Just another example of amusing agent rationalizations

[The Tim]

[Alan]

[redmondjp]

Wait - now hold the phone here.

I live 0.8 miles from the Redmond West Campus of MS. I can walk there in less than 15 minutes. THAT is close to Microsoft.

Now how exactly is Gig Harbor close to Microsoft? Do they have an office over there that I don't know about? Does a four-hour daily commute sound like it's close?

Somebody needs to call the person who wrote that ad and set them straight. There are lots of places in India that are close to Microsoft too!

[rose-colored-coolaid]

[kpom]

Come now, I'm sure you can drive from Gig Harbor to Microsoft in 15 minutes... (does anyone at MS commute from Gig Harbor, I wonder).

I also like the bit about no floods or earthquakes...

[Alan]

I studied AI for five years in grad school. I understand the machine learning viewpoint. One of the most easily made mistakes that results in overfitting is to use data that is not independent. If you read the newspaper and it says that housing always goes up and then you talk to your neighbor and he says that housing always goes up, you might think that is two different pieces of information and increase your confidence that housing always goes up. But your neighbor got his information from the newpaper so you are actually giving double credit to the newspaper. What is worse is that next time you talk to your neighbor you tell him that you agree housing prices always go up. That increases his confidence -- but now he is counting the paper three times (twice from his own reading and once from yours).

AI (and natural intelligence for that matter) is largely about choosing the right biases that are useful for your domain. Taleb claims that the bias humans naturally have for narratives is strong and that you need to be careful that it is not exploited by others. It doesn't mean that narratives aren't useful, just that they aren't as useful as you think they are.

Of course, I used a narrative to make my point. If you believe that narrative you should might make the inductive step that the seattle bubble is an echo chamber where we all continually reinforce our own beliefs and maybe there isn't a bubble in real estate after all. But since you have a natual bias to give too much weight to narratives, you should probably pay less attention to my story than you initially did.

[kpom]

My IPhone informs me that it is currently a 1 hour and 24 minute drive from MS to Gig Harbor with traffic (52 miles, BTW). Of course, it's still early in the evening commute...

[rose-colored-coolaid]

Good response Alan. And I agree that narrative is a powerful function in human decision making.

Reading that gun violence is up 500% might make you shake your head. But having a friend tell you how they were shot at will probably cause you to change your behavior. So I agree 100% as far as that's concerned.

I haven't read the Black Swan book, my worry was that people would read a synopsis of it, hear that big events are often not predicted, and then come to the conclusion that we need to always be on the lookout for these black swans. In reality, there's a difference between rational caution and paranoia, and a lot of people I know tend to prefer paranoia.

[Alan]

I am in the process of reading it. It is a fun and easy read. Taleb uses narratives to make his points and then tells you that it is easy to make mistakes from narratives.

I like to use "the water tank" problem as an example of what he is talking about.

Consider a water tank with a constant flow of water entering it and a drain where the flow of water out is proportional to the level of the tank. You have a bobber that measures the water level, but the spashing and waves caused by the entering water gives you a noisy signal. We model the noise with a Gaussian curve.

If you do a qualitative analysis of this system, you find that one of three things can happen:
- The water reaches a constant level below the top of the tank as the flow in becomes equal to the flow out.
- The water overflows the tank.
- The water reaches a constant level right at the top of the tank (which is kind of a corner case so I'm going to ignore it for the rest of this discussion).

One can measure the water level of the tank and make predictions about the next time step. If the tank does not have a "funny" shape you can even predict fairly early what the steady state of the tank will be. However, if you do not know the capacity of the tank, you have no way of prediction through observation of the water level, when the tank will overflow. You can build a highly predictive exponentially decaying curve with Gaussian noise. You get these fantastic results right up until the point when the tank overflows. Even if you know that the tank might overflow in theory, reading the water level tells you nothing about the actual capacity. And of course, by the time it happens, it is too late. The tank overflowing is a Black Swan.

But maybe you know the capacity. There are still other things you may not know. Maybe the tank isn't strong enough to hold all the water that could fit in it. Maybe its supports aren't strong enough. Maybe it is near a busy intersection and a car crashes into it one day. The guy making predictions from the model says it is impossible. According to the error estimates it happens once in a million years. And yet it happened twice last month.

I think black swans boil down to bad models, but building good models is nearly impossible. How do you account for things that you don't know that you don't know?

[rose-colored-coolaid]

[Alan]

I think one of the lessons is that you don't want to leverage billions of dollars on a model that you know is incomplete in some way that you do not understand.

[kpom]

"If that's the lesson, I find it sad that we even need to teach people this. As a fellow computer science guy, that models are simplifications of real life is so fundamental as to be axiomatic."

A standard problem with economic models is that they tend to use nice, smooth, tractable, twice-differentiable functions - so they aren't very good at predicting when something is going to fall off a cliff.

[rose-colored-coolaid]

[Alan]

The point of my example, and of Taleb's book, is that you cannot see the cliff by using past data. By the time you see the cliff you are already falling.

[rose-colored-coolaid]