Haute Lumière

Commerce · III.10 · MMXXVI · daylight

La Bourse  /  Volume III  /  Nº III.10  /  Ten concept briefs

TEN CONCEPT BRIEFS · Chapter III.10 — Reserves and Resilience

One page each. A reader who reads only these ten pages has the chapter.

---

BRIEF 1 — What a Reserve Is For

The idea. A reserve is not spare capacity. It is a decision, taken in advance, not to use something — and the whole cost of it is that the thing is alive and earning less than everything around it.

That definition does three useful things at once. It says a reserve has a price (the spread between what it earns and what the rest of the balance sheet earns). It says a reserve has a purpose (an event, with a duration). And it says a reserve has a frequency test (the event has to arrive often enough for the price to be worth paying).

Worked example. A firm holds $50m of cash against disruption. It returns 12 percent on its operating capital and 2 percent on that cash. The reserve costs $50m × 10% = $5.0m a year. Nobody put a $5.0m line in the budget called insurance, but that is what it is, and it is the third or fourth largest discretionary expense in the business.

Why it matters. Once a reserve has a price, it can be compared with alternatives that have prices — a committed facility, consigned stock, a second supplier, a different product design. Without a price it can only be compared with virtue, and virtue loses to a working capital target every single time.

You already know this because you keep a spare tyre you have never used, you have decided how much petrol counts as empty, and you have never once thought of either decision as free.

---

BRIEF 2 — Safety Stock, and Whose Variance It Is

The idea. The buffer you need is set mostly by your supplier, not by your customers.

   safety stock  =  z · σ_DL
   σ_DL = sqrt( L·σ_d²  +  d²·σ_L² )

Two variance terms are added: demand variance over the lead time, and demand squared times lead-time variance. The second term is the one that moves.

Worked example. 1,000 units a week, demand deviation 300. Semiconductor lead times ran about 13 weeks in 2019 and 27 weeks at the May 2022 peak; take lead-time deviations of 2 weeks then and 8 weeks at the peak.

The required buffer multiplies by 3.58× and demand never moved. In 2022 the lead time is 96.34 percent of the variance.

Why it matters. It tells you where to spend. A week removed from your supplier's lead-time deviation is worth more than a forecasting project, and it is purchasable — with a contract clause, a second source, or a payment.

You already know this because you leave earlier for the airport when the traffic is unpredictable, not when it is merely heavy.

---

BRIEF 3 — The Loss Function and the Price of the Last Point

The idea. Expected units short per cycle is σ · L(z), where L(z) = φ(z) − z(1−Φ(z)). Service level and safety stock rise steadily; expected shortage falls off a cliff; and the price per unit of shortage avoided rises without limit.

Worked example. On σ = 8,150 units at $12 a unit a year:

The last step buys 2.06 fewer units short for $61,499.

Why it matters. "We should never run out" is not a policy. It is a refusal to pick a number, and the number it implicitly picks is the most expensive one available.

You already know this because you do not insure a bicycle for the same excess as a house, and you would find it strange to be asked to.

---

BRIEF 4 — The Newsvendor, and Where the Insurance Stops Paying

The idea. Francis Edgeworth posed this in 1888 for a bank's cash reserve, not for newspapers. The reserve problem and the inventory problem have been the same problem since before either had a name.

   critical ratio  CR = Cu / (Cu + Co)
   Cu = cost of being one unit short      Co = cost of holding one unnecessary unit
   order the n-th unit while  Cu·P(D ≥ n)  >  Co·P(D < n)

Worked example. Grid spare transformers. A shortage costs $2.4m (a three-day outage); a held spare carries at $90,000 a year. CR = 0.963855. With Poisson demand at four failures a year:

• the 8th spare has a marginal value of +$37,323
• the 9th spare has a marginal value of −$36,805

So Q* = 8 — twice the expected number of failures. The whole reserve costs $720,000 a year, which is 30 percent of one avoided outage.

The continuous version. With a stockout penalty p, demand D and order size Q, the optimum satisfies 1 − Φ(z) = hQ/(pD). At h = $12, Q = 26,000, p = $240 and D = 52,000, that is z = 1.96, a 97.5 percent service level.

Why it matters. The ninth spare is where the insurance stops paying, and it is computable to the unit.

---

BRIEF 5 — Expected Shortfall, and Why the Tail Shape Decides

The idea. Value at Risk asks how bad is the edge of the tail. Expected Shortfall asks how bad is the tail on average. Basel's 2019 market-risk framework moved from 99 percent VaR to 97.5 percent ES for exactly that reason.

Worked example, in units of σ.

Under a normal loss, ES at 97.5% and VaR at 99% differ by 0.49 percent — which is the whole calibration. Under a t with five degrees of freedom and identical variance, expected shortfall is 16.68 percent higher.

Why it matters. A buffer sized on normal-tail assumptions is a sixth short against a fat-tailed world of the same volatility, before anything else goes wrong. Volatility is not the input that matters for a reserve. Tail shape is — and tail shape is the part of the distribution you have fewest observations of.

You already know this because you have watched an "unprecedented" event happen for the third time in a decade.

---

BRIEF 6 — The Break-Even Return Period

The idea. One line decides whether a reserve is worth holding.

        loss the event causes  ×  share of it the reserve prevents
   T* = ───────────────────────────────────────────────────────────
                    all-in annual carrying cost

T is an interval in years. Hold the reserve if the event arrives more often than T. Put your own incident log beside it and the argument finishes.

Worked example — the chips. The 2021 auto chip shortage cost the industry $210bn of revenue (AlixPartners). A three-month chip buffer, on 77m vehicles at $500 of semiconductor each, is $9.62bn held, carrying at 25 percent = $2.41bn a year.

Four correlated shocks in 2011–2022 is an interval of 3.0 years against a break-even of 4.2. On these assumptions, it pays.

Why it matters. It moves the conversation from values to frequency, and frequency is something an operations team already has data on.

---

BRIEF 7 — The Honest Negative: Over-Reserving Loses

The idea. A buffer earns less than the business. Over twenty years that spread compounds, and compounding does not forgive caution.

Worked example. ROIC 12 percent; a competitor holds 20 percent of capital in cash at 2 percent.

   buffered growth = 0.80×12% + 0.20×2% = 10.00%      drag = 2.00 pts/yr
   over 20 years:  1.12²⁰ = 9.646×   vs   1.10²⁰ = 6.727×
   the lean firm ends 43.39% larger

Now make ruin absorbing and solve for the annual ruin probability at which they draw level:

   (1−p)²⁰ × 9.646 = 6.727   →   p* = 1.7857% per year  =  once in 56.0 years

The threshold moves with the buffer's own return:

Why it matters. Read the last row. The trade-off is not efficiency against resilience. It is efficiency against DEAD reserves, and closing that spread is the whole of the second half of the chapter.

---

BRIEF 8 — Ulanowicz's Window, Handled Carefully

The idea. On a flow network, ascendency A measures organised, efficient throughput and development capacity C bounds it. With a = A/C, robustness is modelled as R = −a·ln a, which is zero at both extremes and peaks at a = 1/e = 0.3679.

What is robust about it. An interior optimum exists. A system tuned entirely for efficiency is brittle; a system tuned entirely for redundancy is inert. Both failures appear on the same curve, and that is a genuinely useful shape.

What is not. Three cautions, and they are usually omitted.

1. 0.368 is a property of the chosen function, not a measurement. Any function zero at both ends with one interior peak peaks somewhere.
2. The peak is a plateau. Within 10 percent of maximum, a may run from 0.216 to 0.544 — a band 0.328 wide, 0.89 times a* itself. Within 1 percent it still runs 0.317–0.421.
3. a moves when the analyst redraws the network. Flow diversity on k equal compartments is ln k: 0.693 at two compartments, 2.773 at sixteen. Aggregation is a choice of the observer and it moves the denominator.

Why it matters. Used as a shape, it is one of the best arguments in ecological economics. Used as a target to three decimal places, it is false precision, and false precision is how a good instrument gets discredited.

---

BRIEF 9 — Living Reserves: Buffers with Negative Carry

The idea. Cash, inventory and spare machines cost the spread to hold. Soil, skill, relationship and repairability produce in ordinary years while standing by, so their carrying cost is at or below zero.

Worked example — soil. Each 1 percent of soil organic matter holds about 20,000 US gallons of water per acre (USDA NRCS). Convert it into the unit a farmer forecasts in:

   20,000 gal × 3.785412 = 75,708 litres per acre
   ÷ 4,046.86 m²        = 18.71 mm of rainfall equivalent
   three points          = 56.12 mm  =  11.2 days of a summer crop at 5 mm/day

On the Rodale Institute's drought differential of 31 percent, a 10 t/ha crop at $200/t is $620 per hectare in a drought year — and the organic matter that held the water raises the wet-year yield too.

Worked example — skill. Six million German workers on short-time work in

1. German unemployment rose 0.90 points; American rose 11.20 — a factor

of 12.4. At a replacement cost of seven and a half months' salary on €45,000, that cohort represents roughly $168.8bn of rehiring not done.

The catch, stated plainly. A living reserve is correlated with the shock it insures against. The soil is on the same farm as the drought; the skill is in the same firm as the downturn. That correlation is exactly what makes it cheap and exactly what can make it fail when needed — so it is a layer, never the whole reserve.

---

BRIEF 10 — Who Pays, and Why Reserves Must Be Mandated, Mutualised or Alive

The idea. A reserve's cost lands on the holder. Its benefit often lands somewhere else. That gap is not a lament; it is a design specification.

Worked example. In 2021 General Motors' revenue fell 7.43 percent while its net income rose 48.83 percent — margin from 4.91 to 7.89 percent. Used vehicle prices in the US rose 45.2 percent in the year to June 2021. On 40 million transactions at $25,000, that is an illustrative $452bn paid by buyers in one market in one year, against a chip buffer that would have cost the industry $2.41bn a year to hold.

A shortage is a price event, and the producer stands on the selling side of it. No private carrying-cost calculation will close that gap.

So reserves that survive at scale are one of three things.

1. Mandated — bank capital, EU gas storage fill targets, strategic petroleum stocks.
2. Mutualised — reinsurance, a pooled regional buffer, a cooperative's shared reserve fund, an industry captive.
3. Regenerative — soil, skill, repairability, where the holder captures enough of the ordinary-year return to fund the extraordinary-year cover.

Why it matters. If a proposed reserve is none of the three, it will be cut, and it will be cut by someone behaving rationally. Design for that on day one rather than appealing against it in year four.

---

All figures in these briefs are computed in lib/verify/III_10.py, with every input, unit and source printed, and every assumed parameter labelled.