Gerd Gigerenzer & Nassim Nicholas Taleb: The dichotomy of behavioural economics

complex models need complex problems or vice versa complex uh problems need complex Solutions so we might think and if our complex model doesn't work we'll just make it more complex and so on Albert Einstein once said make everything as simple as possible but not too simple today I will talk about our research at the max Blan Institute for human development about Simple Solutions to complex problems the key message will be that we need to distinguish between situations of certainty and uncertainty in Frank Knight's terminology between situations of risk where we know everything and we fine-tuning pays that means complexity pays and on the other side situations of uncertainty where fine-tuning does not pay makes the model more fragile and leads to surprises again and again so the key message will be there's a distinction between risk and uncertainty good decisions under risk do not transfer necessarily to good decisions under uncertainty and second we should engage systematically in the study of tools for dealing with uncertainty and I will focus on one of these tools namely heuristics a heuristic is a simple rule and heuristics have gotten a bad name in behavioral economics but because they've never really systematically been studied not formalized and not tested competitively so what I want to do with you today is to give you three examples of uh complex problems where simple juristic can do better than optimization models where they can lead to safer Solutions and to faster Solutions and then in the second part I will sketch a more General underlying theory that helps us to understand when less is more that is when ignoring part of the information pays when simpler models pay because that reduce estimation error and overfitting and at the end of hope that some of you will engage in a systematic study of tools for uncertainty and the others at least will have a sleepless night so let's start uh the distinction between risk and uncertainty goes back at least to Frank Knight it is been taught in almost every textbook in economics and finance and immediately forgotten the typical models are made for a situation of risk where a situation of risk is where we know all alternative all consequences and the probability distributions for sure in that world fine tuning pays and this is the world of the standard economics and finance models so if you go to the lot uh if you play the roulette this evening in the cin Cino you can be absolutely certain what will happen and everything that you need is probability Theory and you can calculate how much you will lose in the long run so but most of the problems we deal with have a certain degree of uncertainty and between risk and uncertainty is a Continuum and that is why we have to think about how much calculation and how much data we should ignore and in that world it's no longer true that you can calculate the optimal solution optimization is a fiction second it's no longer true that the uh the models that give you the best answer and the risk are the best answers for uncertainty and here by uh statistical principles of yeah that will explain in a minute uh we are often better off by simplifying and the key problem is how much to simplify and not too much as Einstein said it so I'll give you uh three examples from various domains to illustrate the point and then I go into the details that's the program are you ready then let's start is there anyone who plays baseball or cricket two people soccer oh okay it doesn't roughly a ball is coming in high an experienced player knows immediately where to run to catch or stop the ball how does he or she know it's a complex problem there's wind there's Spin and the typically answer is if it's a complex problem we need a complex model to understand it that's the version that I will not represent today the other alternative is it's a problem under high uncertainty and we need to see whether we can find a simple solution let's start with the complex solution what's that you know probably you might think that a player somehow calculates the trajectory of the ball and runs to the spot where the ball will come down have you ever calculated a trajectory that's how it goes and the problem is not computation the problem is estimation and that's very important to understand the we can compute today almost everything but not estimate for instance Alpha is the initial angle from which the ball was thrown or kicked and that would have been estimated in a second or two and you also May notice that I gave you a simplified version there is no wind no direction of wind and uh speed there is no spin so but how else could a player solve this complex problem a number of studies show that experienced player use a handful of simple humanistics I'll show you the simplest one which works if the ball is already high up in the air and it consists of three building blocks first fix your Gaze on the ball second start running and third adjust your running speed so that the angle of gaze always remains constant this player does exactly that observe how the player is running and the angle of gaze always remains constant what happens then the player will be exactly at the point where the ball is coming down we want to see it again the important thing is that this player can ignore all the information in the formula that you just saw and also he can ignore the information that I left out of the formula and focus only on one variable which is the angle of case this is an example of a heuristic from a larger class of one good re Ison heuristic and it shows you that it can solve the problem actually while the uh the trajectory computation model will fail not because of computation but of the estimation problem and theistic is also safer and faster if you have a dog at home and throw a fish frisbey then you will observe that your dog doesn't try to calculate the trajectory it will run in a way to keep the optical angle constant does exactly the same thing in our research we study experienced decision maker who often make decisions intuitively and work out the underlying juristic for instance a ball player makes these in decisions intuitively have you ever interviewed a soccer player mostly the player doesn't know but the player can do can solve the problem without knowing that's called intuition and we should not denigrate intuition so we study the intuition of experience e uh experts and then work out the underlying juristic and teach them explicitly so that others can use them to make faster safer and better decisions here is an example you may remember the miracle of the Hudson River a plane took off in laguadia airport in New York and after a few minutes something unexpected happened um flock of Canadian geese collided with the plane the engines of modern Chets are designed in a way so that they can digest Birds but not Canadian geese they are too fat and the impossible happened they flew in both engines upon which the engines shut down it became silent in the plane we know from the passenger reports that there was no Panic there was prayer the two pilots salber and SCS turned around and had to make a decision about life and death will we make it back to the airport or do we have to take a more risky alternative going into a river with a Jet Plane how did they make this DEC uh this decision so did they calculate and estimate their trajectory no they didn't they used the same urtic now explicitly and in the case of a sailing plane it goes like that you fix it the Tower of the airport through the cockpit uh windshield and if the tower goes up in your windshield you will not make it you will hit in below and that's can be done in a few seconds and left the pilots time to go into a checklist and checklist going through checklist is the opposite from aestic and the what we usually need is a good collabor ation between simple and complex tools so this example illustrates that first heuristics are not the stupid causes of all human disasters as you can learn in behavioral economics today second juristic can able uh better safer decision making By ignoring much of the information and in this case by reducing reducing the entire estimation problem and third one can teach heuristics explicitly and it's also incorrect that you read in textbooks that heuristics are unconscious no every jistic can be used conscious and nonconscious let's move on to the systematic study of heuristics and here it's important to follow three methodological tools first is formalize theistic even if it's simple you need to formalize it otherwise you will never find out how good or bad it is so it has become a tradition to use words like availability which are never formalized and which explain everything after the fact but cannot predict anything that's not my vision of heuristics um and second is after formalizing test how good they are and test them competitively typically against complex models and that also should hold for those of you who are in love with complex models test your complex models competitively against a simple juristic to learn how they are doing and third testing should not be in fitting your parameters to already given data but in real prediction of the future and these principles are often violated and particular much of success of optim optimization model is claimed by just fitting parameters to data that's not what you need that's hindsight so um consider the following problem you own a large company you have a customer basis of 100,000 of customers and you send information like cataloges or special targeted offers and you don't want to send information to customers who will never buy from you so-called inactive customers question is how to predict which of your customers will be active and which one will be inactive this is a highly difficult problem the usual answer to deal with a complex problem is a complex model and what I show you here is a typical example of such a model in marketing and it's called the Paro negative binomial distribution model it makes a number of assumptions which I'm not going through and again the problem is not the computation the problem is the estimation of its parameters from samples and as long the world is stable and highly predictable and you have large samples estimation is fairly okay but this is a problem which is highly outside on the end of uncertainty what is the alternative in one study uh the observation was made that experience managers rely on a simple heuristic to solve this problem it's called the Hiatus heuristic if a customer hasn't bought for 9 months classify as inactive otherwise active you can change depending on your business the time interval but it's again a juristic that pays attention to only a single variable now you might think the par negative binomial distribution model must be better because it has all the information and more I would advise you to do the test and two professors of marketing who believed in the uh uh that complexity needs complex models did the test and here is what I found there were three companies in the airline the Paro negative binomial distribution model the brown column made 74% correct prediction that's impressive unlike you test competitively and then they found out that the simple heuristic does better in the apperal business the uh advantage of the simple jistic was even larger and in the seed in our business it was the same you just saved all the time about estimating computation uh what you see here is called a less is more effect let's be clear what that means it means that if you have a certain body of data and computation as the Paro negative binomial distribution model uses if you take only a subset of this data and ignore all the estimation in that case or make it simpler in general you may do better not only faster and save money but actually make more accurate predictions that sounds against the common Spirit of big data analytics and it is and uh I'll give you in a minute the explanation why this happens I just want to uh call back so we know a number of situations where making it simple as Einstein said but not too simple actually pays so a final example that you probably know very well here uh assume you have a certain sum of money you want to invest it you don't want to put everything in one basket but to diversify how that's the question how much in each of n assets Harry marwitz from the University of Chicago got his Nobel priz in economics for the solution and uh that's the um mean variance portfolio or mean variance model I assume that everyone of you knows this and again the problem is not the computation the problem is the estimation of all these parameters and they increase exponentially with the number of n but Nobel Prize optimization model problem solved so we might think when Harry marwitz made his own investment for the time of after his retirement he used his Nobel prize winning optimization model so you might believe no he did not he used a simple juristic that is known under the name one n allocate your money equally to n assets so this is a juristic that ignores all data while Mart's optimization needs tons of data to do the uh parameter estimates so how good is one over n compared to Mark's optimization in one study s investment problems were studied such as one of them 10 American industry funds and allocate your money with Mark optimization you need uh many years of data in that example it was 10 years of stock data with one/ n you need nothing you're done what was the result according to standard um measures such as shop ratio one of made more profit than marit's optimization in six out of the seven problems again a less is more effect and uh so one when does this happen the real question is can we identify the environments where less is more or where more is better that is the real question we call this the question of ecological rationality I give you first some qualitative um characters so on the left side we have the situation of risk on the right side the situation of uncertainty and everything in between so if you are in a situation with low uncertainty that means the world is stable and the future is like the past and few Alternatives and high amount of money then make it complex Lex that's the world of complex solutions that the world of optimization such as mean variance but if you're in a world where tomorrow is not like yesterday or at least you don't know it uh when you have many Alternatives so many parameter to estimate relatively VI data make it simple that's the world of of uh tics and there is a large Continuum in between one can put this quantitatively one can ask the questions assume you have 50 um n equals 50 say 50 assets and uh how many years of stock data would you need so that the mean variance model has a good chance to get better than 1/n remember in the study I just reported uh the um there were 10 years of stock data so how much more do you think think the answer to this question can only be given by computer simulation but it helps to think so what do you think how much how many years of stock data would one need so that the markb optimization finally gets its parameter estimations under control and gets that we can expect it it's being better than 1/ n 10 years is too little 12 15 years 20 the best estimate is 500 years that means in the year 2500 people can stop trusting simple juristic and do the calculations provided the same stocks are still around in the stock market in the first place do our banks understand this important connection between Simplicity and uh an uncertain world and complexity and a certain world I'll show you a letter I got from my internet bank and the letter was sent to All customers and it's in German but I'll translate it says with Nobel prize winning strategy to success in investment and then the letter read do you know Harry marwitz no because Germans have no idea about Finance but he should and then they explained that he won the Nobel Prize in economics the bank is now doing marit's optimization a little bit late but anyhow and then there was a warning about your two simple strategies and your intuitions what this bank has not understood is that they sent the letter 500 years too early so let me now come to the second part and give you an idea about a more General understanding of when simple heuristics pay and when complexity pays so far I just gave you three examples why do people use juristic the standard answer to this questions in behavioral economics is because uh juristic are yeah simple and they need less effort less information but you have to pay a price in terms of accuracy your predictions won't be as well this is the few that you can still uh uh read in Caron's book thinking fast and slow and it is not correct it is only correct in a world of risk in a world of risk yeah heuristics are always second class but not necessarily in the world of uncertainty forly this view means that your total error of your prediction instrument has two components bias so a systematic component and noise some kind of unreducible error the moment you make actual predictions under uncertainty the situation is different uh it is characterized by the BIOS variance tradeoff well known in machine learning not so well known in economics and here there's another component that's called variance variance is basically the OV sensitivity of your algorithm to the specific properties of a sample or in other words if you take many samples from a population and estimate your parameters you get a variability across the parameters I'll explain this with a simple uh picture on the left side there's a dboard on the right side there's another one the the player on the left side has a systematic bias he is throwing too low and too much to the right the bias is defined by the difference between the Bull's Eye and the mean dart at the same time the player has a low variance on the right side there's a player who throws all over the board but this player has no bias on average the darts are exactly in the B's eye but only on average the variability is high so that illustrates that a model with a large number of parameters and a high estimation problem even if it's unbiased will lead to something on the right side predictions all over the Place depending on the specific sample and you can do better with a simple model that reduces variance for instance 1 / n has no free parameter so it will always give the same Dart throw so there will be no error due to variance only error to the bias Mark wids may have a smaller bias but it creates depending on the sample size and the number of parameters to estimate much more variance I think uh this is a a a good way to understand that the question is not to just look at the bias and make it as slow as low as possible but there is a a a trade-off between buyers and too much variance a simple juristic with no single free parameter as you just saw in the Hiatus juristic if you keep the 9 months constant or in 1n we'll have zero variance only error due to piers and that's how one can understand that one can do better by designing models that reduce variance not only Biers here's an example uh here you see the temperature in London these are 365 points average temperature and assume uh that could be U yeah Financial indexes uh it doesn't matter here you want to know what's the underlying you want to model this what's the underlying process and assume you think in terms of pols here we have a polinomial of degree 12 and one of the degree 3 so the uh simpler one is the green one the degree three question which of the two pols FIT the data better no shouldn't be so difficult yeah right the red one yeah it has more degrees of freedom and if you want to have a perfect fit use a polinomial of degree 364 but fit is not the goal of science nor of an investor you want to predict so what we do now is we take a sample of 30 out of this entire uh population and uh fit these pols on that or we can do we just take the fitted pols on the year 2000 and predict next year's and what uh you see here is that the error on the y axis H decreases the more complex the polom is and as we know now it if it just would go on at 364 Dee we will have zero the question is now if we now predict what shape will the curve have to make it simple let me ask two questions will the prediction curve so we take every every one of these pols and just saw this 12 degree polinomial fits better than the three polinomial that's what you also see has less error so we take every of these and predict uh will the prediction curve be above or below the red curve what do you think above yeah it's more difficult prediction it's more difficult and if you're socially intelligent you could have seen I left more space above so it must be there but what shape does it have is it also monotonically uh going down so meaning that the more complex model you have the better you will predict or has it a different shape what's the shape give me a a sign H it's go going up yeah you got the half the truth it's u shaped so what do you see now think about the 3 degree polinomial that wasn't as good in fit as the 12 degree polom in prediction it's much better it has less error and you may also see the interesting fact that a one degree polinomial which just a ly which is not a good model of temperature in London has a total error that is smaller than the great fitting 12° polinomial and that's mostly due to error so what does this show it illustrates that by making overly complex models that try to take care of every potential causal variable or whatever it is fre we risk creating model that system atically mispredict the future uh and the alternative is make it simple but not too simple not one degree but here like three so so let me come to the end yeah uh there are I showed you today three versions to make things simpler one is to use equal weights and do not even try to estimate the weights in order to reduce estimation error second one reason juristic such as the Hiatus and lexor graphis didn't show you but uh you can read about that so here's the key message situations of risk are not the same a situation of uncertainty despite almost all economic theories just pretend to deal with risk uh second heuristics are not a tool that explains why we are dump but there are useful tools in a world where optimization is a fiction Nasim Talib calls this the turkey illusion and finally uh robust robustness is a goal is the goal under uncertainty not optimization and complex problems do not always need complex Solutions we should have the courage to look out for Simple Solutions rather than to something that looks impressive mathematically but doesn't do well less can be more thank you for your attention um I'm honored to come back actually I don't was someone here did someone attend risk mind the first one I think was 1999 I gave a lecture in 1999 that's sort of similar to gird's lecture against Marco and I said I'm safe at any speed it was called and I'm going to use some of the graphs here and U and effectively all I'm doing is continuing gird's presentation because he sort of happened to be the only academic I agree with uh let me start you know monani has a saying that he prefers the company of peasants to the you know you prefer the company of the uninformed to that of the misinformed okay so we have this whole CL my War has been historically as a Trader and in 99 when I give the first Lex I was still a Trader so I was rested in jeans by the way now that I'm not a Trader I'm not allowed to wear jeans I said the uh the idea is to debunk pseudo experts and view the world the way we view it ourselves visibly if you are a pseudo expert you have one here so we've been living plagued with that expert problem people who think they can run their life your life people who think you can do things and effectively that was from sent to me this morning from the China Morning Post um it looks like we're in a situation where people are finally realizing that there is a generalized expert problem that effectively many of the people we call experts on paper look very impressive but they can't find a coconut on coconut Island and you can fool voters you know you could fool voters in 1960 you could fool voters in 1980 you could sort the full voters in 2000 2017 it's much harder because there's something called Twitter and also we're going to see why these uh experts aren't that expert let me take um a simple case you know I like to give mathematics in the morning to just wake people up up and then move back to more entertaining stuff so we're going to talk about scientific papers and this talk is pretty much about how your grandmother uh knows more about things that are scientific and is vastly more scientific in many instances than scientific papers because people have a feeling that if the paper is published in nature or in some Journal that it is true okay and they base their studies on something called P values okay so if you be believe in P value right after this I have two Bridges one in Lebanon and one in Brooklyn you know that we can negotiate later if you believe in P values but people tend to believe in P values and the thing blew up in Gs field psychology with his enemies when someone attempted to replicate a 100 papers published and say called the best journals in his field psychology in 2008 of The 100 papers how many replicated in your opinion 95 it was only how many do you think replicated sorry 39 39 replicated and of the 39 that replicated few had the same effect as originally announced okay so in other words it was true and and the reason is that people don't realize what P value means say I take a p value is a stochastic variable it's a mathematic this is simple mathematical application of you know continuation is that you think you have a number like value at risk it is not given by God for future experiments it is something that results from a simple sample so if you take the Ensemble if you repeat the same experiment many times effectively if you if the true P value is something like 0.12 12 12% 53% of the time you're going to have values below 5% and 25% of the time you're going to have value below 1% okay that's straight P value but there's something actually worse is that these experts what they do is not only they already have a flaw metric to tell us if something is statistically significant if the result works but they do something vastly worse they do a lot of experiments okay and take the best that's why they they get funding they're paid to publish papers they're not paid by reality they're paid by some Institute that gets funds from some Institute okay so you publish M trial is exactly like Fidelity in the old days that have 100 funds and then they show you the performance of the top five and somehow the funds would go into cyberspace okay they disappear okay the the one the non-performing ones survivorship bias the same applies here so if your effective P value is2 okay you can bring it down to 1% simply by selecting doing 12 experiments and picking the best of these 12 experiments so this allows us to debunk already some expert Problem by saying your grandmother has conveyed heuristics through generations so if she's better than 50% reliable she's better than the expert because the expert is not reliable we're talking about a a a a war against these experts they've been wrong on stalinism salafism Dynamic stochastic equilibrium modeling I'm sure many of you are familiar with that nonsense okay portfolio Theory Marco wi said is the Iraq wmd uh housing projects selfish Gene country risk recessions predictions all that lobotomies by the way okay so and you want us to believe him next now comes a deeper problem and that's the topic of my next book The Fifth volume of what I call the incerto the black one is the one people uh know me for hopefully they would forget about the black one remember antifragile and the incerto as a overall uh you know uh uh project and and that was what I showed in 1999 in that problem with the expert if you're not paid by reality but paid by perception how good a job you're going to do and say I have a portfolio to manage what do I do I show small gains okay and once in a while a huge loss okay in other words I'm going to short be short some out of the money option for a very simple reason when I make money I get something called a bonus some of you it's this time of a year so some of you are familiar with conversation about a bonus okay so you get a bonus when you make money and when you blow up what do you get no bonus all right do you have something called a negative bonus no no but if you trade your own money you have a negative bonus you see so that asymmetry okay created the whole industry of experts who never penalized for being wrong take the New York Times the New York Times really got us into Iraq with the wmd uh you know thing that they have to write a check for the losses we incurred for the people who starved in Iraq no even the financial no you take a journalist like Thomas Freedman he just wrote some another idic book he wanted us to go into Iraq he was wrong did he pay price no when he's right you know he collects a little bit of bonus and he reminds you of it so really this is what the problem is lack of skin in the game and effectively what skin in the game does to you is it is the only disciplinarian and it's remarkable that someone like like Professor gigerenzer coming from Academia realizes that people who have skin in the game put themselves I mean it's rare maybe 10% of academics understand the concept of skin in the game and only 1% go all the way to look at its impact because those who have skin in the game don't look for complexity what do they look for Simplicity because in the real world you're not judged by some supervisor who's going to write a report you're judged by your accountant okay so there is something you know that old parable of after huge labor the mountain gave birth to a mouse so if you're there are two worlds a world in which you want the mountain something very sophisticated to give birth to something as simple as a mouse and then the other one you want the mouse to give birth to a mountain in Academia is pretty much a b mouth give birth to a mountain so we're going to compare two projects in finance Marco witz versus Thorp I had after drinks with uh gerd last night visibly he Mak the case for Mart okay that Marist doesn't work in practice but works in Academia because it is very complicated as you saw his equation it works he is like a basketball player so we have torp I don't know how many of you have heard of Ed torp at torp is a fellow who wrote first the book called beat the dealer in gambling he was an MIT faculty and mathematician and he met the information Theory people and they started going to casinos and gamble okay and they made tons of money using very simple heuristics and his heuristic was in Black Jack instead of counting cards which you know you know the combinatorics are hard enough for a computer you just take the deck and you look a strong card you plus one and a bad card minus one and and when the deck hits something like plus five you start betting that's it so because you have the edge over the casino in some circumstances and a very simple heuristic that could be applied by anyone who knows to count till five you know how to count till five very good you can get rich so that was his technique and then later on he came to finance and effectively he popularized and formalized something called the Kell Criterion does anyone here use a clar caran but anyone who gambles his own money uses the version of the Kelly Criterion from Warren Buffett to all these people what's the Kelly Criterion here you have to know the future and the joint distribution of return at Infinity okay and with Precision because a little bit of failure of getting you know the Precision blows up your your your optimal allocation in this world all you need to know is two things expected return and how much you willing to lose and you go one step at a time you see every every morning you have coffee and then change your portfolio and that's a Kelly Criterion this is information Theory computer science as as he said again bias variance is well known in machine learning and machine learning is within that school of thought and here is people who Trend to sell either themselves right for something called the Journal of Finance which has never produced any article of note for reality but helps and tenure this is an academic world or the expert world the one that's exploding and as we speak across the world from China not China from India to brexit to the US brexit was mostly a vote against these people these sudo experts or these so-called experts so another case of so-called experts I went to work for the IMF where they do stress testing and I presented about two years ago that paper that technique to test tail risk was a very simp simple heuristic and they wrote a paper and they think they're going to get it in their procedure in 2028 okay to tell you how what it was too simple so they couldn't believe that it works it was very simple all you do is do stress test at 20% 22% 24% see if you have acceleration if you have acceleration you're in trouble if you don't have acceleration of losses you're not in trouble very simple procedure but they it's too simple for them so they're delaying the implementation although it's one of their it's was working paper that published in one of their books and then later on maybe make it to procedures another example I've been in a war against GMO genetic modification which is proposed on grounds of you know we're going to feed the planet okay again same pseudo expert they want to complicated solution we discovered the following if someone shows up in a doctor's office office the first rule we know since the beginning of medicine or since the hypocritic oath is first Do no harm so the first thing you do is try to tell him to have good night sleep if he has a headache all right where the expert will propose immediate brain surgery because it looks more scientific so likewise you know or you give them an aspirin if you have to do something you don't you know give do brain surgery if an aspirin can do or right so that's a simple rule in medicine that took a long time to implement in Western medicine because you had that that fear of not being scientific and now we understand that science doesn't look scientific science is heuristic and science is again in Simplicity because the more complex something is the more side effects it will have so the same thing with they want to do GM rice something extremely complicated okay so instead of give people rice and vitamins you want to combine them in something that has not worked but is very profitable would be very profitable for Monsanto and other firms so instead of doing this okay to solve world hunger just solve a simple Transportation problem the world hunger is a very simple distribution you know uh uh issue it is not something to do with production because most of the the the difficulties come from distribution of food so but no but it's too simple and we have shown actually that a banana is not the same risk as all these complicated genetically modified organism because when you do things naturally at low speed like you know or uh the Apple we breed we do effectively modify crops progressively but the way we modify them is by staying within what we call the viability Island and if you jump a viability Island all bets are off we have no idea scientifically where it can take us to you see so we were taking monstrous amount of risk actually for no return because was shown that basically these uh are more expensive so it's riskless returns but but it's very good for the experts because it looks more scientific so the other the area where there is a big expert problem as I'm saying your grandmother would not like GMOs and she's right another case of don't think that because the experts said they're good that they're good actually we noticed no tail risk study and mathematically if you know mathematics very well then you're probably at par with the grandmother you see to know that mathematically the risk metrics are not sound like mathematically uh marcoz anybody who knows mathematics would realize that Ma Marco cannot work work in practice under fat tals so nonlinearity Nature has mechanism by which it worked which bring us to the general problem of complexity okay when you you you know are trained in linear methods you think that you can solve problems with the tools you have but when you understand complexity Theory realize hey it's not that simple or it could be simple all right but you have to approach things differently using different methods so what the the the characteristics of complexity are first opacity you don't see what's going on in the system the second one is interaction between components they interact and that's quite Central and the third one is fat tales fat tales means of course things that you know I don't have to explain fat tales because I guess this is you know I've done nothing else all my life except study fat tales but we summarize f that rare event dominates the properties but now when we look at interaction you realize you cannot generalize you cannot generalize to school a fish from the observation of a single fish and this has pretty much sunk behavioral economics and let me explain how assuming the biases that we humans have according to these pseudo Specialists or Specialists are right assuming they exist they still cannot generalize to Market Behavior because the market is not driven by the average you know it is driven by people plus interaction so you cannot really see how things aggregate from studying individual Behavior so we can all be stupid and the market can be smart okay or we can be stupid on average and the market can be very smart so can't really claim Market had these biases based on observed biases for individuals and that's mathematically okay can be done mathematically now we're going to hit something that I think is uh uh quite consequential um that I've been thinking about it almost all my life but it was formalized by this fellow here I don't know if you see this fellow he's happy man he's g m Gman I don't know if you know who M Gman is he's an American physicist Nobel Prize winner he's probably the most uh decorated scientist alive uh he discovered the quars uh he is a very very very very smart fellow and he's one of the founders of the complexity uh works at at the Santa Fe Institute and and with a fellow they discovered the following so let me let's talk about something quite Central in finance and portfolio Theory all of us have this illusion of looking of thinking that if you look at average returns in the market and assuming that you have the right metrics that they will apply to you if you trade effectively wrong wrong wrong let's talk about P dependence and why it is Central um I don't know if you can try the experiment but let's do do the following experiment first you iron your shirt and then you wash it okay second case two you wash your shirt and then you iron it okay uh you get different results no ahuh so the sequence matters okay so we're going to see let's talk about sequence this slide actually I had been working on this I've been conveying this I've been giving metaphors about this but didn't hit me till I saw it treated by Mary Gman and and they he and his co-author effectively have Disturbed everything we know about probability in SPO by sh the following let's say that we have 100 people okay 100 people uh uh go to the casino all right 100 people number 29 goes bust you see when he's bust we put him head down okay he's ruined what happens to number 30 if number 29 goes bust will he go bust no and then you can take the average return from all of them and let's say the casino gives you the edge you can take the average return of all of them no divide and so on Reduce by the person who went bust this is called this is what we do it's called State space methods L State spased methods everything in finance is based on these methods which is basically to take the expected return all right from from this unfortunately if you take any Speculator and make him go to the casino 100 times and on number 29 but it's day 29 not speculated number 29 it's day number 29 the Speculator is bust head down what happens to day number 30 there is no day number 30 okay so this is very simple here so that this is called the Ensemble average and this is called the time average they're not the same any Speculator subjected to ruin will eventually be ruined it's only a matter of time it may take 10 years five years or a billion years but it will be ruin okay so you got to look at stopping time of ruin this we know as option traders that all these analysis of option selling don't count because eventually if you are ruined you're going to be ruined Okay so this is what we call so there's something for these two to match you need something called ergodicity and this is only satisfied but what we saw with my friend you see sry Marco witz the ugly guy versus the uh better you know fit guy at torp Kelly Criterion is erodic it's a strategy that allows you to capture the effective Alpha of the market so your Alpha matches the market only under condition of absence of ruin so this is quite annoying actually classical risk Theory got it okay and it's quite annoying because it disturbs a lot of things we do about valuation okay valuation you know by taking Market return doesn't apply to valuation to you individual if you have what we call Uncle point you take the alpha of the market if at some point the value drop some level and you panic because hey you know what you don't want to take additional risk and age 65 people retire in Europe no age in America is age 87 and a half where people retire but okay but you're retired now you depend on retirement income and you want to keep buying the same the Cuban cigars and this is a level you can no longer buy Cuban cigars so you cut your position guess what you cannot Alpha will no longer match the market you see particularly if they are dips and stuff like that so basically this is and what what muray Gman who's a great mathemati great physicist great scientist great mathematician basically he said except for the information guys all you guys Marco Marco Marco all these things and everything that has been done pretty much in L State space method is bogus I'm not he he said it he said it doesn't doesn't work and a lot of people have made that mistake he named them you know decision science dis and Theory so now let's talk about psychology and psychological experiment basically what people don't get is that if your grandmother is paranoid about small risk you cannot tell her no it's only a small risk for the following reason you can never analyze risk taking as one shot experiment the way psychologist have done it they tell you well do you want this bet this bet oh he overestimates small probabilities you can't for his following reason you're always exposed to negative probabilities okay and if you take risk say do this for on a weekend okay but at the same time uh smoking cigarettes and at the same time having knife fights with people in a mafia you know once a while visibly you have to ad upt those risks and what happens to you when you add up all these small little risks we go bust so the way to analyze risk taking is not by taking one static thing but over time assuming how many tail risk you're going to take over your lifetime your remaining balance of Lifetime not a single thing and before M GMA nobody formalized it every Trader knows it every single person who partook of the no ruin thing in Insurance knows it but it's not in the finance literature that's a mother of all expert problems and if you take behavioral finance and biases it assume the agent will never again take the same risk but we take these risks so if you do it that way and assume risk is cumulative then basically there's absolutely nothing irrational about worrying about GMOs and as people want us to P people want to pathologize us for worrying about GMOs this fellow works for Obama and I'm so glad so glad he can still work for Obama but with no impact he actually incidentally is an enemy of good no I mean let's say academically he's an enemy of G he's personal enemy of mine because I'm as you know as you can see the difference ger is much more polite than I am all right so and uh so and and and this guy says oh precaution is irrational because it's unscientific but when you think about it you do real science you realize that effectively effectively we have to be more and more paranoid as we go up this ladder so you have to be paranoid about yourself but that's okay because if I die it's not the end of the world okay I only lose my remaining life expectancy but if uh family friends and pets it's worse my gene pool is worse the tribe is worse self-defined extended tribe Humanity but ecosystem aha so you're losing a lot so the more you go up the more you have to be paranoid so when we are worried about things that impact the ecosystem it's absolutely not subjected to rational decision what we call rational decision making you should be a no no and that's where your grandmother intuitions are right so if we reframe things using this background and I hope for for for hopefully I'd be invited again by gird to go to a center and then try to do work along these lines and we can see exactly where all these analysis of decision making you see are wrong because taale probabilities matter overall not just locally all your life if you you see and though the only way to analyze things is by looking at how much it reduces your life expectancy and if you think life expectancy is an extra my life expectancy is an extra 30 years 40 years then I can take some risks but I don't want to reduce the life expectancy of my dream pool and I definitely do not want to reduce the life expectancy of the ecosystem and every time you take a risk in the system they add up so it allows you to be aggressive in some areas and not in other areas so I'm going to stop here because there are a lot of things to discuss and I'm taking a bonus of about 30 seconds G for the conversation so I'll continue that so we can we can I can uh cut earlier about 30 seconds earlier it's never happened I think to me before uh to continue uh but thank you for listening to me and hopefully uh you will be more skeptical of the experts leaving this room than than than you've been coming in thank you um this is a short discussion section and you are invited to just pose questions and maybe I'll start with one a question I got um the when I mentioned the dog and the Frisbee and uh uh one of uh the colleagues here pointed out that that was the title of Alden uh of Andy Halen's talk the Jacks and H talks in 2012 this is correct and in fact it was Ming King then the governor of the bank in England who put uh Andy and myself together early in 2012 I explained him and the first example I gave to him was the gay juristic that dogs use and I explained him the mathematics below that and Andy Halden uh reacted differently than most people who say oh that's interesting yes wonderful let me go on with my business he said okay now I understand it he did some his own research on that and when we met the second time he said this is my Jackson Hole talk I wanted to give here's the waste paper basket that's it yeah and I will now give a talk on uh these ideas that we have started to work together so how to make the financial system safer by making it simpler and the analogy that he uses in this Jackson hle talk titled the dog and frisbee is the dog how a dog solves a complex problem and under high uncertainty and that could be something that the financial um uh at least the practitioners if not Financial Theory could take seriously and investigate rather than ignoring it and carrying on with the optimization models that nasib is all always uh willing to bash them down I mean they bash themselves one comment I want to make a connection to skin in the game that I didn't quite make tight during the converation during the lecture is that what why why do I uh Advocate SK in the game is it not a disincentive okay it is an evolutionary thing is for the following reason if you drive on a highway any nut can kill 30 people by just uh uh you know going wild but why doesn't it happen what it doesn't happen because these people are dead you see they're dead so so they're no longer because they have skin in a game they tend to exit the dream pool when they make a mistake yeah you see so so this is where so so what fil so we have an evolutionary thing for things to filter things that survive are things that have robust rules robust and not necessarily the simplest but the most robust and way that they can survive different environment and they don't lead to ruin so this is the connection between SK so this is why skin of the game for me is monstrously necessary in in in for example you take hedge funds and by now about 60% of the assets of the big boss at the hedge fund who's effectively the risk managers are in the fund M that's skin in the game and it doesn't doesn't probably make it much wiser it make make them a little wiser they can't hide risk but it eliminates it makes them bankrupt go bank so uh nothing then to take the deutche bankk so how do we get how we change the system so that the managers have skin in the game you don't want that the system has changed automatically because all the risk taking have moved away from Banks move to hedge funds it's like people tell me how do I organize Society to have skin in the game but get rid of of Obama all right less centralization more decentralization the German Way where people who make decision live in a community MH like in Switzerland so so in other words it's much easier to have a structure that is um uh that that commodate skin in the game then try to force skin in the game in existing structures such as dob so are you saying that the solution could be or one of the solutions only have hedge funds and all Banks Banks just do a traditional banking but that's exactly what's happening that's that's exactly what happen the the the the risk taking has migrated progressively into hedge funds for several reasons but and hedge funds tend to have skin in the game and this is why you have a and the Ecology of hedge funds we have about I mean I counted one year where we had 2,800 hedge funds closing without one of them making making it past the 10th page of a newspaper okay so you realize Okay so uh so so that that you have just like restaurant business the restaurant business has a an excellent Ecology of people coming in and people leaving and that skin in the game allow effectively acts as a filter so yeah so let's maybe let's uh give the audience a chance to ask questions to PO talks there were some common uh ideas so the uh idea that complexity should not be fought with complexity always yes please a microphone is coming thank you yeah at the moment there is a big push um towards machine learning especially in the risk models you only have to look at program of this conference and all sorts of uh student projects that I supervise all the time and of course that yet is another jump in complexity and involved solution to a problem of let's say credit scoring and so on so this development is very very strong and it goes completely against of what you are advocating and I'm with you on this but I want to know uh what do you think about it and what is the response you would give to uh people trying to push those very complex nonlinear machine learning methods on this for example credit risk problems or Market risk problems or any other risk related issues MH so um the answer is that if the problem like predicting U credibility of debors if you're in a situation where the you are in a stable world where things are like tomorrow like yesterday where you can actually have enough information to estimate your parameters then there is a chance that the complex uh models would outdo simple rules but I suspect that is exactly not the case in this problem and I would like to see empirical studies about uh cred credit worthiness so for instance we know that experienced Bankers often just look at two things whether if you apply for a loan in the simplest case yeah versus filling out all these pages of information and then running a logistic regression so one needs to have an empirical approach for that rather than believing in complexity and also there is something else uh in my OB ation uh many reasons for using complex models that most people don't even understand but they're impressive because they form okay is uh one reason is defensive decision making defensive decision making means you as a manager protect yourself against being criticized or being sued and this is your first priority so not to to actually do the best for the company but you do do something second best that protects yourself and often relying on Big Data an analytics that many managers do not really understand but they buy it in order to protects if something goes wrong you might be accused why didn't you do this so and defensive decision is is a big problem and uh in my uh analysis of large corporations about u a third to half of all decisions are made huh uh defensively that means that the managers do not follow recommend the best thing for their company but something second best to protects himself a huge waste of money time and intelligence yeah that's a skin in the game problem exactly because let's take a very simple thing outside Finance if I have a doctor to go to a doctor and um what's he going to do you see that graph uh hidden risks delayed and benefits we know that Lipitor if you're don't have hyper choler Aria or all the blood pressure thing if you're slightly hypertensive that these harm you more than they help you but the point is he's going to prescribe these things to you because if you have a heart attack you can sue the doctor but if you have if you have delayed estrogenics means side effects later okay it is so so he doesn't pay for it and so what happened is that they follow not the same the optimal situ ation but the the one that optimizes their legal liability okay and so this is the the the situation is effectively remedied by skin and game and I think gerd had a comment has never asked a doctor what what you should be doing ask him what he would do if you were you he already change one layer of you know mental skin in the game by ah you know I would do this or if if you uh don't ask your doctor what he recommends to you but what he would do if it would be his own uh son his own brother his own wife and just to add a figure to what Nasim is saying um the problem about skinning game here defensive decision making or defensive medicine is depends on the environment it's particular a defensive prone if you live in the US and in one study over 800 American doctors were asked whether they practice defensive decision making typically uh giving you too much uh antibiotics too many unnecessary cancer screening methods or Imaging techniques which also bring in money and U What proportion of American doctors said yes I'm doing defensive medicine so in the study 93% and it's probably an underestimate because not everyone admits it to the other one or even to him or herself so here's a a huge psychological problem that is also relevant in finance because many people who are decision makers defends elves and are R willing to sync their own company it draw to my attention in your in your presentation about the certainty uncertainty and you said there is a somehow degree of low uncertainty in in my understanding that the risk we defined it as the probability of occurrence that the level of uncertainty should be should be the ending line of the certainty that when you have low low uncertainty it mean that you are more certain so there should not be somehow the uncertainty is the ultimate end if it is we are living in a in a word of certainty then there will be no risk the risk is that this is the level of the certainty we don't and the Simplicity you have said out of the experience don't you think that the exper experience in the compilation of the complications and the analysis and all the formulas that we have learned and the anybody would learn by intuitive or by education that to make his last decision simpler or even hortic so this is the what I'm trying to say do you measure the level of uncertainty or you level the measure of certainty okay and your risk so uncertainty has many more Dimensions than certainty you may not know what the alternatives are so if you want to marry there's no way to know what the alternatives are you may not know what the consequences are even if you know all Alternatives there are surprises out there and uh even if you know if your world is so small that you know all consequences that can happen and all Alternatives you may not know the probability distributions you may get in problems that Nasim has been talking about that there may be fat tails but you cannot really estimate them so under these conditions by definition you cannot optimize and that is so self-evidence as almost everyone forgets it about immediately and one one way to deal with these situations is to use to a certain degree heuristics and and emphasize one needs to study them seriously taking heuristics seriously and also taking un certainly seriously that has not been happening for instance behavioral economics criticized neoclassical economics about Homo economics who wanted to make it more psychologically real which was a good intention but they didn't there and it ended up uh for instance in the in the book by saor and sunstein uh the uh as um homo economicus is the ideal for us omnison and we are Homer Simpsons and there's no idea of critique on the model any deviation between the model and what you do the blame is on you never on the model and we need a new behavioral economics that dares I have one comment on on this is that often when you hear people say that we humans are irrational it's a person not knowing what the hell they're talking about okay so this is where go back to if the expert they that sunstein I disc had a little thing going where they use they don't know enough probability to be talk talking about probability that's one first comment the other one I'd like to comment on something I I picked up to connect things with with to Titan things you remember I made a comment that probability in time is not the same as probability in space okay and that's critical it reminds me of a comment I heard you make 15 years ago in Munich uh that was but we had all beer so comment the first time I met good gig renzer he was saying that if you ask a Frenchman what is the pro what does the probability of rain 30% probability of rain he would tell you 30% of experts think that it's going to rain if you ask a German he would tell you that 30 days out of the year is going to rain and if you ask an Englishman he would tell you it's going to rain 30 30% of the time during any given day right so and we see all of these are can be called probability so the word probability itself is not rigorous enough to describe many phenomena that you have to attach what you mean by probability and effectively in finance they didn't they and it's again it's not my idea it's to go back to that giant muray Gman it's so much easier when I present ideas I agree with and and dress them up in a giant who knows definitely more math than anybody in finance and knows more about physics and go back to physical probability because he's he's a big Quantum mechanic guy uh the what Gman said someone talking about probably in time talking a complete different animal that only rarely coincides to probability in space yeah we have another question I'm johanes fory German s Banks Association when professionals let me not call them experts for the moment for example make statements on rare events can we safely assume that there is some heuristics underlying so that it's worth following your program like formalize it uncover it train the experts afterwards or is there something um uh which you might call just statements from the gut which maybe are less valid than those based on tics I mean and how would they differentiate the uh so if there are rare winds that are so rare that you really can have good statistics then you should say that and this is part of the world of uncertainty I think it's an illusion uh if you use probability models on that thing and that's what uh uh N I think what do you call the turkey illusion tury problem so meaning that in the SAT of uncertainty you you still use your probability which is a theory of certainty probability yeah it just works if you know the the probability distribution period yeah and uh so that maybe the I I think it's much more caution and modesty necessary and there can be htics that you might use you might research for these situations rather than probability models that give you a number and suggest an illusion of certainty it's the same there's a classical thing I've noticed in my experience and U and and it's pretty pretty much you see uh uh and we did some kind of uh test uh not testing but counting incident if you have a PhD in finance you uh you stop understanding something that a peasant would understand which is a difference between evidence of absence and absence of evidence so the minute someone have a PhD in finance they cease to understand it okay some people may have a may you know but if you have a PhD in math maybe not you see so there's something about this half knowledge they believe and they start talking about evidence-based all right so there's there's evidence of no harm or no evidence of harm so mixing them the two are very very divers so that's the turkey problem so effectively I've noticed let me give you a very simple example um in 1999 we looked at every single fund manager who blew up in 1998 and we had a list of them uh and then we had uh professors of Finance who went to the or economics who went to the to the hedge fund sector and how many of them blew up in 98 we had 39 names 38 out of 39 blew up okay 38 out of 39 professors of Finance blew up blew up means you know experience an ltcm like draw down except for one and and and and and tells you that it's not you know there some exception and then the other thing also we noticed that when we took the one who didn't blow up okay the one all ones who didn't blow up they were either mathematicians or just uneducated uh uh you know uh fellows so the uh I mean I won't disclose the names now except you know for but we didn't publish it as a study so we have to show the names but you can do it yourself count how many uh you know the ltcm like firm like were there and and you can figure it out there's 600 um units that blew up and you can go from there so it tells you something about about this turkey problem generalized turkey problem that exacerbates with half knowledge yeah there are two more questions over there from group uh I I'm I would be curious about your comments on the the regulatory evolution of internal models so there are lots of discussions now going on is this more a trend now towards hortic what basy is trying to do is going back to standard or what the SSM is actually trying to do with a lots of Publications and kind of micro steering internal models and and getting very much into the details so how would you comment on this Evolution so the um the uh I think Aldi Andy halan has done a great job with his uh Jackson Hall talk to get this issue into this um as I see this so uh the the continuing increasing complexity so Basel one was about 30 pages Basel 2 over 300 Basel 3 over 600 what's coming next we do not know uh the uh the Bas three people at least use now the term Simplicity in their papers uh to which degree that really translates is another issue uh I uh one should I think uh one should realize that the uh take the value at risk uh computations that come from a world of risk not a world of uncertainty a large bank has to estimate thousand of risk parameters a co-variance matrix in order in the order of a million this is mostly guess workor and at the internal models uh and the parameter estimates the border on astrology and that has not prevented not even foreseen any uh uh crisis but it's still done in the absence of an alternative but the Alternatives exist and I didn't have the time I work with the bank of England on designing very simple rules that could yeah get in more safety in the system and I think there is an alternative and in part it is not uh gone or there is still resistance because of this illusion that you need complex models to deal with complex situation it's not true you need sufficiently simple models to have a robust solution and also The Regulators can see more clearly what the banks game it or try to game it than with uh millions of estimates I one comment people don't think don't think that he's saying or that you know in general it would be true that given that we need heuristics that all heuristics are good for the sake of heuristics don't think that that's the first one I mean a lot of people have bad heuristics in in in their practice or start using think hey I got to use theistic and then blow up that's the first one the second one the problem with regulation is again without skin in the game you you you can always beat the system okay and uh when I was I was a Trader for 21 years and I think that my career would have been only 10 years had it not been for Regulators because every time a markets get regulated you hire a lawyer and try to find a way to try to make some money because you know when it's regulated it's an opportunity like uh to you you like you you have regulation you can't go short uh you know on an upt on a downt or something like that then you create a structure that allows you to do that so you start having arbitrages simply coming from regulators and I thank the regulator for having financed half my career okay so I tell you this is they get shocked with tell them yeah you giving money to people who know how to you know maneuver systems and but all you need is a good lawyer so you finance so Regulators Finance good lawyers and finance traders who know that Japan is a market that's over that was over regulated and in Japan you can you know with a good lawyer you can do anything I think our hosts get uh nervous yeah so they want us to leave now but thank you thank you very much for this uh thing and thank you for listening to us

Channel: RiskMindsTV

Views: 41,121

Keywords: Gerd Gigerenzer, Nassim Nicholas Taleb, RiskMinds, Behavioural Economics

Id: 4VSqfRnxvV8

Length: 89min 12sec (5352 seconds)

Published: Wed Dec 07 2016