Market Making Explained: Who Is On the Other Side of Your Fill

The counterparty is a business, and the spread is its revenue line

Start with a question almost nobody asks after a fill: why was anyone willing to take the other side of that trade, at that instant, at that price? You wanted to buy. Somewhere in the book there was no natural seller who wanted to sell exactly your quantity at exactly that moment. Somebody sold to you anyway. That somebody was not doing you a favour and was not lying in wait for you personally. It was running a business, and it took your trade because the terms were good enough to be worth the risk.

The product that business sells is immediacy. The alternative to buying it is waiting, possibly for a long time, for a genuine counterparty with the opposite need and the matching size to show up. Most traders do not want to wait, and the ones who do wait usually discover that the market has moved on while they were being patient. So a participant steps in, quotes a price at which it will buy and a price at which it will sell, and absorbs the mismatch in timing onto its own book. It buys from you now and finds the real buyer later. That gap in time is the entire business, and everything that can go wrong lives inside it.

What the spread is and how it appears in your fill price is covered from the retail side in the guide to the bid-ask spread, and this page assumes it. The question here is the one a level up: given that somebody earns the spread, what does it cost them to earn it, and what does that cost imply about the number you see quoted?

Framing it as a price rather than a fee changes what you look for. A fee is set by whoever charges it. A price is set by cost and competition, which means it is discoverable. If you can work out what it costs to supply immediacy in a particular instrument, you can work out roughly what the spread on that instrument has to be, and you can tell the difference between a spread that is wide because the instrument is dangerous and a spread that is wide because nobody is competing. Those two look identical on a screen and they call for different responses.

The two ways the business loses money

A quoting business has exactly two structural exposures, and they are routinely blurred together even though they behave completely differently.

Inventory risk is about what happens after the trade. When the maker buys from you it now owns something it did not want, chosen by whoever happened to arrive rather than by any view about the price. If a run of sellers arrives, it accumulates a long position. If the price then falls, it loses, and it loses on a position it never decided to take. The direction of that exposure is set by the shape of the flow, and the flow is not something the maker controls.

Adverse selection is about who chose to trade. A quote is a firm price that anybody may take and nobody has to take. That asymmetry is the whole problem. Everyone who arrives gets a free look at your price and trades only if it suits them. The traders who find it most attractive are disproportionately the ones who have noticed something you have not yet priced: a headline, a move in a related instrument, the start of a large order being worked. So the maker collects small amounts from traders with no view and pays out larger amounts to traders with one. There is nothing improper about it. It is arithmetic, and it applies to every standing quote that has ever existed.

The useful way to hold the two apart is this. Inventory risk would exist even if every counterparty were completely uninformed, because random flow still leaves you holding something. Adverse selection would exist even if you could hedge your inventory perfectly the instant you acquired it, because the price you traded at was already wrong. They are separate leaks and they need separate repairs, which is why the model below measures them separately rather than lumping them into a single cost of doing business.

There is a third leak that this page does not model: the plain cost of running the operation, meaning connectivity, clearing, statutory charges, staff and the capital tied up. Leaving it out makes the business look better than it is, and the residual computed below should be read with that in mind. The depth of the book that all of this plays out in is described from the retail side in the guide to market depth.

The model, stated in full so it can be checked

The simulation below is a notional instrument, not any real security, deliberately. A simulated instrument lets every parameter be stated openly, keeps the run reproducible from a seed, and avoids implying anything about how any particular stock or index behaves. Everything on this page comes from one run of the configuration in the table.

Two settings deserve a note before the results. The information lag is doing all the work: if the maker could observe the true value with no delay, no counterparty could ever pick it off and there would be no adverse selection at all. Setting the lag at 8 arrivals gives the maker a mispricing with a standard deviation of about 5.6 basis points, which is 4.7 times the half-spread it is quoting. That single ratio drives almost everything that follows. The inventory skew is switched on by default because a desk that does not manage inventory is not a business, it is a coin toss, and the section on inventory shows precisely how large a coin toss.

Where the spread revenue actually goes

Run the configuration and the top line looks healthy. Across 250 sessions the maker filled an average of 4,335 orders a session and captured Rs 26,011 of gross spread. Then the deductions arrive.

Adverse selection took Rs 22,004, which is 84.6 percent of the gross. Inventory took Rs 68, which is 0.3 percent of the gross and is dealt with separately below because the reason it is nearly zero is the most interesting result on the page. What survived was Rs 3,939, or 15.1 percent of the gross spread revenue.

The per-fill arithmetic makes the mechanism visible. Every fill of 100 shares earns the maker Rs 6.00 of quoted half-spread, win or lose. Of the 4,335 fills in an average session, 913 came from informed arrivals, which is 21.1 percent of them, and each of those cost Rs 24.10 in mispricing. One fill against informed flow therefore undoes what four ordinary fills earned. The business is not thin because the spread is small. It is thin because roughly one fill in five is against somebody who is right.

Put the residual against the volume it took to produce. The maker turned over 21.7 crore rupees of stock in a session to keep Rs 3,939, which is about 0.182 of a basis point, or Rs 1.82 for every lakh of stock intermediated. That is what a competitive quoting business looks like from the inside: enormous throughput, microscopic margin, and no room at all for a mistake in the risk model.

The distribution matters more than the average. The standard deviation of the session result was Rs 2,941 against a mean of Rs 3,939, the best session made Rs 10,727, the worst lost Rs 8,033, and 23 of the 250 sessions lost money outright. A day is almost pure noise. Only the year is a business, and the year came to Rs 9,84,692 on this configuration, which is a modest reward for the amount of risk that had to be carried to earn it.

The spread is set by the toxicity of the flow, not by greed

The obvious next question is why the spread in this model is 2.4 basis points and not ten times that. Nobody chose it as a matter of preference. It is close to the narrowest spread at which the business survives, and that number is computable.

The condition is straightforward to state. The maker earns its half-spread on every fill, including the ones against informed traders. It loses, on each informed fill, the amount by which the quote was stale beyond the half-spread. Informed traders only trade when that amount is positive, so their fills are always losses. Break-even is the spread at which what the uninformed pay exactly covers what the informed take. Solve that condition for the spread, once for each level of informed flow, and you get a curve.

The shape is the finding. With five percent of arrivals informed, the business survives on a spread of 0.447 basis points. At twenty-two percent it needs 1.997. At sixty percent it needs 6.151, which is 13.8 times the first figure, on an instrument with identical volatility, identical order size and identical everything else. Nothing about the maker changed. Only the composition of who was arriving changed.

The open circles are the same break-even found by brute force in the full simulation, with inventory and skew switched on rather than assumed away. They sit consistently a little above the closed-form line: 2.048 against 1.997 at twenty-two percent informed, and 6.521 against 6.151 at sixty percent. That gap is the cost of managing inventory, which the closed form ignores entirely. It is small, and it is real, and reporting it is cheaper than pretending the tidy formula is the whole answer.

For a trader reading a quote screen, the practical translation is direct. A spread is a readout of how dangerous the flow in that instrument is judged to be. Two instruments with the same price and the same volatility can quote spreads an order of magnitude apart purely because of who trades them and what those people tend to know. That is why a spread widening ahead of an announcement is not opportunism: the expected share of informed arrivals in the next few minutes has genuinely jumped, and the quote is repricing the risk. What that costs you as the taker is the subject of the guide to liquidity in trading, from the other side of the same transaction.

Inventory is the position nobody chose to take

The inventory line in the decomposition was Rs 68, small enough to disappear. That is not because inventory risk is trivial. It is because of two things, and both of them are worth stating plainly rather than hiding behind the tidy number.

The first is that in a market with no drift, a position acquired at random is a coin toss, and a coin toss has no expected cost. What it has is variance. Switch the inventory skew off and run the identical flow again: the average session result barely moves, from Rs 3,939 to Rs 2,631, but the standard deviation goes from Rs 2,941 to Rs 37,471, a factor of 12.7, the largest position held in a session rises from 1,538 shares to 9,791, and the number of losing sessions goes from 23 to 88 out of 250. The expected profit is almost unchanged and the business has become unrunnable. Inventory risk costs the average almost nothing and costs the survival almost everything.

The second is that real markets are not driftless. Give the value a trend and the arithmetic changes completely, because a lagging quote in a rising market is systematically too cheap, which means the informed side is systematically the buying side, which means the maker is systematically the seller. It accumulates a short position not by mistake but by construction.

In the representative session drawn above, with the skew off, the maker ended 15,100 shares short, having been as much as 17,100 shares short during the day, and inventory alone cost Rs 70,535 against gross spread revenue of about Rs 28,998 for the day. The session lost Rs 71,034. Nothing broke. No single trade was a disaster. The desk simply took the other side of a one-way market all day, which is precisely what it is there to do, and the position it was left holding cost about 2.4 times everything the spread earned that day.

With the skew on, the same flow left the maker 800 shares long at the close, having never exceeded 1,600 shares in either direction, and the session finished close to flat. Across 120 trending sessions the unskewed version averaged a loss of Rs 97,945 a session, with a standard deviation of Rs 1,17,519 and a worst session of Rs 5,16,124 lost, and it lost money in 96 of them. The skewed version averaged Rs 2,286 with a standard deviation of Rs 3,832 and lost money in 34.

The mechanism deserves a sentence because it explains something you can see on a screen. Skewing does not widen the spread. It moves the whole quote. A maker that is short raises both its bid and its ask: it becomes eager to buy and reluctant to sell. Under price and time priority a two-paise improvement is the difference between the back of a queue and the front of it, so a shift far smaller than the spread changes the fill rate dramatically. When the side of the book you want seems to step away from you, that is very often not a judgement about your price. It is a desk trying to get flat, and you happen to be standing where it wants to go.

The skew is not free. Shading the quote towards getting flat means giving up a little on the fills that reduce the position, and the gross spread captured in the skewed trending session was 9.5 percent lower than in the unskewed one. That is the trade the whole discipline consists of: a small, certain reduction in revenue in exchange for a large reduction in the chance of a session that ends the business.

The round trip, read as a ledger

It helps to see one completed round trip as a set of entries rather than as a price movement. Nobody in the chain is doing anything unusual, and the money does not vanish into an institution. It moves, and the interesting part is where it ends up.

Read down that column and the popular picture inverts. The largest single transfer in a round trip is not from the retail trader to the market maker. It is from the market maker to whoever was better informed. The retail trader funds a part of it, and the compensating service they receive is that their order was filled at all, instantly, at a price within a known distance of fair value.

Your own limit order is a market-making book of one

Everything above becomes personal the moment you stop crossing the spread and start posting a price. A resting limit order is a firm quote that anybody may take and nobody has to take. It has the identical asymmetry, and so it has the identical two risks, scaled down to one order.

The trade-off between the two order types, and the stop-loss variants of each, is set out in the guide to limit orders and market orders. What follows is not that comparison. It is the part that never gets computed: what the selection actually costs when you are the one quoting.

The second model on this page simulates 200,000 attempts to buy passively. The trader decides to buy, posts a limit at the bid rather than lifting the ask, and leaves it working for a fixed window. The instrument quotes 20 basis points, the trade idea carries a stated raw edge of 15 basis points to the exit, and the fill rules are the important part, so they are stated exactly. If the market trades through the resting price by more than 4 basis points, the order fills for certain. If it merely touches the price, the order fills with probability 0.35, because there is a queue in front of it. While the order is still at the touch there is a 1 percent chance per step of an ordinary seller hitting it. Once the mid has risen above the order it does not fill at all.

Those rules are not a caricature and they are not tuned to produce an answer. They are the four things that can actually happen to a resting order, and the selection falls out of them.

The fill rate was 64.1 percent. The average attempt, taken at the decision price, went on to move 15.16 basis points by the exit. The attempts that filled went on to move 1.96. The attempts that never filled went on to move 38.69, which is 2.6 times the average. The passive entry did save the half-spread it was supposed to save, and then handed 13.2 basis points of it straight back through the selection, because the orders that get filled are the ones the market is walking through and the orders that do not are the ones that ran away.

Chasing makes it worse in a specific and measurable way. Take the attempts that did not fill and buy them at market at the end of the waiting window, and their average outcome is 6.08 basis points of loss, against the 38.69 they were worth at the decision price. The opportunity was real. It was simply over by the time the chase started, and the full spread still had to be paid. This is the most expensive habit in the entire comparison and it is the one that feels the most reasonable in the moment.

Paying the spread, or trying to earn it, over a year

To compare the two routes fairly you need a common benchmark, so the accounting below measures each against a frictionless fill: the same idea executed at the mid, with no spread and no fill risk. The difference between that and what each route delivered is the friction, and it is the only number worth comparing.

Crossing the spread produced friction of exactly 20 basis points, which is the quoted spread. That is not a finding, it is a check on the model: a market order in and a market order out costs one full spread and nothing else. Resting produced friction of 13.9 basis points, made up of 8.5 from selection on the fills that happened and 5.4 from the trades that never did. Resting and then chasing produced 16.09, worse than simply crossing at the start.

Scaled to a year on Rs 1,00,000 committed per trade, the spread stops being a rounding error somewhere around a hundred round trips. At 25 round trips a year, crossing costs Rs 5,000. At 250, it costs Rs 50,000. At 750, which is three round trips on an average trading day, it costs Rs 1,50,000 on a position size of one lakh. The resting route on the same assumptions costs Rs 3,476, Rs 34,759 and Rs 1,04,276 respectively.

The width of the quoted spread scales the whole picture linearly and it varies enormously across instruments. On the same one lakh of notional, one round trip costs Rs 50 at a spread of 5 basis points, Rs 200 at 20, and Rs 600 at 60. Two hundred and fifty round trips a year in the widest of the three costs Rs 1,50,000, which is more than most retail traders believe their entire annual cost of trading to be.

The honest reading of this comparison is narrower than it looks. Resting came out cheaper on these assumptions, but not because patience is free. It came out cheaper because it declined about a third of the attempts, and that is only an advantage if you can genuinely walk away from a trade rather than reaching for it a minute later. Change the fill rules, the waiting window or the edge and the ranking can flip. What does not flip is the structure: crossing pays one known cost, and resting refuses that cost and accepts two unknown ones in its place. There is no third door, and anyone who tells you a limit order simply saves the spread has not measured the fills they did not get.

When the quote widens, and what it is telling you

Because the spread is a price for risk, every widening has a cause, and most of the causes are legible if you know what the quoting business is worried about. The table reads each one from the maker's side and then translates it.

The Indian context: an order-driven auction with no obligation to quote

One structural fact shapes everything above in the Indian cash market. Equities trade in an order-driven continuous auction with price and time priority, not in a quote-driven market with an appointed dealer standing between buyer and seller. There is no designated firm whose job it is to make a market in an ordinary stock. Anyone who rests a limit order is supplying liquidity for as long as that order rests, and the participants who do it continuously are market makers by behaviour rather than by appointment.

Three consequences follow. The first is that the supply of quotes can simply leave. Nothing obliges a voluntary quoter to keep a price on the screen, and the conditions that make quoting unattractive, which are volatility and a rising share of informed flow, are exactly the conditions in which the impulse to trade is strongest. The thinning is not a coincidence and it is not a conspiracy. It is the same calculation as the break-even curve above, resolved in favour of standing aside.

The second is the tick. Price and time priority means the queue is ordered by price first, and the smallest permitted price increment is the smallest step you can take to jump that queue. It also puts a hard floor under the spread, because a spread cannot be narrower than one tick. On a low-priced instrument that floor alone can be worth a meaningful number of basis points however many participants are competing to quote, which is why a cheap stock is not automatically a cheap stock to trade.

The third is the price band. Circuit limits cap how far a price may travel in a session, and for a business whose entire model depends on being able to unwind a position, a cap on the exit is a serious constraint. As price approaches a band the quoting calculation degrades quickly, because the inventory you are about to acquire may not be sellable at any price inside the band. Widening or withdrawing is the rational response. The session also does not begin with continuous quoting at all: the opening price is discovered in a call auction, and two-way quoting starts afterwards, which is one reason the first minutes look and behave differently from the rest of the day.

There is one corner of the Indian market where quoting is contractual rather than voluntary. The mutual fund framework requires the issuer of an exchange traded fund to appoint market makers to provide two-way quotes, and those quoters can hedge in the underlying basket rather than carrying naked inventory, which changes the economics considerably. That mechanism, and the arbitrage that keeps an ETF price tethered to the value of what it holds, is set out in the guide to ETF arbitrage.

What the model leaves out, and what to do with what is left

The honest boundary first. This model has one maker, always quoting, in one instrument, against orders of uniform size. Real quoting is a race between several participants for the same fill, in which latency decides who gets the good trades and who gets the leftovers. Real flow arrives in wildly varying sizes, and a single large arrival can do more damage than a thousand small ones. Real makers hedge in correlated instruments rather than carrying the raw position. Real makers can stop quoting, and the option to stop is worth a great deal precisely when the model above says the business is losing money. Every one of those omissions, along with the excluded running costs, makes the residual computed here more generous than the real thing rather than less.

What survives all of that is the structure, and the structure is what changes how you behave. The disappearance of quotes when a strategy expected them is the single largest gap between a tested rule and a traded one, which is worked through in detail in the companion piece on the gap between backtest and broker.

Three habits follow from the arithmetic on this page. Read the spread as a risk premium, not as a fee. A wide spread is information, and the information is about the flow rather than about the greed of whoever is quoting. When it widens suddenly, something in the composition of the arrivals has changed, and you are almost certainly part of what changed.

Decide before you click which of the two risks you are choosing to bear. Crossing the spread buys certainty and pays a known price for it. Resting refuses the known price and takes on selection and absence instead. Both are defensible. Reaching for a limit order to avoid a cost, and then crossing anyway when it does not fill, is the one combination that pays for both.

Put a number on it once a year. Multiply your typical spread by your notional and by your round trips, and compare the answer with what you think your trading costs. Most people find the spread is the largest line and the one they have never measured, because it never appears on a statement. Doing that arithmetic before deciding how often to trade, rather than after, is the sort of unglamorous discipline that separates a process from a habit, and it is the method we teach.

Every figure comes from two deterministic simulations in a single seeded script, so the run reproduces identically. The first models a quoting business against a mixed flow of informed and uninformed arrivals, with the parameters in the assumptions table. Its profit is decomposed exactly into gross spread captured, adverse selection and inventory, and the script asserts that the three components sum to the mark-to-market change in the maker's wealth before any of the numbers are published. The break-even spread is computed in closed form from the same assumptions and then verified independently by bisection on the full simulation, with the two results reported side by side rather than merged. The second simulation runs 200,000 attempts to buy passively at the bid, with the fill rules stated in the text, and its friction decomposition is likewise checked to sum to the whole before being reported.

No real instrument is modelled and no market data is used. Where an Indian structural detail is described, it is described qualitatively, and specific numeric values such as minimum tick increments have been deliberately left off the page because they could not be confirmed from a primary source at the time of writing. All results are illustrative and simulated. They are not a track record, not a forecast, and not an indication of what any participant or any strategy would produce in a live account. The purpose of the exercise is to show how the price of immediacy relates to the risk of supplying it, which is a property of the mechanism rather than of any particular market.

• What slippage is, and where it comes from →
• Reading order flow and the tape →
• Why scalping is a cost problem first →
• The eight ways a backtest can lie →