The CLT allows anyone (including ignorant economists and psychologists) to do statistics by using prepackaged recipes coming from the Gaussian. What are the foundations? How does it work? Where does it not work?
Thank you for taking the time to make and publish these - I normally do not like to learn from videos but the density of the information in these videos is well worth the time investment.
I’m here after finishing my MBA to unlearn the shit I had to study to earn my certificate, thank you Nassim ! I cannot believe that people who spend a lifetime in academia or working with statistics never sat and gave the concepts behind the theories they work with a deep thought, the Incerto makes more sense to me with time
Good Morning Nassim, but the black & scholes do consider that the change of the stock price is a fat tail distribution, because they use the log normal distribution of the change of the stock price (gaussian) and not the distribution of the change of the stock price directly
I agree... I was reading a paper who claims that Student-t Distributions it's a better fit for estimate the daily returns of a stock than Black Scholes (underestimates) and Cauchy Distribution (overestimates). Do you agree? Which distribution is the best estimate considering the fat tails?
I usually keep my mouth shut because I am a fool. But I would like to say thank you so much for the knowledge. I will strive to get better to become less ignorant!
I did some work on Stable distributions (as the error distribution on time series models) on my Stats undergrad back in ~2011. One of the worse things was the very heavy computational cost to compute integrals (which have no closed form most times). There were some pretty good approaches based on the empirical characteristic function, which is somewhat uncommon for the regular statistician. Fun stuff.
Prof. Taleb, would you mind discussing the true meaning of impact factors? A colleague argues that they too are subject to power law. However, our discussion couldn't truly come to a conclusion! It would be greatly appreciated! Thank you for your videos!
Hello Nassim! Have been following the entire series, its great! Just a small request, do a technical presentation of key topics from "Statistical Cons. of Fat Tails" too. Thanks!
Mr Taleb - je vous remercier pour ces videos. J'ai q'un question: connaissez-vous la recherche sur la causalité de Bernhard Scholkopf et al? Qu'est-ce que vous pensez? C'est en quelque sorte frauduleux?
Discrete uniform distribution Bernoulli distribution (with p=0.5) Binomial distribution -> Gaussian distribution The first 3 have équiprobable outcomes Pareto distribution (observations are Not équiprobable)-> fat tailed Gaussian distribution
Thank you. I watch these from behind enemy lines where I work on exposing the frauds in the data science industry. My ex colleagues all privately admit they have never done anything useful in their entire careers. Few bother to ponder why
Are you familiar with phase shifts in thermodynamics? How water turns from a liquid to a solid to a vapor depending on various combinations of temperature and pressure? I've always thought that financial markets behaved in a similar way, where most of the time they are in a "normal" phase, but then under the right conditions of debt/income ratios and uncertainty, they can shift to a "panic" phase. Is there any relationship of this idea of a phase shift to fat tails?
Of course there is a relationship in so far that if you want to model the size of an outcome at a randomly selected point of time, you will need a fat tailed distribution of some kind to capture that really large deviations from the normal phase can happen (even in practice). But if you want to study the phenomenon more closely you should probably look into dynamical systems instead. Check out Steven Strogatz for example.
nice video although i don't see what the second example shows when it's just a special case of the first, maybe if p was not 1/2 it would have been more interesting
@@nntalebproba Thank you. I wrote that comment before watching the whole video. After watching the whole video, I understood "summation". The reason why I asked if "sampling" was implied was because the way I was exposed to CLT in the past was: the plot of sample means (from any distribution, assuming a fixed sample size) should show up as a normal distribution.
To invoke the CLT you need a large enough sample size. Today, bootstrapping (a resampling method) allows the statistician to ignore the CLT and essentially substitute computational power for theory. In bootstrapping the computer repeatedly takes samples from your data with replacement to come up with the true distribution that governs your data, no need to assume normality as per Gaussian.
"here the sum or average, same thing, converge to a gaussian". Well both are inaccurate. The latter degenerates to a point, and the first one one has variance diverging to infinity 🙂. Something important is missing ... 🙂