Haute Lumière

Commerce · II.02 · MMXXVI · daylight

La Bourse  /  Volume II  /  Nº II.02  /  Workbook — the executive

WORKBOOK — THE CORPORATE EXECUTIVE

Chapter II.02 · Markets as Living Systems

For the person with a P&L, a board, a risk committee and a planning cycle. This chapter is not a philosophy of markets. It is an argument that several of the numbers in your board pack refer to quantities that do not exist, and a ninety-day method for finding out which.

---

THE PREMISE, STATED COMMERCIALLY

Your organisation almost certainly sizes risk on a standard deviation.

If the series you are sizing has a tail exponent below 2 — which is common in claim sizes, outage durations, project overruns, customer concentration, and almost everything in financial markets — then the variance of that series is infinite and the standard deviation you are quoting is an artefact of your sample size. Take a longer sample and it goes up. Take a longer one again and it goes up again. It is not converging to anything.

That is not a philosophical objection. It is a measurable property, it takes an analyst about three days to test, and it has a direct consequence for how much liquidity you hold and what your scenario suite contains.

Three commercial claims follow from this chapter, and each of them is a number:

1. A buffer sized on Gaussian 99 percent VaR holds 42 percent less than the average loss in the tail it was built for. The multiple is k = 1.74, and you should compute it on your own data rather than take ours.
2. Averages in heavy-tailed populations are not planning quantities. US firm sizes have Pareto exponent 1.059. If your customer or supplier base looks like that, "the typical account" is a phrase with nothing behind it.
3. Diversity and connection are not free resilience. May's criterion, σ·√(S·C) > 1, says a more diverse, more connected network is less stable in the random case. Supply-chain diversification that raises connectance without lowering interaction strength can increase your exposure.

Nothing here requires your organisation to adopt a worldview. It requires four reports, one analyst, and a treasurer willing to run a test that can tell them not to buy the instrument.

---

PART ONE — DISCOVERY

Days 1–30: find where your own data already says this

Exercise 1.1 — The five series (one week, one analyst)

Pull five series your organisation already keeps, with at least 500 observations each. Good candidates, in order of how often they turn out to be heavy-tailed:

For each, compute six numbers: mean, standard deviation, 99th percentile, expected shortfall beyond the 99th, and two independent tail-exponent estimates — Hill and rank-frequency regression.

Report the disagreement between the two estimates. A single estimate quoted alone is a claim; two estimates with a stated gap is a measurement.

Exercise 1.2 — The concentration sweep (three days)

Three concentrations, each a single number, each already in a system you own:

  Revenue concentration   =  share of revenue from the top 5 customers
  Supply concentration    =  share of critical inputs from the top 3 suppliers
  Capability concentration=  share of critical processes with one qualified owner

Then estimate a, the relative ascendency: the share of total throughput running through your single most optimised path. It is a rough number and it does not need to be better than rough. Compare it to 1/e = 0.368 — not as a target, but as a shape. Above about 0.6 and you are on the steep side of the curve, where every further efficiency gain costs more robustness than it buys.

Exercise 1.3 — The appreciative interview, with your own operators (a week)

Ten conversations, same question, no variation:

"Tell me about a time we got a price, a forecast or an allocation right when the model would have got it wrong. What were you reading?"

Take notes on what they were reading, not the outcome. What you are assembling is the list of local information your models do not carry — which is precisely Hayek's argument, and precisely what a heterogeneous-agent model is built to represent. It is also, immediately and independently of any of this, the highest-value list in your organisation.

---

PART TWO — THE ARITHMETIC

Days 31–50: the four numbers that change a board pack

Exercise 2.1 — Compute your own k (one week)

   k  =  ES(fitted tail, 99 %)  /  VaR(Gaussian, 99 %)

Take your fitted exponent. Compute the expected shortfall at 99 percent under that tail, and the 99 percent VaR under a normal with the same sample standard deviation. Divide.

The chapter's worked case gives k = 1.74 at a cubic tail. Your number will differ and the difference is the point.

The rule that makes this credible: if your fitted exponent comes back above about 4, k will be close to 1 and you should not make the change. Publish that result as loudly as you would publish the other one. An instrument whose own test can veto it is the only kind a risk committee should accept from you.

Exercise 2.2 — Restate one board number as a distribution (three days)

Choose the single number in your board pack that most decisions hang on. Beside it, put four things: the median, the 10th percentile, the 90th, and the fraction of your historical periods that fell outside the range the planning assumption allowed for.

Do not editorialise. Put the four numbers next to the one number and let the committee read them.

Exercise 2.3 — May's criterion on your own network (one week)

Estimate three quantities for your supply or counterparty network:

• S — the number of meaningful nodes
• C — connectance: the fraction of possible node-pairs that actually have a dependency
• σ — typical interaction strength: how much a node's failure moves its neighbours, as a fraction

Compute σ·√(S·C).

Read the second and third rows together, because they are the commercial finding of this whole workbook. At fifty nodes, halving connectance takes the network from unstable to stable while leaving the node count untouched. The lever is not how many suppliers you have. It is how coupled they are — shared sub-tier, shared geography, shared logistics corridor, shared financing. Most diversification programmes add nodes and leave coupling alone, and the arithmetic says that is the wrong half.

Exercise 2.4 — The counter-case, written by you (half a day)

Write 400 words naming the questions where your existing tools are better and should not be replaced:

• single-market pricing over a short horizon with stable demand
• auction and tender design
• matching and allocation problems with a provable stability property
• tax and regulatory incidence, where the counterfactual must be legible
• index and cost-of-living calculations

Circulate this before you circulate anything else. An executive who has already named the boundary of their own proposal is received entirely differently from one who has to be shown it.

---

PART THREE — DESIGN

Days 51–75: the instrument

Exercise 3.1 — Size the facility (one week)

The instrument is a committed, undrawn revolving credit facility, sized on expected shortfall under your fitted tail, with a covenant holiday keyed to a published external dispersion measure.

Why the trigger must be external and published. A trigger keyed to an internal measure will be argued about in the exact month it fires, by people under stress, with the facility's availability as the stake. External, published and non-discretionary removes that conversation from the worst week of the decade.

Exercise 3.2 — The decision inequality (two days)

        k × (annual commitment fee on the increment)
   ------------------------------------------------------   <  1
    P(tail state) × (distressed financing spread
     + forced asset sales + lost options)

Most organisations have never computed the denominator. Compute it from your own history: the last time you were liquidity-constrained, what did the financing cost above normal, what did you sell, and what did you not do? That number is usually large and it is usually never written down.

Exercise 3.3 — The three floors (one week, then annually)

Take the three concentrations from Exercise 1.2 and give each a floor and a ceiling, written into the standing pack and reviewed at the same meeting as margin.

The floor is the whole mechanism. Margin will argue for the ceiling every quarter without help from anybody. Nothing in the standing agenda argues for the floor unless it is written down, which is why resilience programmes lose to efficiency programmes in a fair fight — the fight was never fair, because only one side had a standing item.

---

PART FOUR — DESTINY AND DELIGHT

Days 76–90: make it hold

Exercise 4.1 — Get one exponent into the standing pack

Anything reviewed monthly persists; anything reviewed by exception does not. One tail exponent, one line, on the standing risk page, with its date and its two estimates. Once it is there, removing it requires somebody to explain why.

Exercise 4.2 — Name the second owner

One person who can run the estimation and the model alone. Recruit them before you need them, and recruit them by giving them the credit for the first result. One person is a hobby; two is a practice.

Exercise 4.3 — Publish the back-test annually

If you build a model, publish its misses once a year, internally, with the misses first. A model that is never wrong in public is being used wrongly — its value is that it produces a distribution of futures, and the moment somebody quotes its median as a forecast it has become a worse version of the thing it replaced.

Exercise 4.4 — Delight, deliberately

The pleasure available here is specific: it is the meeting where the extreme case has a number beside it and nobody is guessing. Wondering whether this is the quarter something breaks gets replaced by a figure, and the figure is in the facility, and the facility is committed. Buy the calm; it was always cheaper than the alternative.

---

THE FAILURE MODES, NAMED

So you can see them coming

The model becomes an oracle. Guard: a published back-test on withheld data, annually, misses first.

The metaphor outruns the mathematics. "The economy is a living system" can be made to justify deregulation and intervention equally, and neither inference is licensed. Guard: no analogy is used in a board paper unless you can name the specific result being carried across and the assumption it needs.

The sophistication becomes the product. A model only its author can run will die with its author. Guard: a one-page description a person outside the team can read, and a second owner who has run it alone.

The exponent is fitted once and never again. Tails move. Guard: the eighteen-month hard review is in the facility documentation, not in someone's diary.

---

THE NINETY DAYS ON ONE PAGE

---

BOARD PAPER TEMPLATE

One page. Six blocks. No adjectives.

1. The finding. "Our [series] has a fitted tail exponent of α = [x], by two independent methods giving [a] and [b]. Below 2, the variance of this series does not exist."
2. What that means for a number we already use. One existing planning figure, restated with its distribution.
3. The buffer multiple. "k = [x]. Our current buffer is [y] percent short of the average loss in the tail it was sized for."
4. The instrument. Size, form, trigger, term, counterparty, treatment. Six lines.
5. The decision inequality. One fraction, one comparison, one verdict.
6. What would make us not do this. The exponent above 4, the commitment fee above [x] basis points, or the denominator in block five smaller than [y].

Block six is the one that gets the paper approved. A proposal that names its own veto conditions is read as analysis; one that does not is read as advocacy, and boards are professionally immune to advocacy.

---

APPRECIATIVE QUESTIONS FOR YOUR LEADERSHIP TEAM

1. When has one of our operators been right in a way our models were not? What were they reading, and could we carry it into a model?
2. Where in this business do we already plan for the distribution rather than the average — and what does that let us do that the rest of the business cannot?
3. Which of our supplier relationships is genuinely decoupled from the others, and what made it that way? How would we get three more like it?
4. If every number in our board pack arrived with its tail attached, which decision would change first — and would we be glad?
5. What floor on diversity would we be willing to sign, so that next quarter's efficiency pressure has something legitimate to push against?
6. Who is the second person who could run this analysis alone, and what would it take to make that true by the next cycle?