Some highlights / takeaways / thoughts / comments from #RWRI 14, Day 4 (probably my favorite day thus far):
1. If the payout is convex, you don't care about the average; you care about the second order effects.
1. If the payout is convex, you don't care about the average; you care about the second order effects.
2. The more optimized a system, the more fragile it is. The tradeoff for more optimization is less redundancy. So when hyper-optimized organizations meet a large stressor, they don't have enough safety margin to survive.
Example: a hyperoptimized retail storefront plows ALL available cash into inventory. A less optimized retail store keeps lots of cash in reserve. Which one has a better chance of surviving a massive loss of sales due to COVID (or other stressor)?
One store has a bunch of inventory that can't easily be moved. The other has a bunch of cash that they can use to make masks / hand sanitizer, invest in moving to online retail, or just sit on.
The hyperoptimized may show exceptional returns during the good times, but when there's a major shock to the system, they go bust, negating all of their previous exceptional returns.
3. If you seek antifragility, financial markets are a difficult place to find it, as this industry is much more aware of the concepts of nonlinearity / convexity. Find an industry where these concepts are foreign / people don't get it.
4. "No one without optionality has ever survived in history."
5. More on optimization. If you're built to ignore stressors (often done via optimization), you'll miss out on the information that these stressors provide.
6. The more connected the network is, the more winner-take-all effects. This applies to wealth distribution, Twitter followers, YouTube subscriptions, etc.
The income of an opera singer in the 1800s was capped by (a) the number of seats in the theater, and (b) how much time they spent performing. Now, you can spend the same amount of time making YouTube content, which can be viewed by billions while you take a nap in Tahiti.
7. You are convex when you cap your losses and have exposure to upside. This is why trial and error is so effective: you can gain information very cheaply (capped loss), and then do whatever you want with it (upside exposure / optionality).
Similarly, it's actually good to make small errors, because they may provide information that benefits you (i.e. you learn from your mistakes, or find a surprising result in spite of your error)
8. Philosopher's Stone: having superior convexity is *vastly better* than having superior knowledge. The convex is exponential; knowledge is not. In other words, you can have returns 1,000,000x better than average, but you can't be 1,000,000x smarter than average.
9. "Anything fragile eventually breaks."
If you're concave, over a long enough time span, you will go bust. If you're convex, over a long enough time span, you will hit a fat-tailed payoff.
If you're concave, over a long enough time span, you will go bust. If you're convex, over a long enough time span, you will hit a fat-tailed payoff.
10. Under fat tails / nonlinearities (i.e. most of life), correlation has no meaning.
Pre-COVID, there were likely hundreds of metrics that correlated closely with the unemployment rate. Almost all of these correlations were broken over the past 6 months.
Pre-COVID, there were likely hundreds of metrics that correlated closely with the unemployment rate. Almost all of these correlations were broken over the past 6 months.
You can't draw a correlation line when *one data point* will determine the properties of the entire distribution (and you almost certainly don't have that data point in your sample).
In other words, the correlation from your sample will be vastly different from the correlation from the population, since one (or a few) data points impact the distribution so heavily.
11. Unless you have a strong logical argument for why something is NOT fat-tailed, you cannot rule out fat-tailedness. For example, we know calories eaten in a day is not fat-tailed, because you can't eat 500,000 calories in a day.
A fat-tailed distribution CAN masquerade as a normal distribution.
A normal distribution CANNOT masquerade as a fat-tailed distribution.
Thus, unless there's a strong logical reason for why a distribution is not fat-tailed, the default assumption should be it is fat-tailed.
A normal distribution CANNOT masquerade as a fat-tailed distribution.
Thus, unless there's a strong logical reason for why a distribution is not fat-tailed, the default assumption should be it is fat-tailed.
Example: A firm has incredible returns for a 5 years, then goes bust and blames it on a '10 sigma event' (i.e. 1 in 10^23 probability).
Which is more likely: the CEO is a genius and hit the unlucky 1 in 10^23 event, or the CEO is an idiot because they didn't understand the type of distribution they were playing with?
Hint: the CEO isn't a genius.
