What market sentiment and positioning describe

Sentiment, positioning and price are three different things 

Sentiment is what participants say when they are asked. Positioning is what they hold when they are counted. Price is what the most recent transaction settled at. The three are related, and they are produced by entirely different processes with entirely different limits. A survey asks a panel to describe a mood, and a description costs nothing to give. A positioning report counts contracts that exist, and every one of them was paid for and is still being funded. A price is a single observation, the latest one, and it carries no information about how many participants agree with it or how much they hold at it.

The gap between what is said and what is held is not a failure of honesty. A fund manager can describe conditions as poor and hold a large position anyway, because the position was taken for a reason the survey question did not ask about, because a mandate requires the exposure, or because the cost of closing it exceeds the discomfort of holding it. This is why a survey and a positioning report frequently point in opposite directions in the same week, and why nobody who works with both treats them as one reading. The word sentiment is used loosely for all three in commentary, which is the source of the muddle rather than a shorthand for resolving it.

What is actually published, and what each measure covers 

There is no single measurement of a market's positioning, because there is no register of who holds what. What exists instead is a set of partial counts, each covering whichever slice of participants happens to be observable, usually because somebody is required by a regulator to report it or has chosen to publish it. The families below are the ones a reader will meet.

• Regulated futures reports. Exchange traded futures are counted by the regulator that supervises them, and in the United States the weekly commitments of traders report splits open interest into categories of reporting participant, so the long and short holdings of the managed money category in a currency, an index or a commodity future are public. It counts the futures market only, not the far larger over the counter market beside it.
• Broker book data. A firm may publish the balance of its own clients, conventionally as the percentage of accounts long and short in an instrument. It describes that firm's clients on that firm's book and nobody else's.
• Investor and fund manager surveys. A panel is asked to state a view, and in the better designed surveys to state an allocation. The view questions measure statements. Only the allocation questions come close to measuring commitment, and even those are self reported.
• Options market data. Open interest by strike and expiry is published, and from it the relative cost of protection against a fall and against a rise is derived. It measures what has been paid for, in a market where a participant may be hedging an exposure that is nowhere in the data.
• Fund flow data. Subscriptions into and redemptions out of pooled vehicles are counted. That is a record of money arriving and leaving, which is a slower and blunter thing than a position.

Not one of these covers a whole market, and the differences in coverage are not rounding errors. A major currency trades continuously across banks, corporates, funds, exchanges and central banks. The futures report sees a fraction of that activity, and a single broker's book is a fraction of that fraction. A reading described as the market being positioned in a certain way is always, on inspection, a reading about one measured population.

One further class of sentiment is not a count at all. The published account of a rate setting meeting reports the balance of opinion among the members who attended, so it is read for how many favoured what, and the language it uses is the vocabulary an earlier lesson in this module set out. It is closer to a census of stated views than to a survey, because the participants are named, few and consequential.

A headcount is not an exposure 

The most widely quoted measure is also the most widely misread. A percentage of accounts is a headcount, and a headcount weights an enormous position and a tiny one identically. Nothing prevents a large majority of accounts from sitting on one side of an instrument while most of the money on the same book sits on the other, and the arithmetic that produces that result is not exotic.

A published percentage of accounts answers the question of how many. Any sentence that treats it as an answer to the question of how much has quietly substituted one measurement for another. Regulated futures reports avoid that particular trap by counting contracts rather than participants, which is why they are the reference measure in most professional commentary, and they acquire a different limitation in exchange: they say nothing about how large the reporting entity is or what else it holds.

Every one of these measures describes the past 

A positioning report is a photograph, and it is developed slowly. The count is taken at a stated moment, the data is compiled and checked, and the report is published some days later. Between the snapshot and the publication the market carries on trading, and every position opened, closed or reversed in that interval is invisible until the following report. The regulated futures reports are the clearest case because their timetable is published, which makes the lag measurable rather than merely suspected.

The lag is not a defect to be corrected. It is a property of a count that has to be collected from reporting entities and checked before it is published, and every measure on the list above has some version of it. A monthly survey is a month old at its widest. Fund flow data lags the decisions that produced it. Broker book data is the freshest of the family and is also the narrowest population. The practical consequence is uniform: a positioning reading describes a market that has already moved on, and by an unknown amount.

What a crowded position describes 

A position is described as crowded when a large share of the measured participants already hold it. The word describes a stock rather than a flow, and its consequence is arithmetic before it is anything else: everybody who already holds the position has already bought it. The remaining buyers for that position are the participants who have not taken it yet, so the more of them have taken it, the fewer are left. Crowding is a statement about the size of the pool of available counterparties, and that is the whole of what it is a statement about.

The mirror image is the half most often left out, and it is what makes the concept symmetrical rather than directional. Every open position is a future transaction in the opposite direction, because a position can only be closed by one. A large one sided holding is therefore a large quantity of eventual selling in the case of a crowded long, and a large quantity of eventual buying in the case of a crowded short. Neither statement contains any information about when that transaction occurs, what prompts it, or at what price it is done. A crowded long and a crowded short are the same structure with the signs reversed, and the reasoning applies to both identically.

Crowding also meets the mechanics an earlier module described. A position is closed by its holder, and it is also closed by the counterparty when the equity backing it falls to the published close out level, which is a closure that consults nobody's opinion. Because a margin requirement is a percentage of the full contract value, a loss is measured against the whole contract rather than against the collateral, and is not limited to the amount deposited. Where many participants hold a similar position at similar sizes, that mechanical closing can occur across many accounts in the same short window, which is one description of why moves against a heavily held position sometimes travel further than the news that started them.

Why an extreme reading carries no timing information 

An extreme reading is a statement about a distribution: this observation sits far from the observations before it. A distribution has no clock inside it. Nothing in the arithmetic of a percentile, a standard deviation or a series high specifies an interval, so a measurement can report that a holding is unusually large and cannot report how long it stays that way. The two questions are answered by different objects, and only one of them is being measured.

The convention built on this is the contrarian reading, and it is genuinely held. Traditions that follow positioning data conventionally describe a reading far from its own history as stretched, on the reasoning set out above: the participants who would take the position have largely taken it, so the pool of remaining buyers is thin while the quantity of eventual sellers is large. The convention carries its own contradiction in the same breath, and practitioners who use it say so. Readings that are already extreme become more extreme with some regularity, and nothing in the measurement bounds how far or for how long. A record reading is by definition the largest observation so far, which is exactly what every reading on the path to a larger one also was.

The final row is the entire argument. Because the reading is identical at the start of both paths, it cannot be what separates them. Practitioners who work with positioning data conventionally pair it with something that dates a change, most often the price series itself, and they disagree about whether that pairing adds information or simply relocates the judgement into the second measurement. This page describes the convention and the objection, and puts forward neither as a method.

Risk on, risk off, and what happens to correlation 

Sentiment has a portfolio level vocabulary of its own. Periods in which participants collectively favour the assets that pay more when growth is strong, and periods in which they favour the ones that hold value when conditions deteriorate, are conventionally labelled risk on and risk off. The labels are descriptive shorthand for a pattern observed across many instruments at once, and they are applied after the fact rather than declared in advance.

The consequence that matters is not the label but the correlation underneath it. In those periods instruments that are unrelated by industry, geography or asset class begin to move together, because what is being repriced is not the thing that usually distinguishes them. It is one common question about growth, inflation and the path of interest rates, asked of everything at the same moment. That is the transmission chain this module opened with, and it is why the effect is so broad: an index, a currency, an industrial metal and a bond all sit downstream of the same chain, so when the chain moves, they move with it.

For anyone holding more than one position, that has a structural effect and it does not require a single trade to bring it about. Positions that were taken as separate propositions become one proposition expressed in several instruments, so the number of open positions overstates the number of distinct exposures. The reverse is equally true and equally worth stating: in calmer periods the same instruments separate again and resume behaving as the different things they are.

Correlation is also measured backwards. A correlation figure describes a past window of a chosen length, and choosing a different length gives a different figure from the same data. Relationships that have held for months break, and the periods in which they break are frequently the periods in which they had appeared most dependable. The standard caution about a correlation is therefore the same as the caution about every other measure on this page: it describes a window that has already closed.

Where practitioners disagree 

The first disagreement is about whether the categories in a positioning report describe what their names claim. A regulated report sorts participants by the business a reporting entity registered as, so a category named for commercial activity contains firms hedging genuine physical exposure and also firms whose hedging arm expresses views, while a category named for managed money contains strategies with opposite horizons and opposite reasons for holding the same contract. One camp holds that the categories are still informative in aggregate because their compositions are stable over time. The other holds that a stable mixture of incompatible motives is a stable number that means nothing in particular. Both positions accept the same data.

The second is the direction in which the same reading is read. Contrarian traditions treat a large one sided holding as a description of a depleted pool of further buyers. Momentum traditions treat a growing one sided holding as a description of a flow that is continuing, and read the same rising series as confirmation rather than as a warning. Neither camp can settle the argument with the data, because the identical series is the input to both, and the module's earlier lessons apply here too: the reading is a measurement, and a measurement of what has been done is not an account of what happens next.

The third concerns broker book data specifically. One view is that a firm's own clients are a legitimate sample of a broader participant type and that the balance of their positions is therefore informative about it. The objection is that a single firm's clients are self selected by that firm's pricing, its markets and its marketing, so the sample is not drawn from any defined population and its relationship to the wider market is unknown. There is no test that settles this, because the wider figure the sample would be compared against does not exist.