Fractal Press

I took a few business courses at Stanford. They were pretty entertaining and my favorite was a course on Supply-Chain Management. It was a series of case-studies that all started or ended with: “Rob looked out of his office window overlooking the hills in Palo Alto and wondered how he was going to …”

Invariably, the case-studies had consultants who would come in and save the day. More often than not they would either move the supply-chain from an existing centralized system to a more distributed process or vice versa. In both cases there would be very rational sounding reasons for doing so, and indeed the case study would cite much increased growth and productivity at the host company after the consultants has done their thing.

I never quite understood how one would make the decision before-hand on whether to go with a centralized or distributed process. In any case, it was always presented as a binary decision - it was never framed as a tradeoff.

Because that’s what it really is. Anytime somebody moves the slider on the efficiency scale, the system will become more or less robust. The higher the efficiency, the less robust the system. A highly efficient centralized system is much more vulnerable to single-point-of-failure situations. A more distributed approach, while being more resistant to single-point failures, is less efficient. An efficient system will also fail very efficiently.

Unfortunately, competition pushes systems to move towards higher and higher efficiency, all the while ignoring the massive drop in robustness for each player. Robustness is a difficult quality to measure compared to efficiency, and is thus rarely included in cost/benefit analyses. Efficiency gains can be quoted in percentage terms, but robustness measures only rely on “messy” scenario analysis that are difficult to enumerate and ask people to imagine the unknown.

While efficiency has been reified, robustness remains an elusive measure that few take into account until a failure actually occurs.

I have written about the role of competition and salience before. I argued that the presence of stiff competition can mean that the option is at least priced correctly, if not overpriced. It’s a matter of looking at how skewed the demand-supply distribution is. If demand is much higher than supply (stiff competition), it’s highly likely that the option is priced appropriately or overpriced. I’m talking about real-world options here.

On the other hand, if supply exceeds demand, the situation is ripe for underpriced or “cheap” options. This means that you would know beforehand that there’s a good chance that the ROI is positive. In the case with overpriced options, there is no clear indication beforehand that it’s rational to enter the transaction at all (but salient options with high payoffs will tend to draw hordes of competitors).

I was going through Ben Casnocha’s archives when I read this:

The main thing I’ve learned in my seven years studying and doing presentations is that the standard for presentations / public speaking in the professional world is low, and as such it’s easy to be seen as great. When most people suck at something, all you have to do is suck less. And when I discovered I had a natural knack for communication, I identified speaking / communicating / presenting as one of my natural strengths that, if built upon, could become an unstoppable strength thanks not only to my own capabilities but because of how I would be perceived relative to the masses.

Since public speaking is supposed to be one of the top feared things amongst the population, there is little competition to speak of (relative to less feared things). And since this fear keeps the average quality low, a small edge can be positively leveraged easily.

In competitive environments, a small loss of edge will mean leverage in the opposite direction - being tossed out of the pool. Here, a small edge means an increase in positive leverage.

For me this is exciting because I’m always on the lookout for investing in “cheap” options. I now have a criterion that I can use to ferret out such opportunities. Keeping an eye out for things that people fear.

Of course, it’s very likely that I will initially fear the same things, but if the worst-case scenario is clipped, I will invest in training to overcome the fear. Because I know that I will only have to improve by a little to gain a positively leveraged edge that will cost very little to acquire.

I varied the tail-exponent of a Student’s-T distribution, taking it from a near-gaussian distribution to a near-cauchy, fat-tailed distribution and noted the relationship between kurtosis and information entropy. Here’s the plot:

The implications are interesting. More on that later.

We cannot learn from history in complex systems, except at the meta level (we can learn that we cannot learn, for example). This follows once one accepts that causality is impossible to pin down given an observation of a phenomenon.

Ericsson et al show that learning, and becoming an expert, requires:

1. A tight relationship between action and consequence - i.e causality needs to be discoverable
2. Quick and clear feedback on actions
3. Repeatability

Let us consider whether a brain surgeon passes these criteria. Causality is indeed easily discoverable, nicking an artery in the theatre will make it clear that a mistake has been made - there is a tight relationship between action and consequence. The feedback is often immediate, at least in terms of the surgical procedure itself. Since there is relatively minor variability between human brains, and since the operation theatre is a controlled environment, the entire process of “brain surgery” is highly repeatable.

This makes brain surgery a classic domain for learning from the past and allows the creation of experts.

Now let us look at managing the environment, (enough people are picking on economists now, so I’ll skip them), and see if it is possible to learn from the past and become an expert in this domain. Are the three criteria satisfied?

First, causality is very difficult to identify. Small changes in any number of parameters, ranging from climate to prey availability can have large and sudden impact on predator populations. Given a phenomenon, it hard to say what caused it. If a human action is involved, outcomes are unpredictable and can spiral out of control when coupled with external variations (see the many instances of predator introduction for hilarious examples) .

Second, the results can take years to manifest themselves. The chain of events precipitated by an action can cascade into a series of relatively small changes until it shows up as a big impact a long way down the daisy-chain. Feedback on actions is far from immediate. Often, the people initiating the action are no longer even in charge anymore. Attempts to control Wolf populations in Yellow Stone National Park had impacts decades down the line.

Third, the environment is not controlled. Conditions that held once need not hold after a period of time has passed. Rivers dry up, predator-prey populations change all the time, and climate changes make any assumptions of stationary processes unjustified. Actions that worked earlier need not work now, because the underlying mechanisms may have changed.

All three conditions for learning and expertise are violated. Under these conditions, it is not possible to learn from the past and become an expert.

An “expert” at managing the environment can only be someone who takes a very fallible approach and shows humility and respect in the face of the unknown.

Other domains where experts cannot be present by the nature of the system itself: The movie industry, publishing, economics, stock-markets, climate, politics, wars, … the list goes on.