Orderbook

The two-day gap

Mark Hughes — Sun, 30 Aug 2026 09:00:00 +0000

A trade is agreed in microseconds and settled in days. Between those two moments sits an obligation that exists, is legally binding, and has not yet been performed by either side.

Most of the time this is invisible. When it stops being invisible, it stops gradually and then all at once.

What settlement actually is

Two things have to happen for a trade to be finished: the asset moves, and the money moves. Doing them at literally the same instant is hard, so systems are built to make them conditional on each other — delivery versus payment — so that neither side can end up having performed while the other has not.

The gap between trade and settlement is where the machinery does its work: netting offsetting trades, arranging financing, locating the asset, moving collateral. Shortening the gap does not remove the work. It compresses it.

Margin is the price of the gap

If I owe you an asset in two days, you are exposed to my failure to deliver for two days. That exposure has a value, and the clearing house charges for it in advance, in the form of margin.

Margin is therefore not a tax on trading. It is the funded, prepaid answer to a specific question: if this participant vanished right now, what would it cost to replace their obligations at market prices?

Which produces the awkward feature of the whole arrangement:

Margin requirements rise with volatility. Volatility rises in a crisis. The system therefore asks for the most cash at the precise moment cash is hardest to find.

This is not an oversight, and it is not fixable by wishing. The exposure really is larger when prices are moving. But it means margin is structurally pro-cyclical, and every serious discussion of clearing reform is, underneath, an argument about how much pro-cyclicality to accept and who should absorb it.

The chain nobody draws

The part that consistently surprises people is how long the chain is between “I own this” and the record that actually says so.

A retail account holds a position with a broker. The broker holds it with a custodian. The custodian holds it in an account at a central securities depository. Possibly there are more links, in more jurisdictions, in more time zones. Each link is a promise, and each promise is only as good as the entity making it.

In normal conditions this is a plumbing detail. In stress it becomes the whole question, because the speed at which an asset can be located and moved — not its price, not its liquidity — determines whether it can be used as collateral when collateral is what you need.

Three things worth keeping

Instant trading and slow settlement is a deliberate trade. The gap buys netting, financing and error correction. It costs credit exposure.
Margin is pro-cyclical by construction. Any proposal that claims to eliminate this is really proposing to move the cost somewhere less visible.
Collateral mobility beats collateral quality in a crisis. The best asset in the world is worthless to you on Tuesday if it settles on Thursday.

That is the plumbing. Next week, back above ground.

One book, fourteen venues

Sun, 23 Aug 2026 09:00:00 +0000

Last week the argument was that a quote is an option. This week: what happens when the same instrument is quoted in a dozen places at once, by participants who cannot all see each other at the same instant.

Fragmentation is usually discussed as though it were a policy mistake. It is more useful to treat it as a physics problem that policy then has to live with.

The consolidated book is a fiction you compute

There is no single book. There are many books, each maintained by a venue that knows only its own state, publishing updates that arrive at every other participant at different times.

The “national best bid and offer”, or whatever your jurisdiction calls it, is not a thing that exists somewhere. It is a number each participant calculates from the messages they have received so far. Two participants with different connectivity compute different values for it at the same wall-clock instant, and both are correct given what they know.

Everything awkward about fragmented markets follows from that.

Latency is not a speed problem

The usual framing is that fast participants beat slow ones to the trade. True, but it undersells it.

The real asymmetry is that a fast participant knows the state of the market sooner. If venue A trades at 100.90 and you are quoting 100.85 bid on venue B, your quote is now stale — not wrong when you posted it, but wrong now — and the question is only whether you learn that before someone acts on it.

That is the option from last week, with a new expiry: the time it takes news of one venue to reach another. Shorten it and quotes tighten. Lengthen it and every resting order across every venue must be priced for a longer window of ignorance.

Fragmentation does not create the adverse-selection problem. It sets the clock on it.

What it actually costs

Three costs, in rough order of how much they are talked about versus how much they matter:

Cost	Borne by	Talked about
Connectivity and market data	Intermediaries	Constantly
Stale-quote risk across venues	Liquidity providers	Sometimes
Complexity of proving best execution	Everyone	Rarely, and reluctantly

The first is a real expense and shows up in fee schedules, which is why it dominates the discussion. The second is priced into every spread you have ever paid and appears on no invoice. The third is the one that quietly shapes how brokers route, because a routing decision that is hard to defend afterwards is a routing decision that does not get made.

The case for it anyway

Given all that, why not consolidate?

Because a single venue is a single point of failure, a single fee schedule, and a single operator with no particular reason to improve. Competition between venues is what produced continuous matching, sub-penny economics, and the operational reliability that fragmented markets are rarely credited with. The costs above are real. They are also the price of not having one owner of the only door.

The honest summary is that fragmentation trades a visible, measurable cost for an invisible, structural benefit — which is exactly the trade that markets are worst at defending in public.

Next week: settlement, and the two days where none of the above applies.

What a quote actually costs

Sun, 16 Aug 2026 09:00:00 +0000

Every resting limit order is an option you have written and given away for free. Post a bid at 100.85 and you have granted the entire market the right — not the obligation — to sell you something at 100.85, for as long as the order sits there. Whoever exercises it will do so precisely when it suits them and not you.

That single observation explains most of what bid-ask spreads do, and it is worth taking slowly.

The option nobody prices explicitly

Options are priced off volatility and time. So is a quote.

Volatility raises the value of the option you have written, because the price is more likely to move through your order before you can pull it. Time works the same way: the longer your order rests, the more chances the market has to find a reason to hit it. A market maker who cannot cancel quickly is writing a longer-dated option than one who can, and must charge more for it.

This is why spreads widen into news. Nothing about the asset’s value has necessarily changed at the moment of widening — but the cost of standing still has.

Two kinds of counterparty

The uncomfortable part is that the option is not exercised at random.

Consider two people who might hit your bid. The first needs cash today and does not much care about the next tick. The second has worked out that the price is about to fall. Both trades look identical on the tape. Only one of them costs you money, and you find out which afterwards.

Market makers call the second kind adverse selection, which is a polite name for being systematically on the wrong side. The spread is what gets charged to everybody in order to survive the subset who know more than you do. The uninformed seller subsidises the informed one. That is not a flaw in the design; it is the design.

A spread is not a fee for the service of trading. It is the premium on an option you did not choose to write, priced for the counterparty you hope you do not meet.

Why the top of book lies

A displayed size of 5,000 at the touch does not mean 5,000 is available. It means 5,000 was available at the moment the message left the exchange, from participants who can cancel in microseconds and will, the instant your intention becomes visible.

The gap between displayed and accessible liquidity is not deception. It is the direct consequence of the option: a quote that cannot be pulled is a quote that must be much wider, so a market of fast, cancellable, tight quotes is exactly what you should expect a healthy venue to produce. The liquidity is real. It is just contingent on you not being the person it is afraid of.

What to take from it

Three things follow, and they are the sort of thing worth checking against your own experience of a market:

Spreads are a volatility instrument. When they widen, ask what changed about uncertainty, not what changed about value.
Depth is a snapshot, not a promise. Size that vanishes as you reach for it was never mispriced — it was priced against a different counterparty.
Passive is not free. A resting order earns the spread and pays for it in adverse selection. Whether that trade is good depends entirely on who is trading against you.

Next week: what happens to all of this when the same instrument trades in fourteen places at once.

The cost of being in a hurry

Sun, 09 Aug 2026 09:00:00 +0000

Twenty issues in, and the pieces assemble into one idea.

The spread, market impact, the borrow fee, the haircut, the margin call, the rebate you paid rather than earned — almost every cost in this letter has been a payment for having something happen sooner than it otherwise would.

Trading costs are the price of immediacy. Understanding that reframes the whole problem.

The two costs that move in opposite directions

Suppose you must buy a large position. You have one real decision: how fast.

Trade it all at once and you pay impact. You are demanding immediacy from a market that has only so much of it available at any moment, and you will walk up the book, take every offer, and reveal your intention to everyone watching. The faster you go, the more you pay.

Trade it slowly and impact falls. Each small piece meets a market that has had time to replenish, and you look like ordinary flow. But now you are exposed to timing risk — the price moving for reasons that have nothing to do with you, for as long as you remain incomplete. The slower you go, the more of that you carry.

These two costs move in opposite directions, and there is no setting that eliminates both. Every execution schedule ever devised is a position on that trade-off, and the right position depends on something no algorithm can know: how confident you are, and how quickly you expect to be proved right or wrong.

Impact is what you pay for going fast. Timing risk is what you pay for going slow. You are not choosing between a good outcome and a bad one — you are choosing which of two certain costs you would rather bear.

Why urgency is the real variable

This is why the same order can be traded correctly in two opposite ways.

An order arising from information with a short shelf life should be fast. Impact is a certain cost, and it is the smaller one when the alternative is being right about something the market discovers before you finish.

An order arising from a slow allocation change should be patient. There is no clock. Paying impact to be finished by Tuesday buys nothing, because Tuesday was not important.

Both orders may be for the same stock, in the same size, on the same day. The only thing that differs is why, and why is the input that determines everything — which is exactly the input that never appears in an execution report.

Where the costs came from

Look back and the sources of impact are all mechanisms from earlier issues.

The depth you consume is a market maker’s inventory capacity — finite, and already partly used up by whoever traded before you this morning. The fees you pay or earn depend on whether you supplied immediacy or demanded it. The information you leak is what turns you from anonymous flow into somebody worth pricing against. And the spread you cross is compensation to the person who was standing there, for a risk we have referred to repeatedly and never properly priced.

None of these are frictions in the engineering sense — imperfections to be polished away. Each is a real service, provided by someone bearing a real risk, priced accordingly. A market with no trading costs would be a market where nobody was compensated for standing ready to trade, and the immediacy you were enjoying for free would not be there.

The one asymmetry worth remembering

There is a final piece, and it is the least intuitive.

Your costs depend on what everybody else is doing at the same time. Impact is larger when others are trading in your direction, because you are competing for the same finite capacity. Timing risk is larger when the market is moving fast.

Both worsen together, in the same conditions, for the same reasons — and those conditions are precisely the ones in which portfolios most need adjusting.

The cost of trading is therefore not a constant to be budgeted. It is a variable that peaks exactly when you have the least choice about whether to pay it.

Three things worth keeping

Every trading cost is a price for immediacy. Ask how much you actually need, because that is the only thing you control.
Impact and timing risk trade off against each other. There is no schedule that avoids both, only a choice about which to bear.
Costs rise with urgency and with everyone else’s urgency at once. They are highest when your options are fewest.

Next week: the same cost, from the other side. What does it cost the person on the other end of your trade to have been standing there at all?

Was that a good fill?

Sun, 02 Aug 2026 09:00:00 +0000

You bought a hundred thousand shares at an average of 47.32. Was that good?

The question is unanswerable as posed, because “good” requires a comparison, and every available comparison is wrong in a different way. Which wrongness you pick is not a technicality. It is a decision about what behaviour you are rewarding.

The candidates

The arrival price — where the market was when the order reached the desk. This is the most honest measure of what the trading cost you, because it captures everything that happened after the decision was made, including the impact of your own order.

It is also brutal, and it punishes the trader for market movement they could not control. Buy into a rising market and you look terrible. Buy into a falling one and you look brilliant. Over enough orders that averages out; over one quarter it certainly does not.

Volume-weighted average price — the average price over the period, weighted by volume. VWAP is popular because it is intuitive and gameable in a comfortable direction: trade in proportion to volume and you will match it closely.

That comfort is the problem. A trader measured on VWAP has an incentive to trade slowly and in line with the market, even when the right answer was to get it done immediately. And a large order is part of the volume it is measured against — beat VWAP on an order that was most of the day’s turnover and you have largely beaten yourself.

The closing price — the benchmark for anyone tracking an index, as in issue twenty-five. Perfectly appropriate if the mandate references the close, and perfectly arbitrary otherwise.

Implementation shortfall — the full distance between the decision and the outcome. It is the most complete measure and the least comfortable, which is why it deserves its own section.

What the shortfall contains

The framing dates to Perold’s 1988 paper, and its value is that it decomposes the cost rather than producing a single number.

Spread — you crossed to trade. Small, unavoidable, easy to measure.

Impact — your own order moved the price. The larger and more urgent the order, the more of the cost this is.

Delay — the price moved between the decision and the start of trading. Not the trader’s fault, but real money.

Opportunity cost — the part of the order that never filled. This is the one almost everybody omits, and it is the one that exposes the others.

Because if you leave out the unfilled portion, patience always looks free. A trader who works an order gently, fills sixty percent at excellent prices and abandons the rest scores beautifully on every price-based benchmark. The forty percent they did not buy, in a stock that then rose, appears in no report.

Every benchmark that ignores what you failed to trade rewards giving up. The cheapest execution in the world is the one you never did, and it will win any measurement that only counts fills.

Measurement changes behaviour

This is the reason the topic matters beyond the reporting pack.

Measure on VWAP and you get slow, participatory trading, because that is what matches VWAP. Measure on arrival price and you get fast, aggressive trading, since delay is the enemy. Measure on the close and everything migrates to the closing auction — which is a substantial part of why the close became what it is.

None of those is misconduct. In each case the desk is doing exactly what it was asked to. The benchmark is not a measurement of the trading strategy; over time it becomes the trading strategy.

Which puts the real decision one level up: what did the portfolio actually need? An order arising from a time-sensitive insight needs speed and should be measured against arrival. An order arising from a slow allocation shift needs cheapness and can afford patience. Applying one benchmark to both guarantees that one of them is being traded wrongly.

The unknowable part

A final honesty. All of this compares what happened to a counterfactual — what the price would have been had you not traded. That counterfactual is unobservable. You cannot rerun the day without your order in it.

Estimating impact means modelling that counterfactual, and every impact model is a set of assumptions dressed as a measurement. The numbers are useful in aggregate and across many orders. On any single trade they are an educated guess, and should be quoted with rather less confidence than they usually are.

Three things worth keeping

Pick the benchmark that matches why you are trading, not the one that is easiest to hit.
Include opportunity cost or you are rewarding non-execution. It is the only term that penalises giving up.
Impact is modelled, not measured. The counterfactual does not exist, and the confidence in the number should reflect that.

Next week: the whole cost, from the decision to the fill — and then, the same cost seen from the other side of the trade.

The feed

Sun, 26 Jul 2026 09:00:00 +0000

Everything in this letter so far has assumed you can see the market. Prices, depth, trades, imbalances — all treated as though they simply exist.

They do not simply exist. Each is a product, assembled by a venue, sold on terms, and delivered at a speed that depends on what you paid. Market data is one of the most profitable businesses an exchange runs, and its economics shape what participants can know.

What is actually in a feed

Data comes in tiers, and the gaps between them matter.

Trades — what printed, at what price and size. The minimum.

Top of book — the best bid and offer, and how much is there. Enough to know what a small order would cost.

Depth of book — every price level with resting size. This is what you need to estimate the cost of anything larger than the touch, and it is a different product at a different price.

Full order-by-order — every individual order, its arrival, amendment and cancellation. Orders at the same price are filled in the sequence they arrived, so with this you can reconstruct that sequence and know where in it your own order sits. Without it, your position in the queue is guesswork.

Each tier is a distinct commercial product. A participant seeing only top of book is not seeing a simplified version of the market; they are seeing a different market, in which depth and queue position are invisible.

Direct against consolidated

The second axis is speed, and it produced one of the sharpest structural arguments of the last two decades.

A venue publishes directly to those who connect to it. It also contributes to a consolidated feed that aggregates every venue into one view — the thing that defines the official reference price.

Consolidation takes time. Messages must travel to the aggregator, be combined, and be redistributed. The direct feed, by construction, arrives first.

So a participant with direct connections to every venue can compute the consolidated picture before the consolidated feed reports it. They are not cheating; they are doing arithmetic the aggregator has not finished yet.

The official price is, unavoidably, slightly historical. Anyone taking the trouble to assemble it themselves knows it before the record of it exists — and a rule that references the official price is therefore always referencing the past.

That gap is small and it is the foundation of a great deal of activity. It is also why arguments about consolidated-tape latency are arguments about market structure, not about IT.

Why it costs what it costs

Exchange data pricing is contentious, and the mechanics explain why the dispute is so hard to settle.

The venue’s position is that data is a product of the matching engine, expensive to produce and distribute, and that they compete for order flow and must be allowed to earn a return.

The critics’ position is that the data describes trades made by participants, at a venue with a natural monopoly on its own book. There is no competing supplier for what happened on that venue, and a mandatory input for anyone doing business there is not a competitive market.

Both are true. The output is that the ability to see the market clearly is distributed by willingness to pay, which is a strange property for a mechanism whose whole justification is public price discovery.

What this means for reading anything

The practical consequence sits underneath every issue of this letter.

When someone tells you the spread was a penny, ask which feed, sampled how, and whether that penny existed for a millisecond or an hour. When depth is quoted, ask whether it includes hidden orders — it cannot, because they are hidden. When a trade is analysed against “the price at the time”, ask which time: the direct feed or the consolidated one, and at whose location.

None of this is pedantry. Different observers of the same market, all honest, all looking at real data, will disagree about what happened — because they bought different products.

Three things worth keeping

The tier of data determines what is knowable. Queue position, depth and cost estimates all require products most participants do not buy.
The official consolidated price is always slightly behind the picture available to anyone assembling it directly.
Ask which feed before comparing any two claims about what a market did.

Next week: how to tell whether an execution was any good, and why the honest answer is harder than it looks.

Pulling the plug

Sun, 19 Jul 2026 09:00:00 +0000

Every major market has a mechanism for stopping itself. Prices fall far enough, fast enough, and trading halts — for five minutes, for fifteen, for the rest of the day.

On its face this is peculiar. A market exists so people can trade, and the halt triggers exactly when they most want to. The justification has to be that something specific is going wrong, and it is worth being precise about what.

Two shapes

Halts come in two forms, aimed at different failures.

Single-stock bands apply to one instrument. If the price moves outside a band around a recent average, trading pauses briefly or is confined within the band. This targets the fat-finger error, the algorithm behaving badly, the thin book that momentarily has nothing in it — local failures with no wider meaning.

Market-wide breakers apply to everything at once, triggered by a broad index falling by set percentages. In the US these are 7%, 13% and 20%, with the first two producing a fifteen-minute pause and the third closing the market for the day. This targets a different animal: not an error, but a collective loss of the ability to price anything.

The argument for

The case has three strands, and they are unequal.

The strongest is the pause as a call auction. A halt does not just stop trading; it restarts it with an auction rather than a scramble. During the pause orders accumulate without executing, the venue publishes the price at which the book would clear if it uncrossed now, and trading resumes at a single price that everybody gets. Nobody has to quote first into the dark, and nobody is racing. That is a considerably better way to re-establish a price after chaos than letting a thin continuous book find its own way down — and it is a mechanism worth a whole issue of its own, which it will get.

The second is the operational one, and it is more practical than it sounds. Systems fail under load. Risk managers need to know their positions before they can authorise more trading. Firms need to confirm they can meet margin calls. Fifteen minutes to find out whether your own systems are telling you the truth is not nothing.

The third is the behavioural one — that a pause interrupts panic and lets people reconsider. This is the argument most often given publicly and the weakest of the three. The evidence that halts change the eventual price is thin. Markets that reopen after a halt frequently continue in the same direction, which is what you would expect if the information causing the fall was real.

A circuit breaker cannot make a price wrong. If the news is bad, the price is going there. What a halt can do is change the mechanism by which it gets there — auction rather than cascade — and that is a real difference even when the destination is the same.

The magnet

The strongest objection is a mechanical one and it is not easily dismissed.

If everyone knows the market halts at a particular level, then as the price approaches it, two things happen. Anyone who needs to trade becomes more urgent, because a halt would trap them. And anyone providing liquidity becomes more reluctant, because being filled just before a halt means holding a position through an interval with no exit.

Urgency on one side and withdrawal on the other is a recipe for acceleration. The threshold, intended as a floor, can act as a magnet — pulling the price towards itself faster than it would otherwise have travelled.

This is not hypothetical; it has been observed around single-stock bands. It is also an argument about design rather than existence: it suggests thresholds should be wide enough to be rarely reached and hard to anticipate precisely, not that they should be abolished.

The coordination problem

One more difficulty, and it is the one that makes halts hardest to get right.

An instrument trades in many places at once — a subject this letter will come back to properly — and a related instrument trades in many others. Halting one venue while others continue does not stop trading; it moves it, and to the least suitable places — thinner books, wider spreads, related derivatives standing in for the halted asset.

For a halt to do what it is supposed to, it has to be coordinated across every venue where the risk can be expressed. That is achievable within one jurisdiction and much harder across the derivatives, futures and international listings of the same underlying exposure.

Three things worth keeping

The reopening auction is the strongest argument for a halt — not the pause itself, but what it replaces the scramble with.
Thresholds can act as magnets, because they make demanders urgent and suppliers reluctant at the same moment.
An uncoordinated halt relocates trading rather than stopping it, usually to somewhere worse.

Next week: the data that tells you all this is happening, and why it costs so much.

When someone fails

Sun, 12 Jul 2026 09:00:00 +0000

You trade on an exchange. You have no idea who took the other side and you do not care, because within moments of the trade neither of you is the other’s counterparty any more.

A clearing house has stepped between you. How it survives one of you failing is one of the most carefully engineered arrangements in finance, and it works by being extremely specific about whose money is lost, in what order.

Novation

The starting move is a legal one. The original contract between you and your counterparty is torn up and replaced by two contracts: you with the clearing house, and the clearing house with them. This is novation.

Its effects are immediate and large.

You no longer need to assess your counterparty’s creditworthiness, which is what makes anonymous trading possible at all. Positions become fungible, so an offsetting trade with anyone genuinely closes your position rather than adding a second one. And exposures net down: if you are owed by one participant and owe another, the clearing house sees only your net.

The cost of all this is concentration. Every exposure that used to be spread across a web of bilateral relationships now points at one institution. Everything that follows is about making sure that institution does not fail.

The waterfall

The clearing house holds collateral against every position, sized to a plausible one-day move. When a member defaults, it closes out their positions — hedging, auctioning them to other members, whatever it takes to get flat.

That costs money, usually more than the defaulter’s collateral covers. The order in which the shortfall is met is fixed in advance, and it is the whole design:

The defaulter’s initial margin. Their own money, first.
The defaulter’s contribution to the default fund. Still their money.
The clearing house’s own capital — a slice deliberately placed ahead of surviving members. Skin in the game.
The mutualised default fund, contributed by every surviving member.
Assessments — further calls on surviving members, up to a defined cap.

Steps one and two are unremarkable: the party that failed pays. Step three is there for incentives, not size; a clearing house that risked none of its own money would have no reason to set margin conservatively.

Step four is the one that matters, and it deserves reading slowly.

Beyond a certain point, a member’s loss is paid by every other member of the clearing house — participants who did nothing wrong, had no relationship with the defaulter, and may not have traded that day. Membership is a mutual insurance policy that you cannot opt out of.

What mutualisation buys

That sounds alarming stated baldly, and it is a genuinely good trade.

The alternative is the bilateral world: your protection is your own assessment of each counterparty, and in a crisis every participant simultaneously discovers they cannot evaluate anyone. Markets freeze not because prices are wrong but because nobody will face anybody.

Mutualisation converts a large number of unknowable individual risks into one known, capped, collectively funded risk. It also creates something valuable: surviving members have a direct financial interest in the default being handled well, which is why they participate in default auctions that they might otherwise sit out.

Where the argument is

Two objections are worth knowing, because they are the live ones.

The first is concentration. Novation does not destroy risk; it relocates it. A clearing house is designed to be extremely hard to break, and it is also a point whose failure would be catastrophic in a way that no bilateral failure could be.

The second is pro-cyclicality, which we met with margin and again with haircuts. Clearing house margin models are risk-sensitive, so they demand more collateral as volatility rises. Individually correct. Collectively, they extract cash from the system at exactly the moment cash is scarce — and now they do it through one coordinated mechanism rather than thousands of uncoordinated ones, which makes it faster.

Neither objection has a clean answer. Both are the price of the thing working.

Three things worth keeping

Novation is what makes anonymity possible. Without it, you would need an opinion about every counterparty you ever met.
The waterfall’s order is the design. Own money, then the house’s, then everyone else’s — and knowing you are in the queue changes how you behave.
Clearing relocates risk rather than removing it, from many bilateral links to one hardened node.

Next week: what a market does when it decides it would rather not trade at all for a few minutes.

The rebalance

Sun, 05 Jul 2026 09:00:00 +0000

Last week: what happens when a list changes. This week: what happens because prices move at all.

Any portfolio held to fixed weights drifts. If you hold sixty percent equities and forty percent bonds, and equities rise, you now hold more than sixty percent equities without having done anything. Restoring the weights means selling what went up and buying what went down.

That is rebalancing, and done at scale by enough institutions on a common schedule, it becomes one of the largest sources of predictable flow in markets.

Selling strength, mechanically

The first thing to notice is the direction. Rebalancing is inherently contrarian: it sells the asset that rose and buys the one that fell, every time, without reference to any view about either.

This is genuinely stabilising, and it is one of the underappreciated services that mechanical investing provides. A large pool of capital that automatically supplies selling into rallies and buying into declines is doing what the long-gamma dealer from issue twelve does — leaning against moves, damping volatility, at no charge to anyone.

It also means that the more a market moves in a month, the larger the rebalancing flow at the end of it. The mechanism is self-scaling: violent months produce big rebalances, quiet months produce almost none.

The calendar problem

The trouble is not the rebalancing. It is that everybody does it at the same time.

Month-end and quarter-end are the conventional dates, for reasons that have nothing to do with markets — reporting periods, mandate documents, committee calendars. A pension fund’s policy might say “rebalance quarterly to policy weights”, and thousands of similar documents say something similar.

So flows that would be harmless spread across a quarter arrive in a few sessions. And because the direction is determined entirely by what markets did during the period, it is calculable in advance by anyone with a spreadsheet and a rough idea of assets under management.

A rebalance is a trade decided by a document written years ago, executed on a date chosen by an accounting convention, in a direction anyone can compute from public price data. It is the most telegraphed flow in finance.

Who pays

Predictable flow gets anticipated. If enough capital positions ahead of a month-end rebalance, the price has already moved by the time the rebalancer trades, and the difference is a transfer from the fund’s beneficiaries to whoever was waiting.

That is not a scandal — anticipating a public, foreseeable flow is legitimate — but it is a real cost borne by savers who have no idea it exists, generated by a mechanism they never chose.

The defences all amount to being less predictable. Rebalance on tolerance bands rather than dates, so the trigger depends on drift rather than the calendar. Spread execution over more days. Rebalance with new cashflows — direct contributions towards the underweight asset and pay for the adjustment with money that had to be invested anyway. Use derivatives to adjust exposure without moving the underlying.

Each works by making the flow harder to time. None removes it, because the underlying obligation to hold specific weights has not gone anywhere.

The bigger point

Step back and there is a pattern running through the last three issues.

Index inclusion, rebalancing, and the roll from issue eleven are all the same phenomenon: large, mandatory, scheduled flows generated by rules rather than opinions. In every case the participant’s job is not to trade well but to comply, and in every case the compliance is public.

Markets are usually described as a contest between people with different views. A very large share of what actually trades has no view at all — it is a document being executed. Understanding who is obliged to do what, and when, explains more about short-horizon price behaviour than most analysis of what anybody thinks.

Three things worth keeping

Rebalancing is structurally contrarian and therefore stabilising — right up until everybody does it on the same day.
The cost of predictability is paid by the beneficiaries, invisibly, as a worse execution price.
Rules-driven flow is a category of its own. Ask what is obligatory this week before asking what anybody believes.

Next week: what happens when one of these participants cannot meet an obligation, and who ends up paying.

Joining the index

Sun, 28 Jun 2026 09:00:00 +0000

An index is a list of things and a rule for weighting them. That is all it is. It owns nothing and does nothing.

Yet an announcement that a company is joining one reliably moves its share price before a single index fund has bought anything — and the reason is a clean example of what happens when demand becomes obligatory and public at the same time.

Demand that has no choice

A fund tracking an index has one job: hold what the index holds. Its performance is judged not on returns but on tracking error — how closely it matches the list.

So when the list changes, the fund is not deciding whether the new constituent is attractively priced. That question is not in its remit. It must buy, in the weight specified, by the date the change takes effect.

That is unusual. Almost every other participant is price-sensitive: they buy if it is cheap enough and refuse if it is not. Index demand does not have that option. It is a buyer whose reservation price is infinity and whose deadline is published.

The announcement, not the event

The index provider announces the change some days or weeks before it takes effect, so that trackers can prepare.

The announcement is therefore public information that a large, price-insensitive, deadline-bound buyer will arrive on a known date. Anyone can act on it.

And they do. The price typically moves at the announcement rather than at the effective date, because that is when the information arrives — which is exactly what an efficient market should do with a predictable future flow.

The result is that the tracker, forced to trade at the effective date, buys after the move it caused. The cost lands on the fund’s investors, quietly, as a difference between the index’s paper return and the fund’s actual one.

Index inclusion is one of the few events in markets where everybody knows in advance who must trade, in what direction, in what size, and by when. It is price-insensitive demand with a published timetable.

What happens to the effect

Two forces have worn the classic inclusion effect down over the years, and both are worth understanding as mechanism rather than as history.

The obvious one is that a well-known, predictable flow attracts participants willing to supply the other side. If enough capital positions ahead of the tracker’s buying, the price impact at the effective date is absorbed rather than paid. Some of the cost simply moves from index investors to those participants.

The subtler one is that index providers responded to the problem. Longer notice periods, phased implementation, and rules that spread a change over several days all exist to reduce the concentration that made the flow exploitable.

Note what that means: the design of the index — the thing presented as a passive, neutral measurement — is now partly determined by how the market reacts to it being tracked. The measuring instrument has had to be redesigned around the behaviour of what it measures.

What it is not

A caution, because inclusion is often over-read.

Joining an index changes who owns a company. It does not change the company. There is no new information about the business in the announcement, and there is a respectable argument that any permanent price change from inclusion is a misallocation rather than a discovery.

There are second-order effects that are real — a broader shareholder base, more analyst coverage, easier access to capital — but they are slow, indirect and much smaller than the flow effect around the date.

The move at announcement is best understood as the market pricing a known future transaction, not as a re-evaluation of anything.

Three things worth keeping

Index demand is price-insensitive by mandate. A tracker that buys well and misses the index has failed at its actual job.
The information arrives at the announcement; the flow arrives at the effective date. They are separate events and they move prices for different reasons.
Predictable flow gets front-run, and index rules are now designed partly to make the flow less predictable.

Next week: the same mechanism running continuously, in every fund, forever — the rebalance.

Volatility is a price

Sun, 21 Jun 2026 09:00:00 +0000

“The market is pricing in twenty percent volatility.”

It is said as though the market has made a prediction. It has not. It has quoted a price, and a price is a different kind of object from a forecast.

Two different quantities

Keep these apart and most confusion disappears.

Realised volatility is a measurement. Take the actual price movements over some past window and compute how much they varied. It is a fact about history and there is nothing to argue about except the window.

Implied volatility is a price. Options trade at prices; feed a price into a pricing model and solve backwards for the volatility input that would produce it. The answer is implied volatility — which is not an observation of anything. It is the option’s price restated in more convenient units.

That restatement is genuinely useful. Option prices depend on strike, maturity, spot and rates, which makes two options hard to compare directly. Expressed as implied volatility they become comparable, in the same way that quoting bonds in yield lets you compare instruments with different coupons.

But it remains a price wearing the costume of a statistic, and people read the costume.

Why it is systematically higher

If implied volatility were a forecast, you would expect it to be too high about as often as too low. It is not. Across markets and across decades, implied volatility has on average exceeded the volatility that subsequently occurred.

A persistently biased forecast would be a puzzle. A persistently positive price is not, because of who is on each side.

Buyers of options are overwhelmingly buying protection. A fund with a large equity position, an airline hedging fuel, a business with a currency exposure — they want to cap a loss, and they are willing to pay more than the expected payout for the certainty. That is what insurance is, and nobody expects motor insurance to be priced at expected claims.

Sellers are supplying that protection. They must hold a position that loses money exactly when everything else is going wrong, and they will not do it at fair value. They require a premium.

Implied volatility is not the market’s estimate of future movement. It is the price of transferring the risk of that movement to somebody else — and like every insurance premium, it sits above the expected loss.

The gap between the two is the variance risk premium, and it is the compensation for underwriting.

The shape of the surface

The same logic explains why implied volatility differs across strikes, which under a naive reading of the model it should not.

In equity markets, options struck well below the current price trade at higher implied volatilities than those struck above. Two reasons, both about demand.

Large falls are more likely than a simple bell-curve model allows — markets gap down far more readily than they gap up — so the model’s price for a deep out-of-the-money put is too low, and the traded price corrects it.

And the demand is one-directional. Almost everyone hedging equities is hedging a long position, so almost everyone wants downside protection. That is a crowded side of the market, and crowded sides are expensive.

The skew, then, is not a statistical claim about the distribution. It is a map of where the hedging demand is.

What it is good for

None of this makes implied volatility uninformative. It makes it informative about something other than what it appears to describe.

It tells you what protection costs, right now, in this instrument, at this strike. It tells you where demand for insurance is concentrated. Changes in it tell you that the cost of transferring risk has moved, which is often the earliest visible sign that participants are becoming less willing to hold it.

What it does not tell you is how much the asset will move — and the systematic direction of the gap means that reading it as a forecast will mislead you the same way, over and over.

Three things worth keeping

Implied volatility is a price expressed in unusual units. Comparing it to realised volatility is comparing a premium to a loss.
The excess is compensation for underwriting, not a mistake to be corrected.
The skew maps hedging demand. It shows where people want protection, not where the distribution says danger is.

Next week: what happens to a share when it joins an index, and why the buying is guaranteed before it starts.

The dealer's hedge

Sun, 14 Jun 2026 09:00:00 +0000

Somebody sells you an option. What do they now do?

Not sit and hope. The answer to that question is a continuous trading programme in the underlying asset, and in aggregate it is large enough to change how the underlying behaves.

Hedging the direction away

An option’s value moves with the underlying, but not one for one. The sensitivity — delta — is the amount by which the option’s value changes for a small move in the underlying.

A dealer who has sold a call option is short that sensitivity. To neutralise it, they buy delta’s worth of the underlying. Now a small move in either direction leaves them roughly flat, and they are left holding the risks they actually meant to take a position in: volatility, and time.

That is the business. A dealer does not want a directional view. They want to earn a spread on volatility and hedge the direction away.

Why hedging once is not enough

Delta is not constant. It changes as the underlying moves — and the rate at which it changes is gamma.

So the hedge that was correct at 100 is wrong at 102. The dealer must trade again. And again. Every meaningful move requires an adjustment, and the size and direction of the adjustment depend on the sign of their gamma.

This is where it starts to matter to people who never trade options.

A dealer who is long gamma — typically because they bought options — finds that as the price rises, their delta rises, so they must sell the underlying to get back to flat. As it falls, they must buy. Their hedging leans against the move.

A dealer who is short gamma — having sold options — faces the reverse. As the price rises they must buy to stay hedged; as it falls they must sell. Their hedging pushes in the direction the market is already going.

A short-gamma book is a machine that buys strength and sells weakness, operated by someone with no opinion whatever. It amplifies moves not because anyone decided to, but because staying flat requires it.

The aggregate

For one dealer, this is housekeeping. The interesting question is what happens when the aggregate options market leans one way.

If dealers in aggregate are long gamma, their combined hedging is a stabilising force — a stream of selling into rallies and buying into declines, damping realised volatility.

If they are short gamma, the same mechanism runs in reverse, and hedging flows reinforce moves rather than absorbing them. The market can appear to have momentum that nobody intended.

Two cautions, because this idea gets over-applied.

First, dealer positioning is inferred, not observed. Estimates of it are models built on assumptions about who bought what, and they disagree.

Second — and more important — this is a description of a mechanism, not a forecasting tool. It explains why a move might be larger or smaller than the news justifies. It says nothing about which way the move goes. Anyone using gamma positioning as a direction signal has quietly swapped one claim for a much stronger one.

Where the hedging is impossible

The mechanism also explains why options books blow up in a particular way.

Continuous hedging assumes you can trade continuously — that the underlying moves in small increments and you can adjust as it goes. That assumption is exactly what fails in a gap: the market closes at 100 and opens at 88, and there was no opportunity to rehedge in between.

A short-gamma book takes its losses in precisely the conditions where the hedging programme cannot run. The strategy works until the market stops providing the continuity it depends on, which is another way of saying it works until it is needed.

Three things worth keeping

Delta hedging generates real flow in the underlying, produced by participants with no directional view at all.
The sign of gamma decides whether that flow damps or amplifies. Long gamma leans against the move; short gamma leans into it.
It is a mechanism, not a signal. It can tell you why a move was large. It cannot tell you which way.

Next week: volatility itself, and why the number quoted for it is a price rather than a forecast.

What a future actually is

Sun, 07 Jun 2026 09:00:00 +0000

A forward is the simplest derivative there is: I agree today to buy something from you on a date in the future, at a price we fix now.

Two people, one agreement, no venue required. And two problems that make it nearly useless at scale.

The two problems

The first is that a forward is not transferable. Our agreement is between us. If I want out three weeks later, I cannot simply sell it — I have to find someone willing to step into my side, and you have to agree to accept them. In practice I enter an offsetting agreement with somebody else and now hold two contracts, two counterparties, and twice the credit exposure.

The second is that it is only as good as my credit. If the price moves in my favour and you fail, the contract was worthless. You are exposed to me and I to you, for months, on an amount that grows as the price moves.

A futures contract is a forward with both problems engineered out, and everything peculiar about futures is a consequence of the engineering.

Standardisation

The first fix is to stop letting the two parties agree the terms.

The exchange specifies the contract completely: quantity, quality, delivery month, delivery location, settlement procedure. The only thing negotiated is the price.

That is what makes the contract fungible. Because every contract for a given month is identical, one is interchangeable with any other, and closing a position is simply a matter of doing the opposite trade. There is no need to find the original counterparty. There is no original counterparty.

The cost is precision. A producer whose actual exposure is a slightly different grade, in a different place, in the middle of a month rather than the end, cannot hedge exactly. They accept a residual mismatch — the basis — in exchange for being able to trade at all. Nearly every hedge in existence is approximate for this reason.

Daily settlement

The second fix is to stop letting credit exposure accumulate.

Each day the contract is marked to the settlement price and the change is paid in cash, immediately, from the losing side to the winning side. Variation margin.

The exposure therefore never runs beyond one day’s move, which is the entire point. Add initial margin — posted upfront, sized to cover a plausible one-day move — and the clearing house has collateral against the gap between marks.

This has a consequence people miss until it bites them: a futures hedge has cashflows before it has an outcome. A perfectly correct hedge held to expiry can require large cash payments along the way. If the thing being hedged does not generate matching cash on the same schedule — and physical inventory, or a long-dated liability, does not — a correct position can still fail for want of liquidity.

A future converts counterparty risk into liquidity risk. It is a very good trade, and it is a trade: you have not removed the danger, you have changed which kind you are exposed to and when it arrives.

Convergence and the roll

Two behaviours follow from the delivery mechanism.

Convergence. As expiry approaches, the futures price and the spot price must meet, because the contract is about to become the physical thing. Any gap is arbitrageable by someone willing to make or take delivery, and that possibility — even when rarely exercised — is what pulls the two together.

The roll. Most participants do not want delivery. They close the expiring contract and open the next one, and the price difference between them is not noise: it reflects the cost of carrying the asset over that period — funding, storage, insurance — less any yield it produces.

Someone holding a continuous position through repeated rolls is therefore paying or receiving that carry every time. Over a long horizon in a market with steep carry, the accumulated roll can dominate the return from the price itself.

Which is why a futures position and ownership of the underlying asset are not the same investment, however identical they look on the day you put them on.

Three things worth keeping

Standardisation buys fungibility with precision. Your hedge is approximate by design, and the residual is the basis.
Daily margin buys safety with cash timing. A correct hedge can still demand money before it pays any.
The roll is a real cost, not an accounting artefact. Held long enough, it can outweigh the price move you were trying to capture.

Next week: what happens when the people hedging derivatives have to trade the underlying to stay hedged.

Cash against collateral

Sun, 31 May 2026 09:00:00 +0000

Everything so far has assumed that a participant who wants to hold a position can pay for it. Time to remove that assumption, because the market that relaxes it is one of the largest in the world and is almost never discussed outside it.

The pawnshop

A repurchase agreement is a pawnshop transaction wearing a suit.

You own a bond and need cash. You sell the bond to someone today and agree to buy it back tomorrow at a slightly higher price. The difference between the two prices is, in substance, interest.

Legally this is two sales. Economically it is a secured loan: you have borrowed cash and pledged the bond. Getting the legal form right matters enormously — it is what lets the lender simply keep the bond if you fail, rather than joining a queue of creditors in an insolvency — but the economics are a pawnshop.

Two things make this useful. The lender is protected by holding an asset rather than a promise, so they will lend to counterparties they would never lend to unsecured. And the borrower gets funding at a rate closer to the quality of the collateral than to their own credit.

That combination is what allows an intermediary with a modest balance sheet to finance a large inventory — which is to say, it is what allows the market maker from issue eight to exist at all.

The haircut

The number that governs the whole arrangement is not the interest rate. It is the haircut.

Post a bond worth 100 and you might receive 98 in cash. The two-point difference is the haircut, and it is the lender’s protection against the collateral falling in value before they can sell it.

That two percent also defines the borrower’s leverage. If you must find two percent of every position from your own capital, you can hold fifty times your capital. Move the haircut to four percent and you can hold twenty-five times. The haircut is the leverage ratio of the financial system, set bilaterally, by thousands of risk managers, in a market with no central price for it.

Nobody publishes the haircut. It is agreed privately, adjusted quietly, and it determines how much leverage the system carries. Of all the numbers in finance, it may be the most consequential one that never appears on a screen.

When it moves

The problem is what happens when haircuts change, because they change together.

Volatility rises. Every lender independently and reasonably decides that two percent is no longer enough protection and moves to four. No one of them has done anything wrong.

But collectively they have just halved the borrowing capacity of every leveraged holder of that asset. Those holders must now either find more capital — difficult, in exactly these conditions — or sell. Their selling pushes the price down and raises measured volatility, which is the input that prompted the haircut change in the first place.

This is the same shape as the margin problem: a protection that is individually prudent and collectively pro-cyclical. It demands the most capital at the moment capital is scarcest, and it does so through thousands of uncoordinated decisions that no one participant can see the aggregate of.

General and special

One more distinction, because it connects the borrow to the funding market.

Most repo is general collateral: the lender wants cash secured against something safe and does not care which bond. The rate is a money-market rate, essentially the price of secured cash.

Sometimes, though, someone needs one particular bond — to deliver against a short, to settle a failed trade, to satisfy a specific obligation. Now the bond itself is scarce, and the rate inverts: the cash lender will accept a lower interest rate, sometimes near zero or below, for the privilege of getting hold of that specific security.

The bond has gone special, and the depth of the specialness measures how badly somebody needs it. It is the same signal as the equity borrow fee from last week, in a different market and a different unit.

Why it matters if you never touch it

You may never do a repo. It still sets the terms of the markets you do touch.

The bid-offer a dealer shows you includes their funding cost. The size they will quote depends on how much inventory they can finance. The speed with which a mispricing gets arbitraged away depends on whether an arbitrageur can borrow to put the trade on.

When funding markets tighten, all three degrade at once, in instruments that have nothing to do with each other, for reasons that never appear in their own order books.

Three things worth keeping

Repo is a secured loan wearing the legal clothes of a sale, and the legal clothes are the whole point.
The haircut sets systemic leverage and is negotiated privately, which means the system’s leverage is not observable in real time.
A special repo rate is a scarcity signal for one specific security — the same information as a hard-to-borrow fee, in a different market.

Next week: contracts that standardise a trade so completely that the underlying asset becomes almost incidental.

The borrow

Sun, 24 May 2026 09:00:00 +0000

To sell something short, you must first borrow it. That sentence contains almost everything interesting about short selling, and it is the part most often skipped.

The mechanics

You believe an asset is overvalued. You cannot sell what you do not have, so the sequence runs:

Find a holder willing to lend. Post collateral with them — typically more than the value of the shares, marked daily. Receive the shares, sell them in the market, and hold the cash. At some point, buy the shares back and return them. Your profit is the fall in price, less what the borrow cost you.

Three parties, then: you, the market you sold into, and a lender who is usually a long-term holder — a pension fund, an index tracker, a custodian pooling client positions — earning a fee on an asset they were going to hold anyway.

That lender is the constraint on the whole enterprise, and they have not committed to anything permanent.

The two prices

Borrowing has a cost, and that cost is the most informative number in short selling.

For an asset that is widely held and rarely shorted, borrowing is close to free — a few basis points a year. This is general collateral: nobody is competing for it and the lender is happy with any fee at all.

For an asset that many people want to short and few will lend, the fee rises, and it can rise a long way. This is a special, and an annualised borrow cost in the tens of percent is not unheard of.

That fee is a real, continuous drag on the position. A short that is right about direction but slow can still lose money, because the borrow was expensive and time was not free. It is also a price signal in its own right — arguably a cleaner one than the share price, since it measures demand to be short directly rather than inferring it.

Recall

Here is the structural weakness, and it is not a small one.

The loan is callable. A lender who decides to sell the underlying asset, or who simply wants it back, can recall it. When they do, you must return the shares — which means buying them in the market, whatever the price, whether or not your thesis has played out.

You are therefore running a position whose horizon is not yours to set. A long position can be held through a drawdown by anyone with the patience. A short cannot, because someone else can end it.

A long position can be wrong for as long as you can afford it. A short can only be wrong for as long as the lender allows, and the lender’s patience is shortest exactly when you need it most.

Why squeezes happen

Put the pieces together and the mechanism assembles itself.

Suppose a heavily shorted asset starts rising. Three things now happen at once, and they reinforce each other.

Shorts facing losses need to post more collateral, and some cannot. Rising prices make the borrow more valuable, so fees rise and marginal positions become uneconomic. And lenders — who are often the same long holders enjoying the rally — have more reason to sell, triggering recalls.

Every one of those forces produces the same action: shorts must buy. Their buying raises the price. The higher price intensifies all three forces.

A short squeeze is therefore not primarily a story about sentiment or narrative. It is a funding and availability event. The shorts are not buying because they changed their minds. They are buying because the mechanics of their position require it, and the requirement gets stronger as the price rises.

What short selling is for

It is worth stating plainly, because the practice attracts more heat than analysis.

A market where only optimists can act on their view has one direction of information flow. Negative information can only enter the price through existing holders selling, which is slower and weaker. The evidence broadly supports the mechanical prediction: constraining short selling tends to leave prices higher, and less informative, than they otherwise would be.

That is not an argument that any particular short position is well founded. It is an argument that the ability to take one is part of what makes a price mean anything — the point from issue four, that discovery requires someone able to act on what they know.

Three things worth keeping

The borrow fee is a price signal. It measures demand to be short directly, without the noise in the share price.
Recall risk means a short’s horizon belongs to the lender. That asymmetry, not sentiment, is the structural difference between long and short.
A squeeze is a funding event. The buying is compelled by collateral and availability, which is why it accelerates rather than exhausting itself.

Next week: the market that funds all of this, and the reason a haircut is the most important number nobody quotes.

Inventory

Sun, 17 May 2026 09:00:00 +0000

A market maker shifts their quote down. The obvious reading is that they think the price is going lower.

Usually they do not think anything of the sort. They have simply bought too much, and the quote is doing something other than expressing an opinion.

The position they never wanted

Start from what the job actually involves. An intermediary stands between a seller at two o’clock and a buyer at four, holding the risk in between. They are not paid to be right about direction. They are paid a spread for absorbing the mismatch in timing.

But the flow does not arrive evenly. Some afternoons everyone sells. The intermediary who quotes both sides honestly ends the day long a position they never wanted, financed with money that is not free, exposed to a market that owes them nothing.

Every additional unit is worse than the last: it uses more capital, consumes more risk limit, and increases the loss if the market moves against them. So they need a way to make buying less attractive and selling more attractive — without refusing to quote, because a market maker who stops quoting stops earning.

Skewing

The tool is the quote itself.

Suppose the fair mid is 100.00 and the natural market is 99.99 bid, 100.01 offered. Having accumulated a long position, the maker shows 99.98 bid, 100.00 offered instead.

The spread is unchanged. They are still quoting two-way, still competitive, still in business. But they have moved both sides down: their bid is now less appealing to anyone wanting to sell to them, and their offer is the best in the market for anyone wanting to buy. The flow they attract skews towards the direction that reduces the position.

This is the central insight of the inventory models that go back to the early 1980s: a dealer’s quotes are a function of their own book, and only secondarily a function of what they believe. The midpoint of a market maker’s quote is not their estimate of fair value. It is their estimate of fair value, adjusted by how much they want to stop holding what they are holding.

A skewed quote is not a forecast. It is an advertisement — offering a discount on the trade that helps the maker, and charging a premium for the one that does not.

Two forces, one quote

This gives you a cleaner way to read a moving quote, because there are only two things that move it, and they have different signatures.

Inventory pressure is mechanical and temporary. It reverses as soon as the position clears, and it is stronger in less liquid instruments where a position is harder to unwind. Its fingerprint: the price moves, the trade prints, and then it drifts back.

Adverse selection is informational and permanent. The maker widens or steps away because they think the flow they are seeing knows something. Its fingerprint: the price moves and stays moved, because the new level was correct.

Distinguishing these two in real time is the whole game, and it is worth noticing that the market maker is trying to do exactly the same thing from the other side. They also cannot tell whether they are accumulating a position because the flow is random or because it is informed — and by the time the answer is clear, they own the answer.

Why capacity is finite

The practical consequence is one that surprises people who think of liquidity as a fixed property of a market.

A market maker’s willingness to absorb your order depends on what they are already carrying. The same firm, in the same instrument, in the same conditions, will show you a different price at nine in the morning and at three in the afternoon, because their book is different.

Which means “how much liquidity is there” has no answer independent of what has already traded that day. Liquidity is not a reservoir that gets drawn down and refilled at a constant rate. It is the aggregate risk appetite of a set of balance sheets, and every trade uses some of it up.

Scale that observation to a whole market under stress — every intermediary long the same thing at once, every one of them skewing the same way — and you have the mechanism by which liquidity does not so much decline as vanish.

Three things worth keeping

A moving quote is usually about position, not opinion. Ask what the person quoting is likely to be carrying before reading it as a view.
Inventory effects reverse; information effects do not. Both look identical at the moment of the trade.
Liquidity is balance-sheet capacity. It depends on what has already happened today, which is why it is not a constant.

Next week: how you sell something you do not own, and what has to be true for that to work.

Markets without a book

Sun, 10 May 2026 09:00:00 +0000

Everything so far has assumed an order book: quotes displayed, priority rules, anyone able to see the market and join it.

Now discard that, because most of the world’s financial assets do not trade that way. Bonds, swaps, much of foreign exchange and nearly all credit trade by asking somebody for a price.

Why the book fails here

The order book works when the same instrument trades constantly. A share is one instrument; a company has one line of stock and everyone who wants exposure wants that.

Debt is not like that. A single issuer may have twenty bonds outstanding, differing by maturity, coupon and covenant, each a separate instrument. Multiply by every issuer and you have hundreds of thousands of instruments, most of which do not trade on a given day.

A continuous order book in an instrument that trades twice a month is not a market. It is an empty screen with a stale quote on it, and anyone posting a firm price into it is writing a very long-dated option to a market that has all the time in the world to work out whether it is mispriced.

So the mechanism changes shape. Instead of quotes waiting for orders, orders go looking for quotes.

Request for quote

The workflow is exactly what the name says. A buyer specifies what they want, sends the request to a handful of dealers, and receives prices back — firm, for a short window, in the size requested. They trade on one or none.

Notice what this changes.

The price is for that person, at that size, at that moment. It is not a public fact. Another client asking simultaneously might get a different number, and neither can see the other’s.

The dealer knows who is asking. In an anonymous order book you have no idea who is lifting your offer. In an RFQ you know the client, their history, and what they have been doing lately. That information is worth a great deal, and it is priced in.

And the client faces a genuine dilemma: asking more dealers gets more competition but tells more people what you are about to do. Ask five and you may get a better price. Ask twenty and everyone in the market knows there is a large buyer about, and the price you eventually pay may be worse than if you had asked three.

In an order book, the cost of trading is posted in advance and paid by whoever crosses it. In a dealer market, the cost is quoted to you personally, on the basis of who you are and what your asking has already revealed.

What the dealer is actually doing

A dealer quoting a bond is not matching you with another client. Usually there isn’t one. They are agreeing to hold the position themselves until they can unwind it, which may be days.

So the price they quote has to cover the spread, the funding cost of the inventory, the risk that the market moves while they hold it, and the possibility that they cannot find the other side at all. In an instrument that trades rarely, that last one dominates — and it is why the bid-offer on an illiquid bond can look enormous next to an equity spread. It is not the same quantity. It is compensation for warehousing something, not for a round trip measured in milliseconds.

What is lost, and what is gained

The trade-off is clean enough to state plainly.

Lost: transparency. There is no public book, so there is no reliable reference price. Pre-trade you have dealer runs and evaluated pricing services — estimates, not observations. Marking a portfolio is genuinely difficult, and the mark is somebody’s model until you actually sell.

Gained: liquidity that would not otherwise exist. A dealer will quote a size in an illiquid instrument that no order book would ever show, because they are being paid to take the risk and they know who they are taking it from. Remove the relationship and the anonymity of an order book gives them no way to price it.

Electronic RFQ platforms and post-trade reporting have pushed these markets towards more transparency over the last two decades, and the direction of travel is one way. But the underlying constraint has not changed: you cannot run a continuous public auction in an instrument that trades once a fortnight.

Three things worth keeping

Instrument count decides market structure. Few instruments trading often get an order book; many instruments trading rarely get dealers.
In a dealer market, asking is informative. Every additional dealer you query buys competition and spends secrecy.
A wide bid-offer in an illiquid instrument is not a rip-off by default. It is the price of somebody holding a thing nobody else wants today.

Next week: what a market maker’s quote is actually telling you, and why it is usually about their own book rather than the market.

What a broker actually does

Sun, 03 May 2026 09:00:00 +0000

You press buy. Something happens. A confirmation comes back.

What happened in between is, for most people, a black box — and it contains more decisions, made by more parties, than almost anyone assumes.

Agent or principal

The first fork is the one that determines everything after it.

A broker acting as agent takes your order and represents it in the market. They do not take the other side. Their interest is aligned with yours in a simple way: they are paid a commission whether the trade goes well or badly, so they have no position to protect.

A broker acting as principal takes the other side themselves. You buy from them, out of their own inventory. They are not charging a commission; they are earning a spread. Now your interests are opposed in the direct sense that a better price for you is a worse one for them.

Neither is illegitimate, and principal trading is often better for the client — it can offer certainty of execution at a size the open market would not absorb quietly. But you should always know which one you are getting, because they are different products with the same name.

Internalisation

The version most retail orders meet is a particular flavour of principal trading.

Rather than send your order to an exchange, a broker may match it internally — against another client’s order, or against their own book, or by passing it to a wholesaler who does the same. Typically you get a price at or slightly better than the public quote.

That price improvement is real, and it is the argument for the practice. It is also worth being precise about where the money comes from.

Recall the taxonomy from issue two. A wholesaler will pay for the right to interact with retail flow because retail flow is, on average, uninformed — it is not systematically arriving just before the price moves against them. That predictability is worth money. Some of it is returned to the client as price improvement, some is retained, and some is paid to the broker for the flow.

Order flow is not routed to where it trades best. It is routed to where it is worth most — and the two coincide only to the extent that the rules make them.

What is left on the public market

There is a structural consequence, and it is the live argument about the practice.

If the uninformed flow is matched away internally, what reaches the public order book is disproportionately the flow that nobody wanted to internalise. Which is the flow more likely to be informed. Which makes quoting on the public market riskier, which widens the spread there — and that spread is the reference price used to decide whether the internalised trade was any good.

You can see the loop. The public quote is the benchmark, and the practice being benchmarked against it removes the flow that makes it cheap to maintain.

Nobody has a clean answer to this. Both halves are true: individual retail executions are demonstrably better than the public quote, and the public quote is worse than it would be if those orders arrived there.

What best execution actually obliges

The rules that police all this are less absolute than the name suggests.

Best execution is not “the best price”. It is an obligation to take sufficient steps to obtain the best result taking account of price, cost, speed, likelihood of execution and settlement, size, and any other relevant consideration. A broker may quite properly choose a worse price for a higher certainty of filling a large order.

That flexibility is necessary — a rigid price rule would be trivially gameable — and it is also what makes the obligation so hard to enforce. Almost any routing decision can be justified under some weighting of those factors, and the client is rarely in a position to check.

We will come back to this when we look at measuring execution properly. For now the useful question is not “did I get the best price” but “who chose, and what were they optimising”.

Three things worth keeping

Establish agent or principal. It is the difference between someone representing you and someone trading against you, and both are called broking.
Price improvement is funded by predictability. Uninformed flow is genuinely valuable, and some of that value comes back to you.
Best execution is multi-factor by design. That makes it flexible enough to be useful and vague enough to be hard to enforce.

Next week: markets with no order book at all, where every price is quoted to one person on request.

Maker, taker, rebate

Sun, 26 Apr 2026 09:00:00 +0000

A trade has two sides, and most exchanges charge them differently. Some go further and pay one side to show up at all.

This looks like a pricing quirk. It is a solution to a genuine problem, and its side effects have shaped where orders go for two decades.

The empty room

A new venue has a problem that has nothing to do with technology. Nobody wants to post a resting order on a market with no participants, because it will not get filled. And nobody wants to send an order to take liquidity on a market with no resting orders, because there is nothing there to take.

Each side is waiting for the other. The room stays empty.

The standard solution is to pay one side to arrive first. Since resting orders are the thing that must exist before anything else can happen, the venue pays the participant who makes liquidity and charges the one who takes it. If the rebate paid out is a little less than the fee charged, the venue keeps the difference and the room fills up.

What a rebate really is

It is worth being clear about the economics, because “the exchange pays them” is usually said in a tone of accusation.

A resting order is a standing commitment to trade at a price, available to anybody who wants it, for as long as it sits there — and it will be taken up precisely when doing so suits somebody else. That is a risk, and it is borne for free. A maker rebate is partial payment for bearing it. It does not remove the risk; it offsets a fraction of it. We will put a proper price on that risk in a later issue.

So maker-taker does not conjure liquidity out of nothing. It transfers money from participants who value immediacy to participants who supply it, with the venue taking a cut for arranging the transfer. Whether that is a good bargain depends entirely on which of the two you are.

A rebate is not a subsidy for existing. It is compensation for writing an option the market gets for free — collected from the person who exercises it.

The distortion

Here is where it stops being tidy.

A broker routing a client’s order faces a decision that now has two components: where the order is most likely to get the best outcome, and where the routing economics are best for the broker. Those two are not always the same venue.

If a broker is paid to post and charged to take, they have a standing incentive to post — even when crossing the spread immediately might have served the client better. And if venues compete on rebate levels, order flow can end up sorted by fee schedule rather than by execution quality.

This is a real and well-documented conflict, and it is the reason best execution rules exist in the form they do. It is also why some venues run the schedule backwards — taker-maker, or inverted pricing, paying the aggressor and charging the resting order — precisely to attract the flow that maker-taker pushes away.

The existence of both models at once tells you the honest answer: neither is right, and each is a bid for a different kind of order.

Why the quote is not the price

The subtler consequence is that fees drive a wedge between the price you see and the price you get.

A quote of 100.00 is not 100.00 to everyone. To a taker paying a fee it is slightly worse; to a maker collecting a rebate it is slightly better. Two venues showing the identical price can offer materially different economics, and the better displayed price is not always the better net price.

This matters more as spreads narrow. When the tick was wide, fees were noise against the spread. When the spread is a single minimum increment, a fee of a fraction of that increment is a large proportion of the total cost of the trade — and it is invisible in every chart of where the market traded.

Three things worth keeping

Rebates buy liquidity by redistributing it, from participants who want immediacy to participants who supply it. Nothing is created.
Ask who pays the routing decision’s costs. Where the fee schedule and the client’s interest point at different venues, the rules are the only thing holding the line.
The displayed price is not the net price. Compare quotes after fees or you are comparing two different things.

Next week: the firm standing between you and the market, and what it is actually doing with your order.

Where a price comes from

Sun, 19 Apr 2026 09:00:00 +0000

The last three issues built the machine. This one asks what comes out of it.

A price looks like a measurement — as though somewhere there is a true value and the market is an instrument reading it, with some error. That picture is comfortable and it is wrong, and most confused arguments about markets can be traced back to it.

The marginal trade

Start with what a printed price literally is: the terms on which the last trade happened.

Not the average opinion. Not the consensus. The point at which one buyer and one seller stopped disagreeing, at the size they were doing, at that instant.

Everyone else — the thousands who looked at it and did nothing — contributed nothing directly to that number. They are in it only through their absence: they did not think it was mispriced enough to act.

This is why a price can move a long way on very little volume. The number attaches to the marginal trade, and the marginal trade might be tiny. It is also why the value of a large holding is not its size times the last price. The last price was the terms for the last unit, and the terms for the next thousand are a different question entirely.

What gets aggregated

The useful thing about the mechanism is what it does with dispersed private information.

Suppose a hundred people each know one small, true thing about a company that nobody else knows. No one of them can act on much. But each buys or sells a little, and each of those actions moves the price a little in the direction of what they know. The price ends up reflecting an aggregate that no participant possesses.

That is a genuinely remarkable property and it is the strongest argument for taking prices seriously. It is also easy to overstate, so it is worth being precise about the conditions it needs: that people with information can act on it, that their acting moves the price, and that they are not all wrong in the same direction at once.

Where any of those fails, the aggregation fails with it — and the price is still produced, looking exactly as authoritative as before.

A price does not know how much information is behind it. A number produced by one distracted trade and a number produced by a thousand well-informed ones print identically on the tape.

Two ways to move a price

It helps to separate the two forces, because they look the same in a chart and are not remotely the same thing.

Information changes what participants believe an asset is worth. The price moves and stays moved, because the new level reflects a new belief.

Flow changes who needs to trade, without changing anyone’s beliefs at all. A large seller has to sell; the price has to fall far enough to find willing buyers; and once they are done, there is no reason for it to stay there. That move decays.

Distinguishing them in real time is close to the central problem of trading, and nobody does it reliably. But the distinction explains why “the market fell on heavy selling” is an empty sentence. Every fall involves selling. The question is whether the selling was telling you something.

Discovery is not free

The last piece, and the one that connects back to the machine: price discovery is work, and somebody pays for it.

Finding out that something is mispriced takes research, and acting on it means putting capital at risk against a market that will charge you a spread for the privilege. The information only gets into the price because someone expected to be compensated for putting it there.

Which sets up an awkward loop. If prices already reflected everything, nobody would be paid to do the work, and they would stop — at which point prices would stop reflecting everything. Markets cannot be perfectly efficient, because perfect efficiency would remove the incentive that produces the efficiency.

The practical version: prices are approximately right in proportion to how much it is worth someone’s while to check.

Three things worth keeping

A price is the terms of the marginal trade, not a valuation of the whole. Multiplying a holding by the last print is an estimate, and a generous one.
Information moves prices permanently; flow moves them temporarily. Almost everything difficult about trading lives in telling which you are seeing.
Accuracy is paid for. Where nobody is compensated for checking, expect the price to be worse — and to look exactly as confident as any other price.

Next week: how venues pay people to show up, and what that money buys.

The smallest number

Sun, 12 Apr 2026 09:00:00 +0000

There is a number in every market that almost nobody outside it ever thinks about, and it quietly determines how the whole thing behaves: the smallest amount by which a price is allowed to change.

The tick. It sounds like an accounting convention. It is closer to a constitutional provision.

What a tick actually does

A tick sets the minimum distance between the bid and the offer. If the increment is one penny, then the tightest possible market is one penny wide, and no amount of competition can narrow it further.

That immediately makes the tick a floor under the cost of trading — which sounds straightforwardly bad, and is not.

Consider what happens when the tick is very small. Competing to be at the front of the queue no longer requires patience or capital; it requires only that you improve the price by a trivially small amount. Anyone can step in front of a large resting order for a fraction of a penny, take the trade, and leave the patient participant unfilled.

The rational response is not to post large resting orders. So depth thins, quotes flicker, and the displayed market becomes a very thin film over not much.

The other extreme

Now make the tick very large. The spread cannot narrow below it, so intermediaries earn a comfortable margin on every round trip. Because that margin is protected, it is worth queueing for, and queues at each price level become long and deep.

You get a market that is expensive to cross and reassuringly solid once you do — good for someone with size and time, a straightforward tax on everyone else.

So the tick is a dial between two costs:

	Tick too small	Tick too large
Spread	Very tight	Floored, wide
Depth at touch	Thin	Deep
Queue	Barely matters	Valuable, long
Favours	Small, fast orders	Large, patient orders

Neither end is correct, and the right setting differs by instrument. A liquid, low-priced share and a thinly traded high-priced one do not want the same increment, which is why modern tick regimes scale the increment with price and with how actively the instrument trades.

A tick size is a decision about who a market is for. Make it smaller and you subsidise speed. Make it larger and you subsidise patience. There is no neutral setting, only a choice that someone has to make on purpose.

The evidence, such as it is

This has been tested in the field more than once, because regulators on both sides of the Atlantic have deliberately moved tick sizes and watched what happened.

The findings have been consistent in direction if not in magnitude: widening the increment on names that were previously trading at the minimum increases quoted spreads and increases displayed depth. Trading costs for small orders go up. Trading costs for large orders can go down, because there is something there to trade against.

Which is another way of saying the dial does exactly what the mechanics predict, and that the argument about where to set it is not empirical. It is an argument about whose costs matter more.

Where the tick escapes

Two things complicate the picture, and both are worth knowing about.

The first is that a tick constrains displayed prices, but venues have long offered ways to trade between them — midpoint matching, price improvement, mechanisms that fill you at half a tick better than the quote. Every one of these is, in effect, a partial exemption from the increment, granted to particular kinds of flow.

The second is that the tick applies to price, not to size. Where competing on price is forbidden, competition moves to whatever is left: speed of arrival, willingness to show size, and the fees and rebates a venue offers for turning up.

Which is next week’s subject.

Three things worth keeping

The tick is a floor on the spread and therefore a floor on trading costs for anyone whose order is small enough to cross it in one go.
Small ticks buy tightness by spending depth. You cannot have both from the same dial; you can only choose which one you would rather have.
Constrain competition on price and it reappears somewhere else — in speed, in size, in fees. It never simply goes away.

Next week: where a price comes from, and what it is actually made of.

Who is on the other side

Sun, 05 Apr 2026 09:00:00 +0000

Every trade you have ever done had someone on the other end who thought they were getting the better of it. Both of you cannot be right about the price. But you can both be right about the trade — and understanding how is the beginning of understanding why markets work at all.

The naive version

The intuition most people start with is that markets are a contest of opinion: I think it goes up, you think it goes down, one of us collects.

That version cannot be the whole story, because a market made only of opinion would barely function. If the only reason to trade is disagreement about the future, then every trade is a bet that the other person is stupid, and the rational response to being offered a trade is to wonder what they know that you do not. Markets built purely on that logic tend to freeze.

Real markets work because most participants are not there for the same reason, and several of them are not really expressing an opinion at all.

Four reasons to be there

The hedger has an exposure they did not choose and want less of. A miner with copper coming out of the ground, an airline with fuel to buy, a fund manager with a currency they hold only because the shares happen to be listed there. They are not predicting anything. They are converting a risk they cannot manage into one they can.

The investor is moving money across time. They have cash now and want an asset later, or the reverse. Their horizon is measured in years and the price they get today is a detail, not the point.

The speculator is taking a view, and gets called names for it. But a speculator’s function is precise: they are the person willing to hold a risk that a hedger wants rid of, because they think they are being paid enough for it. Without them, the hedger has nobody to sell to.

The intermediary has no view at all and does not want the position. They are buying and selling almost simultaneously, being paid a spread for standing between two people who arrived at different times.

That last category is the one worth pausing on, because it dissolves the coincidence-of-wants problem from last week. A hedger who needs to sell at two o’clock and an investor who wants to buy at four o’clock never have to meet. The intermediary holds the risk for two hours and charges for it.

A market does not need its participants to disagree. It needs them to want different things — different horizons, different exposures, different tolerances for the same risk. Disagreement is one source of trade. It is not the main one.

The one you have to watch for

There is a fifth participant, and every price in every market is set partly in fear of them: the person who knows something you do not.

They look exactly like everybody else. Their order is the same shape as a hedger’s. The difference only becomes apparent afterwards, in the direction the price moves once they are done.

This is why the composition of flow matters so much to anyone quoting. Trading against a pension fund rebalancing to a policy weight is a pleasant business. Trading against someone who has correctly worked out what is about to be announced is not, and no amount of skill lets you tell them apart in advance.

The whole apparatus of spreads, quote sizes and cancellation exists to manage the fact that you cannot see who is coming.

Why it matters what the mix is

Because the mix is not constant, and it changes exactly when it matters most.

In quiet conditions, flow is dominated by the uninformed and the mechanical: rebalancing, hedging, savings, index tracking. Those are the counterparties an intermediary wants, and their presence is what makes tight quotes affordable.

In stressed conditions, the mechanical flow becomes one-directional — everyone needs to reduce risk at once — and the proportion of trades that carry real information rises. The intermediary’s business gets worse in exactly the moment everyone wants more of it.

The market has not become less efficient. Its population has changed.

Three things worth keeping

Ask what would make the other side happy. If you cannot construct a plausible reason for them to want this trade, you may be the reason.
Speculators and intermediaries are functions, not characters. Remove either and the hedger simply has nobody to trade with.
The mix of participants is a market condition, as real as volatility, and it moves in the same direction as trouble.

Next week: the smallest number in the market, and how much it decides.

What a market is for

Sun, 29 Mar 2026 09:00:00 +0000

Fifty issues of mechanisms, and it is worth going back to the beginning — because the beginning is where the assumptions get smuggled in, and we have been building on them for a year.

A market is not a place where prices are announced. Nobody at an exchange decides what anything is worth. A market is a machine for discovering what two people who have never met, and who disagree, will nonetheless both sign.

Nearly everything else about market structure is a consequence of that.

The problem being solved

Imagine you want to sell something and I want to buy it. Left to ourselves, we face three problems, and they are worse than they look.

First, finding each other. In a world without a venue, I have to locate a person who wants exactly what I have, at roughly the moment I want to be rid of it. This is the coincidence-of-wants problem, and it is why unstructured markets are so slow.

Second, agreeing a price without either of us knowing what anyone else is paying. Any number I propose is either too generous or insulting, and I have no way to tell which.

Third, trusting the settlement. You hand over the asset before I hand over the money, or the other way round, and whoever goes first is exposed to the other one changing their mind.

An exchange is a device that solves all three at once: it concentrates the counterparties, it publishes the prices, and it stands between the two sides so neither has to trust the other.

The double auction

The mechanism at the centre is older than any of the technology around it. Buyers say the most they will pay. Sellers say the least they will accept. Both sets of intentions are collected in one place and sorted.

If the best bid is below the best offer, nothing trades — and the gap between them is the spread. When someone becomes willing to cross that gap, a trade happens, and the price of that trade becomes a fact everybody can see.

This is a double auction because both sides bid. In an ordinary auction, one seller faces many buyers and the price only ratchets one way. In a double auction, everyone can be on either side, and the price moves in both directions continuously. That single difference is what turns an auction into a market.

An exchange does not set prices. It sets the rules under which prices are allowed to be discovered — who can quote, in what increments, in what order they are matched. Everything you can observe about a market is downstream of those rules.

Why the price is information

Here is the part that gets underrated. The output of this machine is not just trades. It is a number that summarises what a very large number of people know and believe, and that nobody involved could have produced alone.

Every participant brings a fragment: an opinion, a cashflow need, a hedge that has to be put on by Friday, a model that says something is cheap. The matching process aggregates those fragments into one figure. No individual knows why the price is where it is, because no individual holds more than a fraction of the reasons.

That is why a price is worth taking seriously even when you disagree with it — and why the interesting question about any market is never “is this price right” but “what kind of information can this mechanism actually aggregate, and what does it systematically miss”.

What a venue is really selling

If you asked an exchange what its product is, the honest answer would not be “trading”. It would be certainty about the rules.

A participant needs to know that their order will be treated the same way as everyone else’s, that the matching logic will not change halfway through the afternoon, that a trade printed at 10:04 actually happened, and that the person on the other side will deliver. None of that is glamorous. All of it is what makes putting capital at risk in public tolerable.

This is worth holding onto, because it reframes the arguments that follow in later issues. Disputes about tick sizes, order types, data fees, speed and priority are not technical trivia. They are arguments about the rules of the machine, and the rules of the machine determine what the price ends up meaning.

Three things worth keeping

Prices are produced, not published. A market is a process, and the number it emits carries the fingerprints of the process that made it.
A double auction is the whole trick. Both sides quoting, continuously, in public, is what separates a market from an auction and from a negotiation.
An exchange sells rules, not trades. Ask what any market-structure rule rewards, and you will usually understand the resulting market better than by looking at its volumes.

Next week: the people on the other side of the trade, and why a market only works if they are not all like you.

Deleveraging

Sun, 22 Mar 2026 09:00:00 +0000

Fifty issues of mechanisms, and this one is where they meet.

A deleveraging spiral is what happens when a fall in prices forces selling, and the selling causes a further fall. Nothing exotic is required. Every component has appeared already in this letter, behaving exactly as intended.

The loop

Prices fall for some ordinary reason.

Margin calls arrive. Positions are marked lower, and collateral must be posted against the difference. Cash is needed today.

Margin requirements rise. Volatility has increased, so the models governing initial margin demand more. Not because anyone changed policy — because the input moved.

Haircuts widen. Lenders in secured funding markets reasonably require more protection, which reduces the borrowing capacity of every leveraged holder of the affected asset.

Risk limits bind. Value at risk has risen with volatility, so positions that were within limits are now outside them, and must be cut.

Redemptions come in. Holders of funds see losses and ask for their money.

Every one of those is individually correct. Each is a prudent institution responding sensibly to a genuine deterioration. And every one of them produces the same action: sell.

That selling pushes prices lower, which raises volatility, which tightens margin, widens haircuts, breaches limits and prompts more redemptions. The loop closes.

Nothing in the spiral requires anyone to behave badly. It requires only that many people behave prudently, using similar rules, at the same time. The prudence is what makes it fast.

Why it spreads sideways

The most confusing feature is that unrelated assets fall together, and the mechanism is one this letter has already described twice.

A participant needing cash sells what they can, not what they should. The impaired asset may be unsellable at any sensible price, so they liquidate the liquid, healthy, unrelated things instead.

That transmits stress into markets with no connection to the original problem, by holders with no view on them, purely because the same balance sheet held both.

It is the same mechanism as collateral-driven selling from issue fifteen, and it is why correlation goes to one in a crisis. The link is not economic. It is that the same people own both and are being forced.

Why it stops

Spirals end, and understanding how matters more than understanding how they start.

The forced sellers finish. There is a finite quantity of leveraged holding. Once it has been liquidated, the mechanical selling stops, whatever the price.

Price attracts unlevered buyers. Participants with cash and no obligation to anyone are the natural counterparty. They are typically slow, because their advantage is precisely that they are not compelled to act.

Someone breaks the coordination. A backstop that removes the fear of being the last holder, as issue four described. This works by changing expectations rather than by buying much.

Note that the first two are self-limiting and painful, and the third is the only one that acts quickly. Which is why the policy response to these episodes has converged on that form, and why the debate about moral hazard is so difficult: the intervention that works is the one that removes the consequence of the leverage that caused it.

The uncomfortable conclusion

Every safeguard in this letter is pro-cyclical.

Margin protects the clearing house and demands cash when cash is scarce. Haircuts protect the lender and cut borrowing capacity when borrowing is hardest. Risk limits protect the institution and force selling when selling is most damaging. Ratings protect the investor and trigger mandatory sales at the worst moment.

Each is correct in isolation. Each was designed by people thinking carefully about one institution’s safety. Together they form a machine that amplifies exactly what it was built to contain.

There is no version of these protections that is individually prudent and collectively neutral. The best that can be done is to make them less sensitive, slower to tighten, and less synchronised — which is the whole of macroprudential policy, and it is a matter of degree rather than solution.

Three things worth keeping

Nobody has to behave badly. Prudence, applied simultaneously with common rules, is sufficient.
Forced sellers sell what is liquid, which is why stress appears in markets that had nothing to do with it.
Every safeguard is pro-cyclical. The trade-off is real and permanent, and can only be managed rather than removed.

Next week: back to the beginning. After a year of mechanisms, what a market is actually for.

Stress testing

Sun, 15 Mar 2026 09:00:00 +0000

The problem with value at risk was that it learns from the past, and the past does not contain the crisis you are about to have.

Stress testing abandons the historical distribution entirely. Instead of asking what usually happens, it asks what would happen if a specified thing occurred.

The method

Define a scenario. Equities fall thirty percent, credit spreads triple, the currency devalues, funding markets close.

Revalue everything under those conditions. Report the loss.

That is the whole approach, and its strength is that it makes no claim about probability. It does not say how likely the scenario is. It says what would happen — which is a question with an answer, unlike the question of what might happen, which does not.

It also handles non-linearity properly. A portfolio containing options behaves very differently in a thirty percent fall than a linear extrapolation from small moves suggests, as issue eleven’s discussion of convexity implies. A stress test prices the actual instruments at the actual level; a volatility-based measure cannot.

Where the scenarios come from

Three sources, with different weaknesses.

Historical replay. Apply a real past episode to today’s portfolio. Concrete, defensible, and easy to explain — and it tests the crisis you already survived, against a portfolio built by people who knew about it.

Hypothetical. Construct a plausible bad state. More flexible, and it depends entirely on the imagination of whoever wrote it. Scenarios tend to be variations on recent fears.

Reverse stress testing. The most useful and least common. Rather than asking what a scenario does to you, ask what scenario would break you — and then judge whether it is plausible.

This last one inverts the exercise usefully. Instead of testing a list somebody chose, it identifies the exposures that actually matter, including combinations nobody would have thought to write down.

A stress test tells you what a scenario does. Reverse stress testing tells you which scenarios matter. The second is a much better use of the same machinery, and it is the one that produces uncomfortable answers.

The failure modes

Three, and they are all about the choice of scenario rather than the arithmetic.

Comfortable scenarios. If the exercise is run by people whose results will be judged, there is pressure towards scenarios the portfolio survives. This is not usually deliberate. It is the ordinary tendency to test what you are prepared for.

Missing the correlation. A scenario that moves one variable while holding others fixed misses the point of a crisis, which is that everything moves together. The most dangerous scenarios are combinations, and combinations multiply faster than anyone can enumerate.

Ignoring second-round effects. A test typically revalues a portfolio at stressed prices. It rarely asks what happens when everyone else is revaluing too, and all of them start selling. The prices in the scenario are assumed, and in reality they would be produced partly by the reaction to the scenario.

That third one is the hardest and the least addressed. A test assuming you can sell at the stressed price is assuming a buyer exists — and in the scenario being tested, the buyer is also being tested.

Where it works

Supervisory stress testing of banks has been genuinely valuable, for a reason that is not about the numbers at all.

A common scenario applied across institutions produces comparable results. That comparability, and the requirement to publish, forces disclosure of exposures that would otherwise be private, and it makes capital adequacy arguable in public rather than settled privately.

The number itself may be wrong. The exercise of producing it — building the systems to revalue a whole balance sheet under specified conditions, and being answerable for the result — is worth more than the output.

Three things worth keeping

It makes no probability claim, which is its main advantage over anything estimated from history.
Reverse stress testing is the better question. What breaks you, rather than what a chosen scenario does.
Second-round effects are usually missing. The scenario assumes a buyer who is also in the scenario.

Next week: what happens when everyone’s model tells them to sell on the same morning.

Three kinds of leverage

Sun, 08 Mar 2026 09:00:00 +0000

Leverage is usually described as borrowing to invest, and measured as debt divided by equity.

That definition captures one of three mechanisms that produce the same behaviour, and it misses the two that are hardest to observe.

Balance sheet leverage

The familiar kind. Borrow money, buy assets, and the position exceeds the capital committed.

Gains and losses are both multiplied. Ten to one leverage turns a ten percent fall into a wipeout, which is the whole of the arithmetic and needs no elaboration.

Its virtue is visibility. It appears on a balance sheet, in a debt figure, in a ratio anyone can compute. Which means it is the kind that regulators constrain most effectively, and the kind least likely to surprise anybody.

Embedded leverage

The second kind requires no borrowing at all.

An option costing a small fraction of the underlying gives exposure to the whole of it. A futures position requires margin, not the full notional. A structured note may reference several times its face value.

In each case the instrument itself contains the multiplication. There is no borrowing, so there is no debt, so the standard measure records nothing.

An investor holding options with no borrowings whatever can have far more exposure than one who borrowed to buy shares — and only the second appears in any leverage statistic.

This is why the total return swap from issue thirty-eight is so effective at concentrating risk invisibly: the leverage is inside the contract, and the contract is not debt.

Structural leverage

The third is the least visible and the most important, and it comes from the liability side rather than the asset side.

An institution that must maintain a position — because a mandate requires it, a liability must be matched, a hedge must be held, or a rule demands a ratio — is leveraged in the sense that matters, which is that it cannot choose when to sell.

Consider two holders of an identical portfolio, neither having borrowed. One can hold for as long as they like. The other faces redemptions, or a covenant, or a regulatory ratio that binds when prices fall.

The second will be selling in a downturn. The first will not. Their leverage ratios are identical and their positions are not remotely the same.

The question is never how much you borrowed. It is whether you can be forced to sell before you want to — and borrowing is only one of the things that can force you.

Why they behave alike

All three produce the same failure mode: a fall in prices compels selling, and the selling causes a further fall.

Balance sheet leverage does it through margin calls. Embedded leverage does it through collateral demands on derivative positions. Structural leverage does it through redemptions, mandates and ratios.

The trigger differs. The behaviour is identical, and it is the behaviour that matters — which is why measuring only the first kind gives such a misleading picture of how much forced selling a market might be capable of producing.

What to look for

Three questions get further than any leverage ratio.

What would force a sale? Enumerate the actual triggers: margin, redemption, covenant, ratio, mandate. If none exists, the position is genuinely unlevered however it was funded.

How fast can the trigger fire? Daily margin acts within a day. A quarterly covenant test acts within a quarter. Speed determines whether an unwind is a disruption or a cascade.

Who else is exposed to the same trigger? One participant forced to sell is a loss. Many participants forced to sell the same asset on the same trigger is next week’s subject.

Three things worth keeping

Debt-to-equity captures one of three mechanisms, and the invisible two behave identically.
Embedded leverage lives inside instruments and appears in no debt figure.
Structural leverage is about who controls the timing of your exit, which is the only thing that ever really mattered.

Next week: asking what would happen, rather than measuring what has.

Correlation

Sun, 01 Mar 2026 09:00:00 +0000

Every claim about diversification rests on one input: how much the things you hold move together.

That input is estimated from the past, treated as a property of the assets, and used to justify position sizes. It is none of those things reliably.

The measurement

Correlation summarises the tendency of two series to move together, on a scale from minus one to one. It is the mechanism by which a portfolio is supposed to be less risky than its components: combine assets that do not move together, and the combination varies less than the pieces.

The arithmetic is sound. The difficulty is entirely in the input.

A correlation is computed over a window. Choose a different window and you get a different number — not slightly different, frequently a different sign. Two honest analysts using the same data and different lookbacks reach opposite conclusions about whether two assets diversify each other.

That alone should make anyone cautious about a number quoted to two decimal places.

Why it moves the wrong way

The deeper problem is not instability. It is that the instability has a direction.

In ordinary conditions, assets are driven by their own particulars. A company’s results, a sector’s demand, a country’s policy. Those are genuinely different things, and the correlations between them are genuinely low.

In stress, one factor dominates everything. Whether that factor is liquidity, leverage, risk appetite or a common macroeconomic shock, it acts on everything at once — and the idiosyncratic differences that produced low correlation become irrelevant next to the common force.

So correlations rise towards one in a crisis. Which means diversification, measured in calm conditions, overstates the protection available in the conditions where protection is needed.

Correlation is not a property of two assets. It is a property of the environment they are in, and the environment changes fastest at the moment your position size was justified by the old number.

This is not a subtlety. It is the mechanism that turned pooled mortgages into a systemic event, and it is the single most important reason that risk models underestimate losses in crises.

Why it happens

Three reasons, and they compound.

Common factors dominate. Sector, geography and idiosyncrasy matter in normal times; in stress, exposure to the single thing going wrong dominates all of them.

Forced selling ignores fundamentals. A leveraged holder meeting a margin call sells what can be sold, not what should be sold. That mechanically links the price of unrelated assets held by the same stressed participants — the point this letter made about collateral, and the one it closes on.

Crowding. If many participants hold similar positions because similar models recommended them, those positions are linked through their holders regardless of what the underlying assets have in common.

That last one is the least measurable and probably the most important. Two assets with nothing in common can be tightly correlated simply because the same people own both and will sell both at the same time.

What to do

Three habits, none of them a solution.

Estimate over long windows including bad periods. A correlation computed over a calm five years describes a calm five years.

Test the assumption directly. Rather than asking what correlation has been, ask what the portfolio does if it goes to one. If that answer is intolerable, the diversification was doing more work than it can bear.

Ask who else holds this. Crowding is not visible in price history at all, and it is the channel through which apparently unrelated things become related.

Three things worth keeping

Correlation depends on the window and can change sign with the choice of lookback.
It rises in stress, so measured diversification is largest precisely when it is least reliable.
Shared holders link unrelated assets. Ask who owns both, not just what they have in common.

Next week: the thing that turns a correlated market into a forced one.

Value at risk

Sun, 22 Feb 2026 09:00:00 +0000

A risk report says the portfolio’s one-day value at risk is four million at ninety nine percent.

That is a precise statement. It is also considerably narrower than the use it is put to, and the gap between the two has been expensive more than once.

What it says

The claim is: on ninety nine days out of a hundred, the loss will be less than four million.

That is a statement about a quantile of a distribution. Nothing more.

In particular it says nothing about the other one percent. It does not claim the worst case is four million. It does not say what happens on the hundredth day. A portfolio losing five million on the bad day and one losing four hundred million can have identical value at risk.

That is not a flaw in the calculation. It is what a quantile is. The failure is in reading it as a worst case, which happens constantly because a single number attached to the word “risk” invites exactly that reading.

Three ways to compute it

The number depends heavily on method, which is worth knowing before comparing any two figures.

Historical simulation applies actual past moves to the current portfolio and reads off the relevant percentile. Assumption-light and honest, and it can only produce events that have already occurred. A portfolio with no bad history has no bad scenarios.

Parametric assumes a distribution — usually normal — and computes from volatility and correlation. Fast and analytically convenient. Financial returns are demonstrably not normally distributed; extreme moves happen far more often than the assumption allows, and the tail is precisely where the number is supposed to be looking.

Monte Carlo simulates many paths under an assumed process. Flexible enough for complex portfolios, and entirely dependent on the assumed process, which is the thing you were trying to measure.

All three are defensible. They give different answers, and comparing figures computed differently is comparing different quantities.

The problem that matters

The deepest issue is not any of those. It is that risk is estimated from recent data, and recent data is a poor guide at exactly the moment it counts.

Volatility clusters. A calm period produces low measured volatility, which produces a low risk number, which permits larger positions. A stressed period produces the reverse.

So the measure gives most permission when conditions are calmest — which is typically when positions are most crowded and valuations most stretched. And when stress arrives, the number rises sharply, forcing position reductions from everybody using a similar framework at the same time.

A risk model calibrated on recent history says a market is safe in proportion to how long it has been quiet. If enough participants use it, that shared judgement becomes the reason a quiet market stops being one.

This is the same pro-cyclicality as margin and haircuts, arriving through a different door — and it will be the whole of the closing issue.

What to do with it

None of this makes the measure useless. It makes it a tool with a specific job.

It is genuinely good at comparing portfolios on a consistent basis, at tracking changes in a portfolio’s risk over time, and at allocating limits across desks. Used as a relative measure with a stable methodology, it does real work.

It is bad at absolute statements about how much could be lost, and it is useless as a crisis predictor because it is calibrated on the absence of one.

The standard responses are to supplement it rather than replace it. Expected shortfall averages the losses beyond the threshold, which at least says something about the tail rather than only where the tail begins. Stress testing abandons the historical distribution and asks what a specified scenario would do — the subject of the issue after next.

Neither removes the underlying difficulty: the future contains events not in the sample, and no statistic computed from the sample can tell you about them.

Three things worth keeping

It is a quantile, not a worst case, and says nothing whatever about the bad day.
Method changes the answer. Historical, parametric and simulated figures are different quantities.
It is pro-cyclical by construction, granting most permission when markets have been calmest.

Next week: the input every risk model depends on most, and the one that is least stable.

The tape

Sun, 15 Feb 2026 09:00:00 +0000

Every trade in a regulated market is reported. Those reports are assembled into a consolidated record — the tape — which is the official account of what traded, where, and at what price.

It sounds like a solved problem. It is one of the most contested pieces of infrastructure in modern markets.

Why it is hard

Three difficulties, none obvious from outside.

Trades happen in many places. Exchanges, alternative venues, dark pools, internalisers and over-the-counter arrangements all produce trades. Each reports, and somebody must combine them into one sequence.

Reports arrive out of order. A trade executed at 10:00:00.001 on one venue and another at 10:00:00.002 elsewhere may reach the consolidator in either order, depending on distance and processing. Reconstructing the true sequence requires accurate timestamps from every contributor, synchronised to a common clock — which took regulation to achieve and is still imperfect at the finest resolutions.

Not everything is comparable. A trade at the midpoint of a dark venue, a negotiated block reported hours late, a trade that was part of a portfolio deal, and an ordinary lit execution are all trades. Treating them as equivalent produces a misleading picture, so each carries condition codes describing what it was — and using the tape properly means understanding those codes rather than counting rows.

What it is for

The tape does several jobs, and they pull in different directions.

Regulatory oversight. Reconstructing what happened, detecting abuse, supervising participants. This wants completeness and accuracy above all, and it can tolerate delay.

Best execution. Judging whether a client was well served requires knowing what was available at the time. This wants accurate timestamps and full coverage.

Price discovery for participants. Knowing where an instrument last traded, and in what size. This wants speed above everything.

Those requirements conflict. A record optimised for completeness is slower than one optimised for speed, and the argument about consolidated tapes is largely an argument about which purpose comes first.

The commercial fight

Which is why the topic generates so much heat.

The venues that produce the trade reports also sell market data, as this letter describes later. A comprehensive, cheap consolidated tape competes with their own products. A slow, expensive or incomplete one leaves the direct feeds more valuable.

So the parties who must contribute to the tape have a commercial interest in it being less useful than it could be. That is not a conspiracy; it is a straightforwardly misaligned incentive, and it is why consolidated tapes tend to require regulatory compulsion rather than emerging on their own.

The record of what happened is produced by the parties whose commercial interest lies in that record being less useful than it could be. Nothing about that arrangement is going to fix itself.

The European experience is instructive: for years, trade reports existed but were scattered across venues in inconsistent formats at prices that made assembly uneconomic. Nobody was hiding anything. The data was simply not consolidated, because nobody with the ability to consolidate it had a reason to.

What it means for reading anything

Practical guidance, and it echoes the caution in this letter’s later issue on market data.

Delayed reports distort the sequence. A block reported hours after execution appears at the wrong point. Any time-series analysis needs to use execution timestamps rather than report timestamps, and needs to know the difference exists.

Volume totals depend on definitions. Whether to include negotiated trades, give-ups, or technical prints changes the total materially. Two honest sources disagree because they counted different things.

Not every print is a price. A trade at a price away from the market may be perfectly legitimate — a portfolio transaction, a delayed report, an option exercise. Treating every print as evidence of where the instrument was trading will mislead.

Three things worth keeping

Consolidating trades across venues is genuinely hard, and clock synchronisation had to be mandated before it was possible.
Speed, completeness and accuracy conflict. Which the tape optimises for determines who it serves.
Read the condition codes. Not every print means what a naive reading assumes.

Next week: the number risk managers report, and the question it does not answer.

Obliged to quote

Sun, 08 Feb 2026 09:00:00 +0000

A market maker quotes because it is profitable. When it stops being profitable they stop, which is precisely when everyone else needs them most.

Venues have tried to fix this with contracts: designated participants who agree to quote continuously in exchange for benefits. Whether those contracts do anything useful is one of the more interesting unresolved questions in market structure.

The bargain

The obligations are typically specific.

Quote two-sided prices in named instruments. Maintain a spread no wider than a stated maximum. Show at least a minimum size. Be present for a defined proportion of the trading day.

In exchange the firm receives reduced fees or enhanced rebates, sometimes preferential treatment in the matching queue, and occasionally direct payment from the issuer — common for smaller listed companies who need someone to make a market in their shares at all.

For thinly traded instruments this arrangement is genuinely valuable. Without it, some shares would have no continuous market, and the obligation to be present is what creates one.

Where it fails

The difficulty appears in stress, and it comes from the parameters.

Every obligation has a maximum spread and a minimum size, and both are finite. A market maker meeting a requirement to quote at no worse than a stated spread, in a stated size, has satisfied the contract completely — even if that spread is several times normal and that size is trivial relative to what is being traded.

So the obligation binds where it does not matter and gives way where it does. In calm conditions the firm quotes far better than required, because competition forces it. In stress it retreats to the contractual minimum, which is fully compliant and close to useless.

An obligation to quote is only as good as its worst permitted quote. In the conditions where it matters, that is exactly the quote you will get — because anything better would be voluntary, and nobody is volunteering.

Widening the requirement does not obviously help. A firm required to quote tightly in all conditions is being required to lose money in a crisis, and it will either demand compensation exceeding the value provided or decline the designation entirely. There is no free obligation.

The older model

Some markets historically went further, with a single participant responsible for each instrument — a specialist, holding an affirmative obligation to maintain a fair and orderly market, including trading against the trend when there was an imbalance.

That model concentrated obligation and privilege in one firm. It had a genuine advantage: a participant with a monopoly on the order flow in an instrument can afford to lose money stabilising it, because the franchise is worth defending.

It also had the obvious weakness of any monopoly, and it created conflicts between the specialist’s own book and its clients’ orders. Electronic competition largely replaced it — and what replaced it was a set of participants with no obligation to anyone, competing on speed, free to withdraw at will.

Whether that was an improvement depends on which condition you weight. Ordinary trading is unambiguously cheaper. Stressed trading arguably has less standing behind it.

The honest position

There is no settled answer here, and it is worth saying so plainly.

Obligations that bind meaningfully in stress require compensation that someone must pay. Obligations that are cheap enough to attract willing participants are, by construction, weak enough not to matter when tested.

Every market structure sits somewhere on that trade-off, mostly without having chosen deliberately. It is a design decision that is usually made by default, and it determines what happens on the days that matter.

Three things worth keeping

Designated market making creates continuous quotes in instruments that would otherwise have none. For illiquid names this is its whole justification.
The binding constraint is the worst permitted quote, and in stress that is what you will be shown.
There is no free obligation. Meaningful commitments in crisis cost money, and the question is only who pays.

Next week: the official record of what happened, and why it is harder to produce than it sounds.