Suppose we know someone’s objective and also know that half the time that person correctly figures out how to achieve it and half the time he acts at random. Since there is generally only one right way of doing things (or perhaps a few) but very many wrong ways, the “rational” behavior can be predicted but the “irrational” behavior cannot. If we predict the person’s behavior on the assumption that he is rational, we will be right half the time. If we assume he is irrational, we will almost never be right, since we still have to guess which irrational thing he will do. We are better off assuming he is rational and recognizing that we will sometimes be wrong. To put the argument more generally, the tendency to be rational is the consistent (and hence predictable) element in human behavior. The only alternative to assuming rationality (other than giving up and assuming that human behavior cannot be understood and predicted) would be a theory of irrational behavior—a theory that told us not only that someone would not always do the rational thing but also which particular irrational thing he would do. So far as I know, no satisfactory theory of that sort exists.
David Friedman, Price Theory, An Intermediate Text
Particularly, what we know about cognitive biases precisely is “theory of irrational behavior” — we just don’t have a, complete, theory of irrational behavior.
Friedman continues, but I shortened the quote to make it punchier. Essentially he says that, (1) given a large number of individuals irrationality will average out in the aggregate, (2) In most cases that an economist would be interested in (eg. investors, CEOs) the individuals have been selected to be good at the task they are performing, i.e. not irrational in that domain.
In some contexts it makes sense to talk about errors in opposite directions canceling out but in others it does not as errors only accumulate. Suppose one person overestimates how much they’ll enjoy having an iPad and buys one when they’d be better off without one, and another person underestimates how much they’ll enjoy having an iPad and doesn’t buy one when they’d be better off with one. Looking at the total number of iPads sold, these errors cancel out. But looking at total human welfare, the errors just add up—two people are each less happy than they could be, which is doubly bad. Similarly, if one person gets too much medical care and another gets too little, then they both lose, one from being overtreated and the other from being undertreated.
If you look at the market as a means of aggregating information (as in prediction markets) then errors can cancel out, but when you evaluate the market as a means of distributing products to people then errors just accumulate.
Friedman continues, but I shortened the quote to make it punchier. Essentially he says that, (1) given a large number of individuals irrationality will average out in the aggregate,
This is the part that sounds (and is) wrong. It would perhaps be correct if it was “given a large number of individuals selected from mind space via a carefully crafted distribution of deviations about some mind the irrationality will average out in the aggregate”. The irrationality of a large number of human individuals will not average out.
This seems to be an argument about definitions. To me, Friedman’s “average out” means a measurable change in a consistent direction, e.g. significant numbers of random individuals investing in gold. So, given some agents acting in random directions mixed with other agents acting in the same (rational) direction, you can safely ignore the random ones. (He argued.) I don’t think he meant to imply that in the aggregate people are rational. But even in the simplified problem-space in which it appears to make sense, Friedman’s basic conclusion, that markets are rational (or ‘efficient’), has been largely abandoned since the mid 1980s. Reality is more complex.
Upvoted because it provoked interesting thoughts, even though I disagree with it.
I can actually say in advance which irrational things I am likely to do on a given day. (For example, be up at 1 AM posting on Less Wrong instead of sleeping). If I know enough about a person to know their goals and approximate level of education as relates to those goals, I usually also know enough to have a sense of what types of irrational things they tend to do.
Even when errors are only random noise, modeling people as rational is different from modeling people as rational on average with random errors. If people are rational, that implies that someone with a dangerous job has properly taken the risks into account when choosing the job. But if people are rational on average with random errors, then the person who ends up with a dangerous job is probably someone who underestimated/underweighted the risks (which is a case of the winner’s curse).
It’s standard econometric practice to assume (at the very least) an error term independent of the predictor variables. That error term can be a function of any number of unobserved factors. If unbiased human error were a major component in the variance of our actions, it would be picked up in this error term.
David Friedman, Price Theory, An Intermediate Text
This sounds wrong. Biases have predictable direction, that’s why they’re called biases and not variance (ahem).
Particularly, what we know about cognitive biases precisely is “theory of irrational behavior” — we just don’t have a, complete, theory of irrational behavior.
Friedman continues, but I shortened the quote to make it punchier. Essentially he says that, (1) given a large number of individuals irrationality will average out in the aggregate, (2) In most cases that an economist would be interested in (eg. investors, CEOs) the individuals have been selected to be good at the task they are performing, i.e. not irrational in that domain.
In some contexts it makes sense to talk about errors in opposite directions canceling out but in others it does not as errors only accumulate. Suppose one person overestimates how much they’ll enjoy having an iPad and buys one when they’d be better off without one, and another person underestimates how much they’ll enjoy having an iPad and doesn’t buy one when they’d be better off with one. Looking at the total number of iPads sold, these errors cancel out. But looking at total human welfare, the errors just add up—two people are each less happy than they could be, which is doubly bad. Similarly, if one person gets too much medical care and another gets too little, then they both lose, one from being overtreated and the other from being undertreated.
If you look at the market as a means of aggregating information (as in prediction markets) then errors can cancel out, but when you evaluate the market as a means of distributing products to people then errors just accumulate.
This is the part that sounds (and is) wrong. It would perhaps be correct if it was “given a large number of individuals selected from mind space via a carefully crafted distribution of deviations about some mind the irrationality will average out in the aggregate”. The irrationality of a large number of human individuals will not average out.
This seems to be an argument about definitions. To me, Friedman’s “average out” means a measurable change in a consistent direction, e.g. significant numbers of random individuals investing in gold. So, given some agents acting in random directions mixed with other agents acting in the same (rational) direction, you can safely ignore the random ones. (He argued.) I don’t think he meant to imply that in the aggregate people are rational. But even in the simplified problem-space in which it appears to make sense, Friedman’s basic conclusion, that markets are rational (or ‘efficient’), has been largely abandoned since the mid 1980s. Reality is more complex.
Both claims are implausible. Is there some kind of substantiation?
Upvoted because it provoked interesting thoughts, even though I disagree with it.
I can actually say in advance which irrational things I am likely to do on a given day. (For example, be up at 1 AM posting on Less Wrong instead of sleeping). If I know enough about a person to know their goals and approximate level of education as relates to those goals, I usually also know enough to have a sense of what types of irrational things they tend to do.
Even when errors are only random noise, modeling people as rational is different from modeling people as rational on average with random errors. If people are rational, that implies that someone with a dangerous job has properly taken the risks into account when choosing the job. But if people are rational on average with random errors, then the person who ends up with a dangerous job is probably someone who underestimated/underweighted the risks (which is a case of the winner’s curse).
It’s standard econometric practice to assume (at the very least) an error term independent of the predictor variables. That error term can be a function of any number of unobserved factors. If unbiased human error were a major component in the variance of our actions, it would be picked up in this error term.
Are you thinking of something more specific?