Publishing is a hit-driven industry and we rely on positive rare-events. We attempt to expose our authors and ourselves to as much positive uncertainty as possible.
The Black Swan is a complicated concept, here I’ve tried to illustrate some of these in a more intuitive manner.
1. Negative Skew
Negatively skewed businesses are short a put. They make a little but of money very often and lose all of it (or more) very rarely. Banks have this property. Many negatively skewed businesses can lose more than they have earned in their lifetime. Large scale data storage providers also face similar risks. Catastrophic data loss will not only invalidate their business, the loss can result in millions of dollars of intellectual property being wiped-out.
Competition and consolidation prevent systems from adequately including the price of catastrophe insurance in their pricing.
Publishing is a positively skewed business. So is Venture Capital, most forms of Mining, and pharmaceutical research. There are other examples that I’ll go into in the future.
2. Risk Dimensions
Risk is often shoveled into other dimensions that are difficult to pin down. Only after a rare-event takes place is the new dimension given a name (a la “liquidity,” “correlation risk”).
3. The Problem of Induction
Every day in every way I’m getting better and better…not.
The biggest problem here is that from the turkey’s perspective, the same data from history confirms both theories: the eventual death before thanksgiving, and the notion that humans will never harm a turkey.
The day before thanksgiving, the turkey has the most number of data points to support the theory that humans will keep feeding it forever!
4. Quality of Knowledge
There is knowledge, and there is confidence on that knowledge. Error rates and confidence levels matter and need to be included in all measurements, forecasts, and estimates.
Rare-events, by their very nature, are unpricable. How confident can one be of statistical knowledge if the sample size is small? The rarer the even the fewer episodes we have to work with. Banks trying to estimate the probability of default are trying to come up with a number based on a very small sample set of historical default rates.
Would you cross a river if I told you it was 4 feet deep on average?
5. Mediocristan and Extremistan
Certain properties fall in Mediocristan - the world where the gaussian distribution rules. Human height is an example. No one person can distort the average height of a large number of people.
Personal wealth on the other hand can take on almost arbitrarily large values. Bill Gates’ net worth can distort the average by a large amount, even with a very large number of people (I can’t stress this enough, this has an impact on tradition notions of diversification). A single data point can vastly change the average or sum. That is extremistan.
6. Model Errors
Picking the wrong model can have dangerous consequences. There are domains where spending energy on being prepared is better than spending scarce resources on trying to arrive at a model through pattern recognition or other techniques. See my earlier post on “Jaywalking Models” for a more detailed look at this concept.
7. Reverse Process
Estimating the shape of the puddle of water given the properties of the melting ice cube is simple (relatively). But arriving at the shape of the ice-cube given the properties of the puddle of melt-water is not even possible.
History is opaque. What we see today is only the puddle
8. The Map is Not the Territory
Models and theorizing (aka “simplifications”) always result in loss of information. So does categorizing. These are Black Swan generators. Maps are useful, but must never be confused with the territory.
This post will be updated as I come up with different perspectives.
Post a Comment