloppy language that is confusing you. There are better teachers of this material.
@@nntalebproba It might be clearer if you introduced the precise definitions of "population" and "sample". Also it's spelled "dependent".
This sorta thing I covered in a first year regression unit. It can be hard to believe that researchers can be that uninformed. You might have to showcase your rogues gallery of incompetent science in a compilation at some point. Cheers!
Technically when P is large you need to do some kind of correction for multiple comparisons. FDR controlling works best when P is huge( genome for example). You accept some false discoveries in exchange for capturing most of the true ones.
In science, you can't publish without indicating how you control for multiple comparisons.
What about RSquare? Where does it fits ?
In a way, this is the reason why replicating an experiment is important. You get another draw of the correlation random variable.
There was this comment that correlation between two random sets should be 0.
Seems like there two different correlations with same name - one being mathematical correlation, other is "everyday correlation". Those are not same things, and mistaking one for other is the problem. There is nothing in mathematical correlation definition that requires correlation between random sets to be 0.
Similar to how "Consistency" in CAP theorem and ACID DB properties are defined differently, but are sometimes mixed up because same word is used for both.
Yes!!!
One of the most abused concepts in research. Karl Pearson would be sad. Thanks prof!
Thank you!
Phenomenal :D
The king!!!
Great video as always. Always I have no idea what this means. If you have something that seems correlated say copper/gold ratio and the 10-year us bond. How do you know this isn't just a lucky correlation and actually meaningful and that they react to one another? Please someone with a brain help me out.
From what I understand, correlation is more significant the more you have a simple model (few variables) and more data.
You can just use p-values with Fisher's transformation for example to test correlations (I don't know why Taleb pretends like those don't exist in this video), but perhaps time series methods like Granger causality in a VAR model are better for that kind of time dependent data.
This lecture misses an opportunity to discuss rank correlation such as Kendall's Tau which is the truly fundamental measure of dependence, separate from the shape of the distributions of individual variables. Pearson's rho tangles the effect of the marginal distributions with the effect of "dependence". (Best shown by sensitivity of rho to the location of an extreme outlier, whereas tau is unaffected.) Notice that I use the word "dependence", not "correlation". "Correlation" is often used in an imprecise an confusing way as in this video. Taleb never mentions rank correlation particularly in multivariate statistics (aka copulas) so I suspect that he doesn't appreciate it's power and utility.
Merci
Hi, I didn't get the last part. "Not having enough data, but having too many variables", I can intuitively understand, but what's the solution? Partial correlation? Partial derivatives?
The sum of reflected light, is directly proportional to the non readability of the blackboard. A cover or angled blackboard would increase the usable blackboard space by +60 percent.
Are you saying CB policy is not correlated to stock market performance, but media says input A, gives output B. One moves, the other also. My world has crumbled, give me a new religion.
You're my favorite Arab teacher, thank you, it was very informative
I don't identify as "Arab", so please...
@@nntalebproba You're my favourite Levantine teacher. I'd use a more localist adjective here were it not for the implications.
Share the notebook!
Yes, give me 2 days.
@@nntalebproba Dear Taleb, we are waiting!!! Thanks and regards from Chile.
Its incredible how he gets tidied up for the videos with his cardigan but hasn't solved the simple problem of the chalkboard reflecting the light of the window that makes it hard to read.
finally underestand correlatione
give us the code!
New camera!
I come here for the "friends"
Nassim, how is this any different from Hume's view that all causality is merely habit or convention (or arbitrary)? I can see you are arguing against the very common mistake of spurious attribution of causality to things that are only artificially correlated, but I don't see you saying what is the correct way to attribute causality? Yes causality is the key problem.. it seems to me you're just articulating the problem, not the solution..
Statistics superman
Can you make a video about cryptocurrencies? Why is it a good way to go bankrupt? I like the principle of crypto, decentralized, less reliance on the state. I can even imagine a near future without the state, or anarchy (covid-20) where trade exclusively uses cryptocurrencies?
HaHa, bias in research?
Far too much $$ involved to not get the "right" answer.
new camera
I am not mathematically literate enough to discuss the correlation between random variables. I am however literate enough to know that the “soft” “sciences” are our only hope. For example the science of “BS” which is being foisted upon the least literate of society will ultimately have a much more profound impact in the near term than any of the observations of mathematicians like Dr. Taleb. Big picture, you would have a much greater impact working to manipulate the feeble minded than to explain to the choir why all hope is lost. In other words, figure it out, use your freaking voice to affect a freaking positive outcome. Make your time horizon 100 years!
Thank you!