A day with SFI learning "Optimality vs Fragility"

Recently I had the privilege of attending Santa Fe Institute's latest joint conference with Morgan Stanley. This time, the topic was "Optimality vs Fragility: Are Optimality and Efficiency the Enemies of Robustness and Resilience?" The topic was both intriguing and timely, and the speakers were interesting, informative and a little bit more controversial than in years past. This made for an outstanding day. The audience in the room included some big names in finance and science alike, setting the stage for fascinating Q&As and stimulating conversations during the breaks.

This year, rather than writing one big post covering all of the lectures, I will break each down into its own entry. Here are the subsequent posts in order (and their respective links). Let this serve as your guide in navigating through the day:

Cris Moore--Optimization from Mt. Fuji to the Rockies

Nassim Taleb--Defining and Mapping Fragility

John Doyle--Universal Laws and Architectures for Robust Efficiency in Nets, Grids, Bugs, Hearts and Minds

Rob Park--Logic and Intent: Shaping Today's Financial Markets

Juan Enriquez--Are Humans Optimal?

Dan Geer--Optimality and Fragility of the Internet

I like to think about are how the lectures relate to what I do in markets and where there is overlap and dissention between the speakers. Further, I like to analyze how some of these lectures fit (or don't) with my preexisting views. I would love to hear what others think. Here are a few of my observations to get you all started:

  • Cris Moore's point that "best" is not necessarily optimal, and a confluence of models (what he calls data clusters) can yield better outcomes is extremely important in financial markets.
  • Nassim Taleb's suggestion that stress tests should focus on accelerating pain, rather than spot analysis is a powerful one that all risk managers should think about.
  • John Doyle's observation about the tradeoffs between robustness and efficiency is directly applicable to portfolio construction.
  • Rob Park's explanation of how algorithms are designed to express human intent, and the areas in which that can go has me rethinking my understanding of the risks from HFT.
  • Juan Enriquez opened everyone's eyes to how big the advances are in life science and the consequences this holds for the "secular stagnation" debate.
  • Dan Geer's explanation for why we have a choice between two of "security, convenience and freedom" online is both an enlightening and frightening call to action.

Again I will caution that these are my notes from the sessions. There is no guarantee of accuracy or completeness. I specifically focused on points that were intriguing to me, and purposely left out areas where the subject matter and terminology were too far removed from my competency. 

Nassim Taleb at SFI

Nassim Taleb

Defining and Mapping Fragility


  • Black swans are not about fat tail events. They are about how we do not know the probabilities in the tail.
  • The absence of evidence vs evidence of absence is very severe
    • Too much is based on non-evidentiary methods
  • Financial instruments (options) are more fat-tailed than the function suggests
    • P(x) is non-linear
    • Thus the dynamics of exposure are different than the dynamics of the security
    • To that end law of large numbers doesn’t apply in options
  • “Anyone who uses the word variance does not trade options”
    • The measure of a fat tail is a distribution’s kurtosis
  • There was a great chart of 50 years of data across markets
    • In the S&P in particular, 80% of the kurtosis can be represented by 1 single day (1987 crash)
    • This would not converge in your data studying a broad look at the S&P
    • One  can only talk about variance if the error coefficient of the variance is under control
    • In Silver, 98% of its 50-year variance comes from 1 observation
  • EVT-extreme value theory is very problematic because we don’t know what the tail alpha is.
    • In VAR, a small change can add many 0000s
    • There is no confidence at all in the tails of these models
    • The concentration of tail events without predecessors means that such events do not occur in the data. Tails that don’t occur are problematic.
  • A short option position pays until a random shock. Asymmetric downside to defined, modest upside. This bet does not like variability (dispersion), volatility.
  • Look at the level of k (believe kurtosis??) and see sensitivity to the scale of the distribution. This is fragility.
    • Volatility = the scale of the distribution
    • The payoff in the tail increases as a result of sigma
  • If you define fragility, you can measure it even without understanding the probabilities in the tail
    • Nonlinearity of the payoff in the tail means that the rate of harm increase disproportionately to an instance of harm
    • What is nonlinear has a negative response to volatility
  • Fragility hates 2nd order effects. For example: if you like 72 degree room temperature, 2 days at 70 degrees is better than 1 at 0 and the next at 140.
  • Lots of nature demonstrates “S” curves
    • In the convex face of the s-curve, we want dispersion. In the concave face we do not (stability)
  • How to measure risk in portfolios: takes issue with IMFs emphasis on stress tests looking at a “worst” past instance, which is a stationary point in time.
    • Dexia went out of business shortly after “passing” such a stress test
    • Solution: do 3 stress tests and figure out the acceleration of harm past a certain point, as conditions get worse. 
      • We should care about increasing levels of risk, not degree
      • Risk increases asymmetrically, so if the rate of acceleration is extreme, this is stress.
  • Praised Marty Liebowitz for “figuring out convexity in bonds”
  • “Convex losses, concave gains --->thin tials ---> robust”
  • Antifragile=convex, benefits from variability
  • Can take the past to see the degree of fragility. You get more information and more measurable data from something that went down and then back up in the past, than something that went down and stayed at 0. 
  • Adding information is concave (N). Convex is when we add dimensions (D), spurious correlations increase.
    • There is a large D, small N problem in epidemiology.
    • NSA is one of the few areas that uses data well, but this is so because they are not interesting in many things, only the few that have value to what they’re trying to do.
  • PCA analysis, variations are regime dependent. 
  • We can lower nonlinearity of a price (buying options) 
    • Hard to turn fragile into anti-fragile, but can make it robust (tea cup can put lead in it).
    • Robust requires the absence of an absorption barrier – no O, no I in transition probabilities. Don’t stay or die in a specific state.
  • “Small is beautiful”
  • Q that “VAR is the best we have”:
    • A pilot says on a flight to Moscow: “We don’t have a map of Moscow, but we do have one of Paris.” You get off that plane. We don’t take random maps for that reason, and same logic applies to VAR.
    • Using VAR under this logic is troubling because it encourages people to take more risk than they really think they are taking. They anchor to the probabilities of VAR, not reality.
  • Liquidation costs are concave. There are diseconomies of scale from massive size.