Guest Post by Milan Adams
A pension fund in the Midlands runs a cash sleeve its investment committee still calls “dry powder,” a phrase that has survived eleven consecutive quarterly reports. The label flatters the arrangement. Cash inside a written mandate behaves as a quantity only while the rule governing it stays distant; approach the floor and the sleeve turns into a constraint, and a constrained holder makes prices rather than taking them.
Every manager of other people’s money under a published rule shares that architecture, which explains why they sell in the same week. Compliance manuals describe four separate rulebooks. Economically they reduce to one.
Contractual floors are binary and stated. Risk-budget floors compute themselves: a volatility overlay divides a target by realised variance and holds the quotient as leverage, so the cash share climbs on its own as the tape deteriorates. Legal floors operate through deliberately narrow definitions — the liquidity coverage ratio counts reserves and sovereign paper and little else, leaving a bank’s regulatory “cash” far smaller than anything its treasury desk would recognise as such. Discretionary appetite floors bind hardest in bad weather, because the official who wrote them sits on the committee asked to explain a breach.
Each family produces its own elasticity, and diversification across them runs close to zero. Mandates stop dead and then scramble past the floor. Volatility maps accelerate continuously, since the leverage function is convex in the forecast. The coverage ratio breeds hoarding, the numerator being scarce by construction. Appetite statements breed imitation, because risk committees read the same sell-side note on the same Tuesday afternoon.
Convergence shows up cleanly in the data. Funds with near-identical mandates hold cash ratios spread across two or three percentage points in quiet years; across the second and third weeks of a stress episode that dispersion collapses onto the floors. Coordination played no part. Identical arithmetic pointed everyone at the same number, and the arithmetic sat in the mandate.
Volatility targeting rewards study because the algebra is public. Exposure follows
Lt=min(σ∗σ^t, Lmax)Lt=min(σ^tσ∗,Lmax)
so doubling the volatility forecast halves the book and doubles the cash share before anybody forms a view. Three properties of that map are separate failure modes. Convexity makes the tenth volatility point cost twice the fifth. Endogeneity means the selling lifts realised variance, which lifts the required cut, which produces more selling. The cap at $L_{max}$ lets leverage fall freely and rise back only to the ceiling, so the strategy ratchets downward through a crisis and enters the recovery underweight.
Aggregate size is unknowable to the dollar, and the ignorance is structural. Vol-targeted books hide inside wider mandates, rarely register as a strategy, and get inferred from flow-return elasticities. Those elasticities cluster anyway: published work puts the implied de-risking at 1.5 to 3 times the change in target leverage during volatility shocks, with the multiple widening in the tail. Whatever the population, the resulting demand curve slopes the wrong way, and falling prices repel the very capital that should be catching them.
Risk parity generalises the same construction across asset classes and concentrates leverage where variance sits lowest, which means duration. When bond volatility tripled in the opening week of March 2020 the structure had to de-lever, and had to de-lever the long end hardest, because that is where the borrow lived. Estimated unwinds run from $50 billion to several hundred billion depending on the assumed population. Direction survives the uncertainty: those cuts landed in the sector simultaneously absorbing record coupon supply, and Darrell Duffie’s order-book work at Stanford put the collapse in Treasury market depth near 80 percent from pre-crisis levels by mid-March.
Value-at-Risk replaces variance with a quantile and adds two wrinkles. Estimates come off a lookback window, so the measure lags the shock and then overshoots once the shock enters the window. Layered beneath it sits a stress or expected-shortfall overlay binding at a lower probability and a heavier loss, planting a deeper kink under the first.
Nesting is what commentary usually misses. Portfolios carry stacks rather than rules — mandate, VaR limit, stress limit, liquidity bucket, appetite statement — layered like floors of a building. The stack binds in sequence as stress deepens, each layer adding a slab of selling at a defined stress level. Measured daily, liquidation out of a leveraged book reads as a staircase, and the steps are large. Post-mortems and the occasional fund disclosure have attributed a quarter or more of gross exposure to single constraints.
Regulation built a floor with a pathology of its own. A bank at 115 percent coverage holds fifteen points of headroom before it must act, and acting means selling liquid assets or refusing to roll funding, both of which tighten conditions for the next bank. Supervisors reward a preserved buffer; the market remembers who sold. Banks therefore behave procyclically at the exact hour their balance sheets carry most value, a trade-off chosen deliberately in the calibration and then forgotten in the commentary.
Eligibility narrows the problem further. High-quality liquid assets form a short ladder of sovereigns, reserves and a thin band of agency paper, with corporates, equities and most securitisations excluded or haircut away. Regulatory floors concentrate liquidity demand onto a handful of instruments instead of widening it. During March 2020 dealers absorbed record issuance while deposit drawdowns ate their buffer stacks and volatility compressed their internal limits, and the Federal Reserve’s answer — hundreds of billions a month in purchases plus the temporary exclusion of Treasuries and reserves from the supplementary leverage ratio — amounted to an intervention against the ratio as much as against the pathogen.
That exclusion lapsed in March 2021, and the same weakness resurfaced two years later in purer form: a group of regional banks running close enough to the minimum that a modest mark on a held-to-maturity book turned into a solvency question. Duration killed those institutions. The ratio fixed the distance at which duration becomes fatal.
Money market funds sit where the rule meets the run. Their cash allocation is a regulatory artefact — weekly liquid asset minimums, a board’s power to levy fees, redemption at par on demand — and the buffer doubles as the trigger. Shareholders of a fund approaching the threshold face a simple trade: redeem at par today and leave the cost of selling illiquid paper to the stayers. Rational individually, a stampede collectively.
Prime funds shed roughly a fifth of their assets across two weeks in March 2020 as buffers slid toward the threshold. A fee-and-gate framework built as a circuit breaker turned into an incentive to leave before the gate shut. The Money Market Mutual Fund Liquidity Facility attacked the buffer rather than the credit, lending at par against fund assets so the sale became unnecessary. Amendments in 2023 traded gates for mandatory fees and lifted minimums, which shifted the timing of the run while leaving its direction intact. Par redemption plus a threshold buffer equals a run.
Paul McCulley named the underlying arrangement “shadow banking” in 2007: maturity transformation outside the deposit insurance perimeter, funded at par, backed by assets that are anything but. Nearly two decades later the perimeter has migrated and the arrangement persists. What changed is the density of rules packed inside it.
Density is where the analysis becomes estimable. Spread holders along a cash-ratio axis, give each a floor $\underline{c}$ drawn from a density $g$, then let a price decline cut cash ratios by $\Delta$. Forced sellers total $G(\underline{c}_0+\Delta)-G(\underline{c}_0)$, and price impact scales with how much of $g$ the interval traverses. Call $g$ the threshold density. Dense regions turn a two percent move into a flood; sparse regions absorb the same move without a murmur.
Lumpiness is manufactured. Floors follow convention and imitation, clustering at round numbers — six, five, ten — so $g$ emerges multimodal with spikes sitting on the conventions. A shock landing between spikes triggers little; one large enough to reach the next spike triggers a discontinuity. Calibrated simulations throw sharp local maxima at the conventional levels, each maximum scaling with the share of the population that adopted the convention. Where a single mandate dominates — public pension minimums, insurer reserve rules, the coverage ratio — the spike grows tall enough to appear in the impact function.
Data requirements are specific: a cash-ratio panel, a floor list, and flow observations fine enough to time a crossing. Form N-PORT supplies monthly holdings for registered funds and the ratio follows arithmetically; Form PF gives quarterly private-fund exposure; call reports deliver bank balance sheets a quarter late. Prospectuses and regulatory text yield the floors. Daily flow proxies exist for exchange-traded vehicles and weekly assets for money funds, allowing interpolation between quarterly anchors with a stated error term. Crossings land inside a week or so, which suffices at monthly resolution and fails at anything faster.
Three contaminations need handling. Floors depend on history, since a manager who has run eight percent for a decade writes five rather than six. Reported cash carries collateral and settlement balances and overstates the discretionary buffer. The panel is censored from below, funds that breach and die dropping out, which drags the estimated density down near the minimum. A Tobit correction with the floor as the censoring point recovers a latent profile badly — floors vary by fund, survivorship interacts with censoring — yet something usable emerges: a broad mass between eight and fifteen percent, a steep climb below seven, a thin tail at three to five occupied by the already-sick.
Impact is convex in the volume meeting it, which explains why the closing tranche of a forced sale costs more than every earlier tranche combined. Written out, $\Delta p = \lambda (Q/ADV)^{\theta}$, with $\theta$ around 0.5 to 1.0 in liquid conditions and above one under stress. Risk models calibrated in calm periods understate that exponent by a wide margin, and the miss is one reason they fail.
Inventory supplies the machinery. A dealer absorbing flow holds it, finances it, hedges it; position variance grows with the position while the hedge gets harder to place. Past the inventory limit quotes widen, then disappear, and the flow must find a natural counterparty or die unfilled. Kyle’s lambda linearises the relationship and the Amihud ratio — absolute return per dollar of volume — serves as the low-frequency workhorse because it demands only daily closes. Both blend information with inventory and both spike mechanically on large-return days, yet their peaks land on the right weeks.
Amihud’s practical message concerns the tail. A fund raising five percent of net asset value clears most of it cheaply and then pays dearly for the final percent, the point where the dealer sits full and the natural bid has already been filled. Marginal cost overtakes the marginal benefit of holding, and the sale runs past the constraint that caused it. Overshoot is where the benchmark loop ignites: the forced print becomes the valuation, and the valuation sets the next margin call.
Late in 2025 the eligible universe acquired a competitor that few had modelled. Dollar stablecoin issuance, channelled by the GENIUS Act’s reserve requirements into T-bills and repo, turned a corner of the crypto market into a structural buyer of the same short paper banks count as liquid. Whether that demand behaves as a stabiliser or as a second procyclical sleeve depends on redemption mechanics nobody has yet stress-tested, and on a Treasury stock whose composition is being quietly rewritten by the same legislation.
Crude sitting in a pipeline belongs to somebody and trades for nothing. Neither does the two metres of oil at the base of a tank, the cushion gas holding a salt cavern above the pressure where deliverability dies, or the grain keeping an elevator’s belts loaded. All of it enters the published stock figure. Little of it can move, and the distance between the printed number and the movable one explains every squeeze on the record.
Pressure keeps linefill in place. Draw a line below its minimum and hydraulic capacity goes with it, restoring it takes days of pushing product through low-pressure pipe, and the tariff passes the cost to the shipper. Heels persist because the outlet sits above the floor plate and a tank drained to zero entrains sediment and loses its roof seal; drawing past the heel costs a maintenance event rather than a discount, so operators treat it as a wall and breach it only when the alternative is worse. Cushion gas keeps back-pressure in a reservoir that would otherwise deliver a fraction of nameplate through the withdrawal season — a shortfall calculable from pressure curves and discovered the hard way by European operators in the winter of 2022.
Geometry and contracts set these volumes, and neither yields to price. Throughput moves slowly, connectivity barely at all, take-or-pay obligations on annual or multi-year cycles. A supply shock therefore compresses the gap to the floor with nothing giving way, and the commodity grows more fragile as it grows more expensive — the reverse of an equity.
Run the arithmetic on a real site. Forty tanks at two percent heel and a hundred thousand cubic metres of nameplate leave eighty thousand cubic metres immovable before a barrel trades, a full day of a large refinery’s run. Linefill and loading-rack buffer add several days of cover on top. Analysts reading the total stock print and calling the market well supplied describe oil already bolted to the machinery.
Behaviour near the wall gives the game away. The marginal barrel above the heel carries the maintenance event avoided plus the optionality surrendered, so spot spikes against the deferred and the spread quotes an invoice rather than a forecast. Many small tanks and heavy throughput throw steep backwardation, because maintenance cost per barrel runs high; few large tanks throw shallow backwardation for the opposite reason. Cleaning the tank and refilling collapses the spread, and any model reading the inversion as demand will be wrong twice.
Metals add exchange warehouse rules and a tighter second wall. Deliverable shapes in approved locations form a sliver of global stock, and part of that sliver sits under financing that will hold at any price. The float is the commodity analogue of a share free float and it decides squeezes. Global copper stocks run to millions of tonnes; registered London tonnage in a tight month runs to tens of thousands; the share neither financed nor held by the dominant long has fallen to a few thousand, days of world consumption.
Ownership stays hidden. Exchanges publish totals and cancelled warrants, and cancellation reveals only that registered stock has been spoken for and will leave. How much sits with the party also long the prompt — the single question a short cares about — remains invisible. Hunts in silver, copper through 2005-06, tin in 2021, nickel in 2022, alumina and copper in 2024, and London silver again in October 2025: each episode put the float below prompt open interest while shorts were still deciding.
Warehouse economics push the same way. Rent accrues on stored metal and full sheds beat empty ones, so load-out rules are drafted to protect the rent. London requires a minimum daily outturn once metal is cancelled, a figure tightened after the queue scandals of the 2010s, yet a long queue still means months of waiting. Queued tonnage trades at a discount equal to the present value of the wait, and in a squeeze the discount inverts for accepted shapes and locations because the short has no say over provenance. The spread between tiers can exceed the metal’s value.
Procyclicality follows from the queue itself. A nearly full warehouse has little room to receive, so queues lengthen fastest when stocks run highest; falling stocks shorten the queue and shrink the deliverable discount at the precise moment the short needs it to persist. Nothing about rebuilding aggregate stock restores the float once accepted producers are disrupted.
October 2025 made that point in silver. London free float thinned to a fraction of the paper claim sitting against it, lease rates exploded to double digits annualised, and metal crossed the Atlantic by cargo plane to settle arbitrage against New York. Front spreads inverted hard enough to make borrowing uneconomic, and the squeeze fed on the physical logistics of moving bars between two vault systems rather than on any change in mine supply. Silver had traded above $50 an ounce once before, in 1980 and again briefly in 2011; the 2025 version ran on plumbing.
Nickel in March 2022 fused both walls. A single large short sat concentrated in prompt dates against a float of one accepted briquette made in one disrupted country. The price doubled within a session and touched $100,000 a tonne intraday while deliverable stock moved by zero. Margin calls ran to the billions, the exchange suspended trading, cancelled $3.9 billion of prints above the prior close, and reopened under limit-down that took days to grind through.
Argument over the cancellation persists, and each side holds part of the truth. Longs lost gains, shorts lost losses, and the question was whether those prints qualified as prices. Liquidation artefacts, said the exchange. A squeeze print remains a print, said the counter, and retroactive repricing teaches the market that squeezes can be rewritten. Tighter margin methodology, mandatory off-exchange position reporting and a revised deliverable specification followed. The float problem outlived all three, the specification still leaning on a narrow producer set, and by 2026 the exchange was designing a replacement contract around a different deliverable entirely.
Cocoa in 2024 ran on genuine deficit and reproduced the anatomy. Disease, ageing trees, illegal mining on farmland and weather cut West African output; certified stock hit a multi-decade low; grinders held much of the remainder for their own runs. Front spreads inverted past what a short without physical cover could carry, the price travelled from about $2,500 to above $12,000 a tonne inside a year, and margin rose several-fold on the way. Two harvests later, prices had come off the peak while certified stock stayed historically thin — the deficit eased, the float did not.
Information sat in the open throughout. Crop forecasts, disease maps and weekly certified stock were all published. The market possessed the data and produced a violent move regardless, because the data described the physical balance while fragility lived in the float and the margin grid. Supply shortness says little about where a price stops; near the floor the price quotes the cost of avoiding default, and production cost becomes irrelevant. Grinders ran to working inventory and paid, shorts hit margin walls and were carried out, and the spread drove the capitulation.
Carry formalises the relationship:
Ft,T=St⋅e(r+w−y)(T−t)Ft,T=St⋅e(r+w−y)(T−t)
with $y$ the convenience yield — negligible when stocks run ample, large when they run scarce. At the operational minimum $y$ stops being latent and turns nearly observable, because the marginal yield equals the maintenance event avoided plus the optionality surrendered. Backwardation there quotes an invoice for a tank cleaning, scaling with the count of tanks at heel and the throughput dependent on them.
Clearing houses watch the front spread for that reason. Inversion past a threshold triggers margin on calendar positions, unwinds follow, and the inversion steepens. The expiring May 2020 WTI contract pushed the front spread beyond $50 and brokers began refusing new spread business. Storage decisions near the floor stop being marginal and turn binary — keep the tank in service or take it out — and small price changes move the answer by zero.
April 2020 remains the cleanest demonstration that the floor is mechanical. On the 20th the expiring WTI contract settled at minus $37.63. Delivery at Cushing fell due within the month, Cushing sat effectively full, and longs at expiry holding neither tankage nor connection nor refinery had one move: pay somebody who did. The payment became the price. Heels drive the sign, because a receiving tank needs room above its heel and aggregate room above the heels across the complex had dropped below open interest at expiry.
Fraud played no part, exchange failure played no part, and the demand collapse was weeks old and already priced. A wall bound while a delivery obligation ran, and the price detached from every positive value. Smaller versions recur in hub gas, nodal power and harvest grain where elevator capacity runs dry. Any commodity model unable to print a negative number at full utilisation carries a specification error, because the true cost function holds no floor at zero.
Zoltan Pozsar, then at Credit Suisse, wrote about plumbing through the same weeks rather than about barrels: reserve scarcity, the repo market’s structural shortfall, the $3-3.5 trillion reserve stock beneath which settlement starts to seize. That figure belongs to the same species as a tank heel, fixed by settlement mechanics rather than preference, and it explains why a central bank balance sheet became a market variable instead of a policy residual. Four years on, with QT finished and the standing repo facility drawing routine use, the threshold had moved into open view.
Analysis usually separates cash from inventory, and collateral is where they collide. Futures positions demand margin in cash or haircut securities, calls land within hours, and the trader funds, sells collateral or cuts. Which of the three happens follows from the trader’s cash ratio against its own floor.
Physical traders meet the compound version. Inventory hedged short is economically flat and operationally ruinous: the stock loses, the future gains, and the gain stays unrealised while the call turns real. Cash must be raised while goods sit unsold, selling into the market that produced the call realises the loss the hedge deferred, and the position unwinds at the worst hour available. Traders who grasp this hold buffers larger than exposure implies, which lifts their cost of capital and slowly drives them out in favour of better-funded rivals. Physical trading concentrates into fewer, larger hands and the marginal buyer loses variety.
Central counterparties occupy the middle and run procyclical by design. Initial margin covers a close-out over a holding period at a high quantile; holding periods lengthen when liquidity looks poor while the quantile stays fixed and bites hard in the tail. Models look backward by necessity, so a sixty-day window fills with the shock only after sixty days of it and the largest calls follow the largest moves. Margin symmetry has occupied regulators since 2018 and the lag has narrowed without closing.
Nickel made the lag visible. London’s nickel requirement multiplied inside days, increases landed on shorts already bleeding, intraday calls ran to hundreds of millions against members who passed them to clients, and at least one client book failed to meet them inside the cycle. Trading halted partly to give the margin system time to catch the price. Follow the ordering, because the ordering carries the mechanism: price moved on a thin float, margin rose on the price, shorts sold on the margin, price moved further. Deliverable stock held constant throughout while a physical constraint passed through a financial amplifier and emerged an order of magnitude larger.
Variation margin opens a second channel by acting as transfer rather than buffer. A large move can push the transfer past the loser’s liquid resources while net worth stays positive, so the failure arrives as illiquidity, which moves faster than insolvency because daily settlement permits no negotiation. Every cash floor in the system doubles as a variation-margin floor.
Dealers form the hinge and capital governs their capacity. The supplementary leverage ratio pins total exposure, gross repo book and reserves included, against tier one capital, so intermediating customer flow expands the balance sheet one-for-one. Binding ratios leave the dealer choosing between intermediation and capital, and the price of intermediation climbs until the marginal trade clears at the required return on leverage. Repo shows the price directly: overnight rates touched 10 percent intraday and SOFR jumped to 5.25 percent in September 2019 as tax dates, settlement and a depleted reserve stock collided with the ratio binding across the dealer community. Bill purchases relieved it by adding reserves.
Excluding Treasuries and reserves from the ratio in 2020 achieved the same result in one stroke, dealer Treasury inventories recovering inside weeks and off-the-run spreads narrowing behind them. Lapse a year later coincided with a fresh climb in the balance-sheet cost of intermediation and widening in the sectors carrying the most inventory. Identical customer flow clears at wildly different prices according to the capital price, which tracks the leverage denominator rather than the asset’s liquidity. By late 2025 the calibration had become a live policy question, with proposals to soften the ratio’s treatment of low-risk assets and the dealer community pricing the probability of relief into its willingness to warehouse.
Hyun Song Shin’s BIS work on global liquidity describes the offshore leg of the same machine. Dollars fund most world trade, non-US banks fund dollar assets through swaps and repo, and the cross-currency basis quotes the premium. A negative basis means scarcity, arriving whenever the balance sheets that supply dollars are pinned. March 2020 drove the euro-dollar basis to several hundred basis points at the short end; swap lines and the FIMA repo facility supplied dollars against local collateral and collapsed it within days.
Asymmetry makes the basis worth studying. Widening raises the arbitrage profit and the balance-sheet intensity together, so closing capacity falls as the incentive climbs, and past a threshold the trade stops and the basis quotes pure scarcity. That threshold follows from the arbitrageur’s leverage ratio and its own cash floor. Profitable in expectation and unfundable in the interim, the trade demonstrates limits to arbitrage in its textbook form, quoted live. Swap lines work by supplying the one input private desks cannot manufacture — unconstrained official balance sheet at a known price — and they need little volume, since credibility alone stops the hoarding that makes the basis a scarcity price.
Three weeks of March 2020 compressed the whole apparatus and remain the best natural experiment available. An exogenous shock travelled almost entirely through threshold mechanics: volatility tripled, systematic programs cut, prime funds shed a fifth of assets as buffers neared the fee threshold, dealers absorbed record issuance with constrained balance sheets and ballooning inventories, the basis blew out, commodity traders met hedge calls by selling stock, and gold fell because leveraged holders were raising cash against it. Off-the-run Treasuries dislocated from on-the-run by tens of basis points, worst in issues held most heavily by levered relative-value funds, and repo in those issues seized. Early purchases were too small to matter. Later ones were enormous, and the facilities that did most of the work — the MMLF, the credit facilities — bought modest volume and moved the market on announcement, a credible backstop removing the need to sell.
Britain’s LDI episode two years later serves as the control, holding no physical component at all. Funds ran long gilts against derivative overlays, levered in repo and total return swaps, with collateral waterfalls demanding cash at short notice and buffers sized to a decade of low volatility. The mini-budget moved thirty-year yields about 160 basis points inside days. Calls outran the pace of gilt sales, since selling into the fall realised losses and the funds were the largest holders; trustees faced their own floors and began redeeming. The Bank of England bought from 28 September with a stated pace and a stated end date, and the date was the clever element, an open-ended facility inviting waiting while a dated one forces the trade. Yields fell within hours. Roughly £65 billion nominal was taken and mostly unwound at trivial cost, and the fragility had stayed invisible in every conventional measure until the week it detonated. Final buffer standards for the sector followed from the regulator and took effect through 2025-26, raising the cost of leverage across the pension system — a genuine trade-off the post-crisis commentary largely skipped.
European gas in 2022 mirrors the physical side and was managed with an explicit threshold. Cushion gas forms a hard engineering floor and a soft commercial one, breached rather than failing the system. The supply cut gutted the injection season, storage entered autumn far short, and TTF ran from about €30 a megawatt-hour in early 2021 to above €300 in August 2022 with the front spread inverting on the impossibility of refilling. Governments legislated a target — 80 percent by 1 October, 90 percent by 1 November — with a duty on operators and money behind it. Storage hit the number and the winter cleared.
Three legs made the design work. Set the floor above the engineering minimum with margin, date it, fund it. A target without a date invites deferral, a date without money invites default, a floor without margin invites a breach. Payment came in shuttered fertilizer plants, smelters and glass furnaces, and some capacity never returned, which separates a physical floor from a financial one: defending the physical floor destroys something. By the 2025-26 winter the target itself had been relaxed to 83 percent, the buffer having done its job and the cost of holding it having become the binding consideration.
Four loops recur across every case, and naming them makes comparison mechanical. Price move, margin call, cash draw, sale, further move forms the collateral loop, its gain given by the margin-to-buffer ratio across the leveraged population. A forced print at an index weight becomes the valuation, the valuation sets the next call and the next redemption — the benchmark loop, strongest where the seller is large against the book. Poor performance drives outflows, outflows force sales, sales worsen performance: the redemption loop, damped by gates and notice periods that make a fund less attractive in calm weather. Low stocks lift convenience yield, yield steepens backwardation, backwardation discourages storage, stocks fall — the physical loop, whose gain runs high in normal conditions and collapses near the floor, where storage becomes maintenance.
Compounding separates a sharp move from a disorderly one. A collateral loop inside a thin float turns into a benchmark loop, the forced seller setting the print. A redemption loop inside a fund holding physical stock turns into a physical loop, the sale being inventory. Damping one while others stay open can beat damping none, and can also lose to it: a facility supporting prices without relieving margin becomes exit liquidity, while one relieving margin without touching the float clears the financial leg and abandons the physical one.
Index construction is feasible on soft parameters and still worth doing. Four components, each inverted so higher reads worse: distance from the cash floor, taken as the cross-sectional median gap between current ratios and stated or implied floors; distance from the physical floor, taken as deliverable stock in days of cover; intermediation thinness, taken as dealer inventory against average daily volume plus spread against its trailing median; collateral stress, taken as the margin-to-buffer ratio plus the cross-currency basis. Standardise on a rolling window, weight by inverse historical volatility, aggregate, and publish the components separately, since the composite conceals which loop is live.
Three calibration choices carry the index. A window of three to five years spans a cycle without blurring a regime, and belongs beside the printed value. Weights should come from each component’s historical contribution to drawdown rather than from taste, which makes the index partly in-sample and demands an annual refit. Distance terms need confidence bands, resting as they do on floors that stay partly unobservable and stock series that overstate what can move.
Ten diagnostics merit tracking alongside it, a composite being worth only what its parts are worth:
Ranking is what the index does well. Top-decile readings have preceded every disorderly episode in the sample; bottom-half readings have preceded none of them, which justifies publication provided the lag and the error sit beside the number. Timing stays out of reach, the trigger arriving from outside the model and arriving smaller than expected. Inside a fragile state the outcome belongs to the state rather than the trigger: May 2020 went to minus $37 because tanks were full, while the shock itself was weeks old and already priced. Nickel ran on a thin float. Cocoa went vertical on a multi-decade low in certified stock, and each move surprised only those who had skipped the tonnage count.
Policy therefore holds three real levers and a long list of false ones. Raising minimums charges the whole book a carry cost for tail protection, blunt and honest. Making buffers usable targets the protection, but only where supervisors tolerate use, a frowned-upon buffer being decoration. Standing facilities price insurance at the point of use and are the most efficient, and they are the most politically exposed because use is visible while absence is invisible. The Fed’s 2021 Standing Repo Facility is the model and its flaw is stigma: the rate sits above market in calm weather, routine use stays rare, fluency stays thin, and the first crisis use reads as a signal. A rate near market would fix it.
Physical policy is thinner. Reserve release moves slowly and depends on politics; deliverable-specification reform at the exchanges cuts deep, broadening shapes and locations widening the float and lowering the squeeze premium at the price of a wider contract basis; position limits work only on deliverable-adjusted rather than gross exposure, which almost nobody does. Every version trades the same pair: tighten and the market turns slow and expensive in calm weather, loosen and it turns fast and cheap and prone to the tail. Stated tail frequency with a quantified and accepted carry cost beats an aspiration.
Aim beats calibration as a design principle. Attack the loop, leave the price alone. Buying the asset relieves the benchmark loop and leaves collateral; lending against the asset relieves collateral and lets the price find its level. The second wins almost always, preserving the signal while removing the constraint.
Data limits everything above, and the gaps cluster where the framework needs most: who holds the float and on what terms, and what the margin-to-buffer ratio looks like across the leveraged population. Both are known afterward and unknown during. Cash ratios arrive quarterly for most institutions, floors get disclosed for registered products and inferred elsewhere, deliverable stock appears only for exchange-registered tonnage and gets surveyed with multi-week lags for the remainder, dealer inventory exists weekly for primary dealers and nowhere else, and margin reaches clearing houses and supervisors and practically nobody in public.
Three disclosures would move it: weekly cash reporting with the floor stated for systemically important funds, a delivery-intent reporting requirement at the exchanges published with a lag, and an anonymised quarterly margin-to-buffer disclosure for clearing members. Each reveals something an institution would rather keep, and none is technically hard.
Floors move as well, which no index captures. Committees waive mandate minimums at two in the morning. Decrees lift storage targets. Specifications get broadened over a weekend. Capacity to move the floor amounts to resilience, spread unevenly and correlated with size and political access, and it stays out of the risk report until somebody uses it.
The useful question is therefore never what will happen, but how far the floors sit and how thin the layer between them runs. Both are measurable, both surface in fragments, and both get ignored until the print arrives. At Cushing in April 2020 the room above the heels fell below open interest at expiry and the price of a barrel became a bill. That is what a floor does when it is reached. The screen read minus $37.63, and every tank holding the number had been telling anyone who counted where the price was headed for a fortnight.
Tweet
Continue reading...
[ H/T The Burning Platform ]
A pension fund in the Midlands runs a cash sleeve its investment committee still calls “dry powder,” a phrase that has survived eleven consecutive quarterly reports. The label flatters the arrangement. Cash inside a written mandate behaves as a quantity only while the rule governing it stays distant; approach the floor and the sleeve turns into a constraint, and a constrained holder makes prices rather than taking them.
Every manager of other people’s money under a published rule shares that architecture, which explains why they sell in the same week. Compliance manuals describe four separate rulebooks. Economically they reduce to one.
Contractual floors are binary and stated. Risk-budget floors compute themselves: a volatility overlay divides a target by realised variance and holds the quotient as leverage, so the cash share climbs on its own as the tape deteriorates. Legal floors operate through deliberately narrow definitions — the liquidity coverage ratio counts reserves and sovereign paper and little else, leaving a bank’s regulatory “cash” far smaller than anything its treasury desk would recognise as such. Discretionary appetite floors bind hardest in bad weather, because the official who wrote them sits on the committee asked to explain a breach.
Each family produces its own elasticity, and diversification across them runs close to zero. Mandates stop dead and then scramble past the floor. Volatility maps accelerate continuously, since the leverage function is convex in the forecast. The coverage ratio breeds hoarding, the numerator being scarce by construction. Appetite statements breed imitation, because risk committees read the same sell-side note on the same Tuesday afternoon.
Convergence shows up cleanly in the data. Funds with near-identical mandates hold cash ratios spread across two or three percentage points in quiet years; across the second and third weeks of a stress episode that dispersion collapses onto the floors. Coordination played no part. Identical arithmetic pointed everyone at the same number, and the arithmetic sat in the mandate.
Volatility targeting rewards study because the algebra is public. Exposure follows
Lt=min(σ∗σ^t, Lmax)Lt=min(σ^tσ∗,Lmax)
so doubling the volatility forecast halves the book and doubles the cash share before anybody forms a view. Three properties of that map are separate failure modes. Convexity makes the tenth volatility point cost twice the fifth. Endogeneity means the selling lifts realised variance, which lifts the required cut, which produces more selling. The cap at $L_{max}$ lets leverage fall freely and rise back only to the ceiling, so the strategy ratchets downward through a crisis and enters the recovery underweight.
Aggregate size is unknowable to the dollar, and the ignorance is structural. Vol-targeted books hide inside wider mandates, rarely register as a strategy, and get inferred from flow-return elasticities. Those elasticities cluster anyway: published work puts the implied de-risking at 1.5 to 3 times the change in target leverage during volatility shocks, with the multiple widening in the tail. Whatever the population, the resulting demand curve slopes the wrong way, and falling prices repel the very capital that should be catching them.
Risk parity generalises the same construction across asset classes and concentrates leverage where variance sits lowest, which means duration. When bond volatility tripled in the opening week of March 2020 the structure had to de-lever, and had to de-lever the long end hardest, because that is where the borrow lived. Estimated unwinds run from $50 billion to several hundred billion depending on the assumed population. Direction survives the uncertainty: those cuts landed in the sector simultaneously absorbing record coupon supply, and Darrell Duffie’s order-book work at Stanford put the collapse in Treasury market depth near 80 percent from pre-crisis levels by mid-March.
Value-at-Risk replaces variance with a quantile and adds two wrinkles. Estimates come off a lookback window, so the measure lags the shock and then overshoots once the shock enters the window. Layered beneath it sits a stress or expected-shortfall overlay binding at a lower probability and a heavier loss, planting a deeper kink under the first.
Nesting is what commentary usually misses. Portfolios carry stacks rather than rules — mandate, VaR limit, stress limit, liquidity bucket, appetite statement — layered like floors of a building. The stack binds in sequence as stress deepens, each layer adding a slab of selling at a defined stress level. Measured daily, liquidation out of a leveraged book reads as a staircase, and the steps are large. Post-mortems and the occasional fund disclosure have attributed a quarter or more of gross exposure to single constraints.
Regulation built a floor with a pathology of its own. A bank at 115 percent coverage holds fifteen points of headroom before it must act, and acting means selling liquid assets or refusing to roll funding, both of which tighten conditions for the next bank. Supervisors reward a preserved buffer; the market remembers who sold. Banks therefore behave procyclically at the exact hour their balance sheets carry most value, a trade-off chosen deliberately in the calibration and then forgotten in the commentary.
Eligibility narrows the problem further. High-quality liquid assets form a short ladder of sovereigns, reserves and a thin band of agency paper, with corporates, equities and most securitisations excluded or haircut away. Regulatory floors concentrate liquidity demand onto a handful of instruments instead of widening it. During March 2020 dealers absorbed record issuance while deposit drawdowns ate their buffer stacks and volatility compressed their internal limits, and the Federal Reserve’s answer — hundreds of billions a month in purchases plus the temporary exclusion of Treasuries and reserves from the supplementary leverage ratio — amounted to an intervention against the ratio as much as against the pathogen.
That exclusion lapsed in March 2021, and the same weakness resurfaced two years later in purer form: a group of regional banks running close enough to the minimum that a modest mark on a held-to-maturity book turned into a solvency question. Duration killed those institutions. The ratio fixed the distance at which duration becomes fatal.
Money market funds sit where the rule meets the run. Their cash allocation is a regulatory artefact — weekly liquid asset minimums, a board’s power to levy fees, redemption at par on demand — and the buffer doubles as the trigger. Shareholders of a fund approaching the threshold face a simple trade: redeem at par today and leave the cost of selling illiquid paper to the stayers. Rational individually, a stampede collectively.
Prime funds shed roughly a fifth of their assets across two weeks in March 2020 as buffers slid toward the threshold. A fee-and-gate framework built as a circuit breaker turned into an incentive to leave before the gate shut. The Money Market Mutual Fund Liquidity Facility attacked the buffer rather than the credit, lending at par against fund assets so the sale became unnecessary. Amendments in 2023 traded gates for mandatory fees and lifted minimums, which shifted the timing of the run while leaving its direction intact. Par redemption plus a threshold buffer equals a run.
Paul McCulley named the underlying arrangement “shadow banking” in 2007: maturity transformation outside the deposit insurance perimeter, funded at par, backed by assets that are anything but. Nearly two decades later the perimeter has migrated and the arrangement persists. What changed is the density of rules packed inside it.
Density is where the analysis becomes estimable. Spread holders along a cash-ratio axis, give each a floor $\underline{c}$ drawn from a density $g$, then let a price decline cut cash ratios by $\Delta$. Forced sellers total $G(\underline{c}_0+\Delta)-G(\underline{c}_0)$, and price impact scales with how much of $g$ the interval traverses. Call $g$ the threshold density. Dense regions turn a two percent move into a flood; sparse regions absorb the same move without a murmur.
Lumpiness is manufactured. Floors follow convention and imitation, clustering at round numbers — six, five, ten — so $g$ emerges multimodal with spikes sitting on the conventions. A shock landing between spikes triggers little; one large enough to reach the next spike triggers a discontinuity. Calibrated simulations throw sharp local maxima at the conventional levels, each maximum scaling with the share of the population that adopted the convention. Where a single mandate dominates — public pension minimums, insurer reserve rules, the coverage ratio — the spike grows tall enough to appear in the impact function.
Data requirements are specific: a cash-ratio panel, a floor list, and flow observations fine enough to time a crossing. Form N-PORT supplies monthly holdings for registered funds and the ratio follows arithmetically; Form PF gives quarterly private-fund exposure; call reports deliver bank balance sheets a quarter late. Prospectuses and regulatory text yield the floors. Daily flow proxies exist for exchange-traded vehicles and weekly assets for money funds, allowing interpolation between quarterly anchors with a stated error term. Crossings land inside a week or so, which suffices at monthly resolution and fails at anything faster.
Three contaminations need handling. Floors depend on history, since a manager who has run eight percent for a decade writes five rather than six. Reported cash carries collateral and settlement balances and overstates the discretionary buffer. The panel is censored from below, funds that breach and die dropping out, which drags the estimated density down near the minimum. A Tobit correction with the floor as the censoring point recovers a latent profile badly — floors vary by fund, survivorship interacts with censoring — yet something usable emerges: a broad mass between eight and fifteen percent, a steep climb below seven, a thin tail at three to five occupied by the already-sick.
Impact is convex in the volume meeting it, which explains why the closing tranche of a forced sale costs more than every earlier tranche combined. Written out, $\Delta p = \lambda (Q/ADV)^{\theta}$, with $\theta$ around 0.5 to 1.0 in liquid conditions and above one under stress. Risk models calibrated in calm periods understate that exponent by a wide margin, and the miss is one reason they fail.
Inventory supplies the machinery. A dealer absorbing flow holds it, finances it, hedges it; position variance grows with the position while the hedge gets harder to place. Past the inventory limit quotes widen, then disappear, and the flow must find a natural counterparty or die unfilled. Kyle’s lambda linearises the relationship and the Amihud ratio — absolute return per dollar of volume — serves as the low-frequency workhorse because it demands only daily closes. Both blend information with inventory and both spike mechanically on large-return days, yet their peaks land on the right weeks.
Amihud’s practical message concerns the tail. A fund raising five percent of net asset value clears most of it cheaply and then pays dearly for the final percent, the point where the dealer sits full and the natural bid has already been filled. Marginal cost overtakes the marginal benefit of holding, and the sale runs past the constraint that caused it. Overshoot is where the benchmark loop ignites: the forced print becomes the valuation, and the valuation sets the next margin call.
Late in 2025 the eligible universe acquired a competitor that few had modelled. Dollar stablecoin issuance, channelled by the GENIUS Act’s reserve requirements into T-bills and repo, turned a corner of the crypto market into a structural buyer of the same short paper banks count as liquid. Whether that demand behaves as a stabiliser or as a second procyclical sleeve depends on redemption mechanics nobody has yet stress-tested, and on a Treasury stock whose composition is being quietly rewritten by the same legislation.
The inventory nobody can touch: linefill, tank heels, cushion gas and the deliverable float — why the physical floor is an engineering fact that prices cannot move, and why it detonates through margin
Crude sitting in a pipeline belongs to somebody and trades for nothing. Neither does the two metres of oil at the base of a tank, the cushion gas holding a salt cavern above the pressure where deliverability dies, or the grain keeping an elevator’s belts loaded. All of it enters the published stock figure. Little of it can move, and the distance between the printed number and the movable one explains every squeeze on the record.
Pressure keeps linefill in place. Draw a line below its minimum and hydraulic capacity goes with it, restoring it takes days of pushing product through low-pressure pipe, and the tariff passes the cost to the shipper. Heels persist because the outlet sits above the floor plate and a tank drained to zero entrains sediment and loses its roof seal; drawing past the heel costs a maintenance event rather than a discount, so operators treat it as a wall and breach it only when the alternative is worse. Cushion gas keeps back-pressure in a reservoir that would otherwise deliver a fraction of nameplate through the withdrawal season — a shortfall calculable from pressure curves and discovered the hard way by European operators in the winter of 2022.
Geometry and contracts set these volumes, and neither yields to price. Throughput moves slowly, connectivity barely at all, take-or-pay obligations on annual or multi-year cycles. A supply shock therefore compresses the gap to the floor with nothing giving way, and the commodity grows more fragile as it grows more expensive — the reverse of an equity.
Run the arithmetic on a real site. Forty tanks at two percent heel and a hundred thousand cubic metres of nameplate leave eighty thousand cubic metres immovable before a barrel trades, a full day of a large refinery’s run. Linefill and loading-rack buffer add several days of cover on top. Analysts reading the total stock print and calling the market well supplied describe oil already bolted to the machinery.
Behaviour near the wall gives the game away. The marginal barrel above the heel carries the maintenance event avoided plus the optionality surrendered, so spot spikes against the deferred and the spread quotes an invoice rather than a forecast. Many small tanks and heavy throughput throw steep backwardation, because maintenance cost per barrel runs high; few large tanks throw shallow backwardation for the opposite reason. Cleaning the tank and refilling collapses the spread, and any model reading the inversion as demand will be wrong twice.
Metals add exchange warehouse rules and a tighter second wall. Deliverable shapes in approved locations form a sliver of global stock, and part of that sliver sits under financing that will hold at any price. The float is the commodity analogue of a share free float and it decides squeezes. Global copper stocks run to millions of tonnes; registered London tonnage in a tight month runs to tens of thousands; the share neither financed nor held by the dominant long has fallen to a few thousand, days of world consumption.
Ownership stays hidden. Exchanges publish totals and cancelled warrants, and cancellation reveals only that registered stock has been spoken for and will leave. How much sits with the party also long the prompt — the single question a short cares about — remains invisible. Hunts in silver, copper through 2005-06, tin in 2021, nickel in 2022, alumina and copper in 2024, and London silver again in October 2025: each episode put the float below prompt open interest while shorts were still deciding.
Warehouse economics push the same way. Rent accrues on stored metal and full sheds beat empty ones, so load-out rules are drafted to protect the rent. London requires a minimum daily outturn once metal is cancelled, a figure tightened after the queue scandals of the 2010s, yet a long queue still means months of waiting. Queued tonnage trades at a discount equal to the present value of the wait, and in a squeeze the discount inverts for accepted shapes and locations because the short has no say over provenance. The spread between tiers can exceed the metal’s value.
Procyclicality follows from the queue itself. A nearly full warehouse has little room to receive, so queues lengthen fastest when stocks run highest; falling stocks shorten the queue and shrink the deliverable discount at the precise moment the short needs it to persist. Nothing about rebuilding aggregate stock restores the float once accepted producers are disrupted.
October 2025 made that point in silver. London free float thinned to a fraction of the paper claim sitting against it, lease rates exploded to double digits annualised, and metal crossed the Atlantic by cargo plane to settle arbitrage against New York. Front spreads inverted hard enough to make borrowing uneconomic, and the squeeze fed on the physical logistics of moving bars between two vault systems rather than on any change in mine supply. Silver had traded above $50 an ounce once before, in 1980 and again briefly in 2011; the 2025 version ran on plumbing.
Nickel in March 2022 fused both walls. A single large short sat concentrated in prompt dates against a float of one accepted briquette made in one disrupted country. The price doubled within a session and touched $100,000 a tonne intraday while deliverable stock moved by zero. Margin calls ran to the billions, the exchange suspended trading, cancelled $3.9 billion of prints above the prior close, and reopened under limit-down that took days to grind through.
Argument over the cancellation persists, and each side holds part of the truth. Longs lost gains, shorts lost losses, and the question was whether those prints qualified as prices. Liquidation artefacts, said the exchange. A squeeze print remains a print, said the counter, and retroactive repricing teaches the market that squeezes can be rewritten. Tighter margin methodology, mandatory off-exchange position reporting and a revised deliverable specification followed. The float problem outlived all three, the specification still leaning on a narrow producer set, and by 2026 the exchange was designing a replacement contract around a different deliverable entirely.
Cocoa in 2024 ran on genuine deficit and reproduced the anatomy. Disease, ageing trees, illegal mining on farmland and weather cut West African output; certified stock hit a multi-decade low; grinders held much of the remainder for their own runs. Front spreads inverted past what a short without physical cover could carry, the price travelled from about $2,500 to above $12,000 a tonne inside a year, and margin rose several-fold on the way. Two harvests later, prices had come off the peak while certified stock stayed historically thin — the deficit eased, the float did not.
Information sat in the open throughout. Crop forecasts, disease maps and weekly certified stock were all published. The market possessed the data and produced a violent move regardless, because the data described the physical balance while fragility lived in the float and the margin grid. Supply shortness says little about where a price stops; near the floor the price quotes the cost of avoiding default, and production cost becomes irrelevant. Grinders ran to working inventory and paid, shorts hit margin walls and were carried out, and the spread drove the capitulation.
Carry formalises the relationship:
Ft,T=St⋅e(r+w−y)(T−t)Ft,T=St⋅e(r+w−y)(T−t)
with $y$ the convenience yield — negligible when stocks run ample, large when they run scarce. At the operational minimum $y$ stops being latent and turns nearly observable, because the marginal yield equals the maintenance event avoided plus the optionality surrendered. Backwardation there quotes an invoice for a tank cleaning, scaling with the count of tanks at heel and the throughput dependent on them.
Clearing houses watch the front spread for that reason. Inversion past a threshold triggers margin on calendar positions, unwinds follow, and the inversion steepens. The expiring May 2020 WTI contract pushed the front spread beyond $50 and brokers began refusing new spread business. Storage decisions near the floor stop being marginal and turn binary — keep the tank in service or take it out — and small price changes move the answer by zero.
April 2020 remains the cleanest demonstration that the floor is mechanical. On the 20th the expiring WTI contract settled at minus $37.63. Delivery at Cushing fell due within the month, Cushing sat effectively full, and longs at expiry holding neither tankage nor connection nor refinery had one move: pay somebody who did. The payment became the price. Heels drive the sign, because a receiving tank needs room above its heel and aggregate room above the heels across the complex had dropped below open interest at expiry.
Fraud played no part, exchange failure played no part, and the demand collapse was weeks old and already priced. A wall bound while a delivery obligation ran, and the price detached from every positive value. Smaller versions recur in hub gas, nodal power and harvest grain where elevator capacity runs dry. Any commodity model unable to print a negative number at full utilisation carries a specification error, because the true cost function holds no floor at zero.
Zoltan Pozsar, then at Credit Suisse, wrote about plumbing through the same weeks rather than about barrels: reserve scarcity, the repo market’s structural shortfall, the $3-3.5 trillion reserve stock beneath which settlement starts to seize. That figure belongs to the same species as a tank heel, fixed by settlement mechanics rather than preference, and it explains why a central bank balance sheet became a market variable instead of a policy residual. Four years on, with QT finished and the standing repo facility drawing routine use, the threshold had moved into open view.
Where the two floors meet: collateral spirals, clearing-house margin, dealer balance sheets and the offshore dollar — the four loops that turn a small trigger into a disorderly market, and the only three levers that damp them
Analysis usually separates cash from inventory, and collateral is where they collide. Futures positions demand margin in cash or haircut securities, calls land within hours, and the trader funds, sells collateral or cuts. Which of the three happens follows from the trader’s cash ratio against its own floor.
Physical traders meet the compound version. Inventory hedged short is economically flat and operationally ruinous: the stock loses, the future gains, and the gain stays unrealised while the call turns real. Cash must be raised while goods sit unsold, selling into the market that produced the call realises the loss the hedge deferred, and the position unwinds at the worst hour available. Traders who grasp this hold buffers larger than exposure implies, which lifts their cost of capital and slowly drives them out in favour of better-funded rivals. Physical trading concentrates into fewer, larger hands and the marginal buyer loses variety.
Central counterparties occupy the middle and run procyclical by design. Initial margin covers a close-out over a holding period at a high quantile; holding periods lengthen when liquidity looks poor while the quantile stays fixed and bites hard in the tail. Models look backward by necessity, so a sixty-day window fills with the shock only after sixty days of it and the largest calls follow the largest moves. Margin symmetry has occupied regulators since 2018 and the lag has narrowed without closing.
Nickel made the lag visible. London’s nickel requirement multiplied inside days, increases landed on shorts already bleeding, intraday calls ran to hundreds of millions against members who passed them to clients, and at least one client book failed to meet them inside the cycle. Trading halted partly to give the margin system time to catch the price. Follow the ordering, because the ordering carries the mechanism: price moved on a thin float, margin rose on the price, shorts sold on the margin, price moved further. Deliverable stock held constant throughout while a physical constraint passed through a financial amplifier and emerged an order of magnitude larger.
Variation margin opens a second channel by acting as transfer rather than buffer. A large move can push the transfer past the loser’s liquid resources while net worth stays positive, so the failure arrives as illiquidity, which moves faster than insolvency because daily settlement permits no negotiation. Every cash floor in the system doubles as a variation-margin floor.
Dealers form the hinge and capital governs their capacity. The supplementary leverage ratio pins total exposure, gross repo book and reserves included, against tier one capital, so intermediating customer flow expands the balance sheet one-for-one. Binding ratios leave the dealer choosing between intermediation and capital, and the price of intermediation climbs until the marginal trade clears at the required return on leverage. Repo shows the price directly: overnight rates touched 10 percent intraday and SOFR jumped to 5.25 percent in September 2019 as tax dates, settlement and a depleted reserve stock collided with the ratio binding across the dealer community. Bill purchases relieved it by adding reserves.
Excluding Treasuries and reserves from the ratio in 2020 achieved the same result in one stroke, dealer Treasury inventories recovering inside weeks and off-the-run spreads narrowing behind them. Lapse a year later coincided with a fresh climb in the balance-sheet cost of intermediation and widening in the sectors carrying the most inventory. Identical customer flow clears at wildly different prices according to the capital price, which tracks the leverage denominator rather than the asset’s liquidity. By late 2025 the calibration had become a live policy question, with proposals to soften the ratio’s treatment of low-risk assets and the dealer community pricing the probability of relief into its willingness to warehouse.
Hyun Song Shin’s BIS work on global liquidity describes the offshore leg of the same machine. Dollars fund most world trade, non-US banks fund dollar assets through swaps and repo, and the cross-currency basis quotes the premium. A negative basis means scarcity, arriving whenever the balance sheets that supply dollars are pinned. March 2020 drove the euro-dollar basis to several hundred basis points at the short end; swap lines and the FIMA repo facility supplied dollars against local collateral and collapsed it within days.
Asymmetry makes the basis worth studying. Widening raises the arbitrage profit and the balance-sheet intensity together, so closing capacity falls as the incentive climbs, and past a threshold the trade stops and the basis quotes pure scarcity. That threshold follows from the arbitrageur’s leverage ratio and its own cash floor. Profitable in expectation and unfundable in the interim, the trade demonstrates limits to arbitrage in its textbook form, quoted live. Swap lines work by supplying the one input private desks cannot manufacture — unconstrained official balance sheet at a known price — and they need little volume, since credibility alone stops the hoarding that makes the basis a scarcity price.
Three weeks of March 2020 compressed the whole apparatus and remain the best natural experiment available. An exogenous shock travelled almost entirely through threshold mechanics: volatility tripled, systematic programs cut, prime funds shed a fifth of assets as buffers neared the fee threshold, dealers absorbed record issuance with constrained balance sheets and ballooning inventories, the basis blew out, commodity traders met hedge calls by selling stock, and gold fell because leveraged holders were raising cash against it. Off-the-run Treasuries dislocated from on-the-run by tens of basis points, worst in issues held most heavily by levered relative-value funds, and repo in those issues seized. Early purchases were too small to matter. Later ones were enormous, and the facilities that did most of the work — the MMLF, the credit facilities — bought modest volume and moved the market on announcement, a credible backstop removing the need to sell.
Britain’s LDI episode two years later serves as the control, holding no physical component at all. Funds ran long gilts against derivative overlays, levered in repo and total return swaps, with collateral waterfalls demanding cash at short notice and buffers sized to a decade of low volatility. The mini-budget moved thirty-year yields about 160 basis points inside days. Calls outran the pace of gilt sales, since selling into the fall realised losses and the funds were the largest holders; trustees faced their own floors and began redeeming. The Bank of England bought from 28 September with a stated pace and a stated end date, and the date was the clever element, an open-ended facility inviting waiting while a dated one forces the trade. Yields fell within hours. Roughly £65 billion nominal was taken and mostly unwound at trivial cost, and the fragility had stayed invisible in every conventional measure until the week it detonated. Final buffer standards for the sector followed from the regulator and took effect through 2025-26, raising the cost of leverage across the pension system — a genuine trade-off the post-crisis commentary largely skipped.
European gas in 2022 mirrors the physical side and was managed with an explicit threshold. Cushion gas forms a hard engineering floor and a soft commercial one, breached rather than failing the system. The supply cut gutted the injection season, storage entered autumn far short, and TTF ran from about €30 a megawatt-hour in early 2021 to above €300 in August 2022 with the front spread inverting on the impossibility of refilling. Governments legislated a target — 80 percent by 1 October, 90 percent by 1 November — with a duty on operators and money behind it. Storage hit the number and the winter cleared.
Three legs made the design work. Set the floor above the engineering minimum with margin, date it, fund it. A target without a date invites deferral, a date without money invites default, a floor without margin invites a breach. Payment came in shuttered fertilizer plants, smelters and glass furnaces, and some capacity never returned, which separates a physical floor from a financial one: defending the physical floor destroys something. By the 2025-26 winter the target itself had been relaxed to 83 percent, the buffer having done its job and the cost of holding it having become the binding consideration.
Four loops recur across every case, and naming them makes comparison mechanical. Price move, margin call, cash draw, sale, further move forms the collateral loop, its gain given by the margin-to-buffer ratio across the leveraged population. A forced print at an index weight becomes the valuation, the valuation sets the next call and the next redemption — the benchmark loop, strongest where the seller is large against the book. Poor performance drives outflows, outflows force sales, sales worsen performance: the redemption loop, damped by gates and notice periods that make a fund less attractive in calm weather. Low stocks lift convenience yield, yield steepens backwardation, backwardation discourages storage, stocks fall — the physical loop, whose gain runs high in normal conditions and collapses near the floor, where storage becomes maintenance.
Compounding separates a sharp move from a disorderly one. A collateral loop inside a thin float turns into a benchmark loop, the forced seller setting the print. A redemption loop inside a fund holding physical stock turns into a physical loop, the sale being inventory. Damping one while others stay open can beat damping none, and can also lose to it: a facility supporting prices without relieving margin becomes exit liquidity, while one relieving margin without touching the float clears the financial leg and abandons the physical one.
Index construction is feasible on soft parameters and still worth doing. Four components, each inverted so higher reads worse: distance from the cash floor, taken as the cross-sectional median gap between current ratios and stated or implied floors; distance from the physical floor, taken as deliverable stock in days of cover; intermediation thinness, taken as dealer inventory against average daily volume plus spread against its trailing median; collateral stress, taken as the margin-to-buffer ratio plus the cross-currency basis. Standardise on a rolling window, weight by inverse historical volatility, aggregate, and publish the components separately, since the composite conceals which loop is live.
Three calibration choices carry the index. A window of three to five years spans a cycle without blurring a regime, and belongs beside the printed value. Weights should come from each component’s historical contribution to drawdown rather than from taste, which makes the index partly in-sample and demands an annual refit. Distance terms need confidence bands, resting as they do on floors that stay partly unobservable and stock series that overstate what can move.
Ten diagnostics merit tracking alongside it, a composite being worth only what its parts are worth:
- Median cash ratio against the stated or implied floor, by fund category.
- Share of the population within one percentage point of its floor.
- Realised volatility against the level where the leverage map’s second derivative turns.
- HQLA headroom at the largest dealers, and its change over five sessions.
- Deliverable stock in days of cover, separated from total stock.
- Cancelled warrant ratio and queue length at the accepted locations.
- Front spread against the maintenance cost of the marginal tank.
- Initial margin as a multiple of the trailing quarter, by contract.
- Cross-currency basis at the three-month tenor.
- Dealer inventory divided by average daily volume, inverted.
Ranking is what the index does well. Top-decile readings have preceded every disorderly episode in the sample; bottom-half readings have preceded none of them, which justifies publication provided the lag and the error sit beside the number. Timing stays out of reach, the trigger arriving from outside the model and arriving smaller than expected. Inside a fragile state the outcome belongs to the state rather than the trigger: May 2020 went to minus $37 because tanks were full, while the shock itself was weeks old and already priced. Nickel ran on a thin float. Cocoa went vertical on a multi-decade low in certified stock, and each move surprised only those who had skipped the tonnage count.
Policy therefore holds three real levers and a long list of false ones. Raising minimums charges the whole book a carry cost for tail protection, blunt and honest. Making buffers usable targets the protection, but only where supervisors tolerate use, a frowned-upon buffer being decoration. Standing facilities price insurance at the point of use and are the most efficient, and they are the most politically exposed because use is visible while absence is invisible. The Fed’s 2021 Standing Repo Facility is the model and its flaw is stigma: the rate sits above market in calm weather, routine use stays rare, fluency stays thin, and the first crisis use reads as a signal. A rate near market would fix it.
Physical policy is thinner. Reserve release moves slowly and depends on politics; deliverable-specification reform at the exchanges cuts deep, broadening shapes and locations widening the float and lowering the squeeze premium at the price of a wider contract basis; position limits work only on deliverable-adjusted rather than gross exposure, which almost nobody does. Every version trades the same pair: tighten and the market turns slow and expensive in calm weather, loosen and it turns fast and cheap and prone to the tail. Stated tail frequency with a quantified and accepted carry cost beats an aspiration.
Aim beats calibration as a design principle. Attack the loop, leave the price alone. Buying the asset relieves the benchmark loop and leaves collateral; lending against the asset relieves collateral and lets the price find its level. The second wins almost always, preserving the signal while removing the constraint.
Data limits everything above, and the gaps cluster where the framework needs most: who holds the float and on what terms, and what the margin-to-buffer ratio looks like across the leveraged population. Both are known afterward and unknown during. Cash ratios arrive quarterly for most institutions, floors get disclosed for registered products and inferred elsewhere, deliverable stock appears only for exchange-registered tonnage and gets surveyed with multi-week lags for the remainder, dealer inventory exists weekly for primary dealers and nowhere else, and margin reaches clearing houses and supervisors and practically nobody in public.
Three disclosures would move it: weekly cash reporting with the floor stated for systemically important funds, a delivery-intent reporting requirement at the exchanges published with a lag, and an anonymised quarterly margin-to-buffer disclosure for clearing members. Each reveals something an institution would rather keep, and none is technically hard.
Floors move as well, which no index captures. Committees waive mandate minimums at two in the morning. Decrees lift storage targets. Specifications get broadened over a weekend. Capacity to move the floor amounts to resilience, spread unevenly and correlated with size and political access, and it stays out of the risk report until somebody uses it.
The useful question is therefore never what will happen, but how far the floors sit and how thin the layer between them runs. Both are measurable, both surface in fragments, and both get ignored until the print arrives. At Cushing in April 2020 the room above the heels fell below open interest at expiry and the price of a barrel became a bill. That is what a floor does when it is reached. The screen read minus $37.63, and every tank holding the number had been telling anyone who counted where the price was headed for a fortnight.
Tweet
Continue reading...
[ H/T The Burning Platform ]
