Market making

Quote both sides, earn the spread on uninformed flow, manage inventory, and survive adverse selection. The single biggest live commercial family: structural, still alpha for the equipped.

See it move

What is market making?

Market making is continuously posting a two-sided quote (a bid and an ask) on one instrument, standing ready to buy from sellers and sell to buyers. The maker earns the spread between its quotes as compensation for providing immediacy (others can trade now, without waiting for a natural counterparty) and for bearing the risk of holding inventory and being adversely selected.

Picture a stall in a currency bureau. You post "I'll buy dollars at 99, sell them at 101". Tourists who want dollars now pay your 101; tourists offloading dollars take your 99. If buys and sells roughly balance, you pocket 2 per round trip and never carry a large position: that 2 is the spread, your pay for being always-on. The whole difficulty is what happens when they don't balance, and when the people trading with you know something you don't.

A market maker is a liquidity provider (a maker: it posts resting limit orders into the limit order book), as opposed to a liquidity taker who crosses the spread with marketable orders (see maker vs taker). You are paid the spread (and often a rebate) precisely because you take on the risks the taker offloads onto you.

This is principal activity: you quote with your own capital and own the resulting position and its P&L. That is the difference from agency execution, where you trade a client's order to minimise their cost. A market maker is the counterparty an execution algorithm trades against. The firms that do this at scale (designated market makers, electronic liquidity providers and principal trading firms such as Citadel Securities, Virtu, Jane Street, Optiver, IMC and Jump) sit alongside, on the open frontier, crypto market makers and independent quoters on prediction markets (the full taxonomy is in market participants).

How does a market maker earn the spread?

The maker posts a bid below a fair-value estimate and an ask above it. When a seller hits the bid and, later, a buyer lifts the ask, the maker has bought low and sold high without taking a view, pocketing the spread. The catch: the two fills rarely arrive paired and balanced, and the price moves in between.

The spread is the maker's gross margin per round trip (one buy plus one sell). If the mid is 100.00 and you quote 99.99 / 100.01, a completed round trip nets the full quoted spread (about 0.02) before any costs, minus fees, plus rebates, minus whatever the price did to your inventory in between. The plain-English model is a half-spread $\delta$ placed either side of a fair value $s$.

Why being filled is ambiguous news: a fill means someone chose to trade against your quote. Often that is a noise trader handing you the spread; sometimes it is an informed trader who knows the price is about to move through you. Distinguishing the two after the fact (never at fill time) is the core of adverse selection and of where the P&L actually goes (see spread vs adverse selection). The richer your fair-value estimate, the less you are picked off, which is the whole subject of Market Making II and the microprice.

What are the three forces a market maker must balance?

Market making is a three-way tension, and you cannot maximise all three at once: you choose a point in the trade-off.

Force 1 – spread capture (the reward). Widen your quotes and each round trip pays more, but you fill less often and slip down the queue; tighten them and you fill more but earn less per fill and expose yourself to being run over. There is an optimal spread, not a free one.

Force 2 – adverse selection (the counterparty risk). A fraction of the flow is informed, trading the right way just before the price moves. You systematically buy from informed sellers right before the price falls and sell to informed buyers right before it rises. This is a cost that scales with how tight you quote, and it sets a floor below which you cannot profitably make a market (Glosten–Milgrom 1985 – see adverse selection).

Force 3 – inventory risk (the position risk). Fills arrive unpaired, so you accumulate a long or short position. That inventory has price risk you weren't paid to take: a maker wants to earn the spread, not bet on direction. Left unmanaged, inventory random-walks to dangerous levels (see inventory management). The standard answer is to skew your quotes away from your inventory so the market flattens you (Avellaneda–Stoikov 2008 – see the model).

These three pull against each other. Quote tighter and you earn more spread P&L and suffer more adverse selection. Hold inventory longer and you pay less skew cost but carry more price risk. Skew harder to flatten and you give up spread on the side you are leaning. Every market-making model is a way of pricing this one trade-off; the rest of this sub-tree is the detail.

How does a market maker manage inventory risk?

By skewing its quotes. When the maker is long, it shades both quotes down (making its ask more attractive and its bid less) so the market preferentially sells it back to flat; when short, it does the reverse. If you are long 50 lots and want to get flat, you make it easy to buy from you and hard to sell to you: lower both quotes so your ask gets lifted. You are willing to give up a little spread to shed the position, and the size of that lean grows with how much inventory you hold, how volatile the asset is, and how risk-averse you are.

The principled form of this skew is the Avellaneda–Stoikov reservation (indifference) price: your private fair value, which sits below the mid $s$ when you are long ($q \gt 0$) and above it when short. You quote symmetrically around $r$, not around the mid, so a long book quotes to sell.

Simpler heuristics exist and are widely used: linear inventory skew (shift quotes by a constant times $q$), hard position limits (stop quoting the side that grows your position past a cap), and hedging the inventory in a correlated instrument. The naive ladder of these is the inventory-management page; Avellaneda–Stoikov is the principled version they approximate, treated in full on its own page.

The map of the market-making sub-tree

Market making splits into two families. Market Making I – inventory (this guide) is about managing the position you are left holding: the canonical model, the naive heuristics, and where the P&L goes. Market Making II – order flow / information is about quoting around a better fair value using the information in the book and the flow, so you are picked off less.

Market Making I – inventory (the guides in this topic):

Avellaneda–Stoikov (2008) – the canonical inventory model: reservation price, optimal spread, and the HJB problem behind them. Inventory management – symmetric vs skewed quoting, linear skew, position limits, hedging. Spread vs adverse selection – the P&L decomposition and the break-even spread. Liquidity provision as a service – the business of being paid to quote.

Market Making II – order flow / information (the companion guide):

Order-flow / information-based market making – quote around a fair value that reads the book and the flow. Adverse selection (Glosten–Milgrom 1985) – the maker's core risk, modelled. Order-flow imbalance (OFI) – net pressure at the top of book predicts the next move (Cont–Kukanov–Stoikov 2014). PIN / VPIN – measuring flow toxicity. The microprice – Stoikov's imbalance-weighted fair value, a better quote centre than the mid.

The same maths, three venues: why portability is the point

The market-making problem is venue-independent: quote two-sided, capture the spread, manage inventory, survive adverse selection. What changes across equities, crypto and prediction markets is the environment (tick size, fees and rebates, book depth, payoff bounds and hours) not the maths. That portability is the whole commercial promise: learn the model once, apply it where the competition is thin.

Equities and futures (equities & futures): small tick, maker-taker rebates, deep books, Reg NMS / MiFID II structure, and intense competition. The maths is textbook here, but the edge is largely a latency-and-scale utility (2026): a hard arena for an independent.

Crypto (crypto market making): 24/7, you run your own infrastructure, CEX/DEX fragmentation, taker fees rather than rebates, variable depth, no circuit breakers. The same quoting applies, but inventory risk runs around the clock and venue risk is real. This is the most accessible arena for an independent quoter.

Prediction markets / Polymarket (prediction-market microstructure): bounded payoffs in $[0,1]$, event-driven, thin books, and a terminal resolution that settles your inventory to 0 or 1. Inventory risk is bounded but binary; adverse selection spikes near resolution as informed traders price the outcome. Thin competition, but unique risks.

This is why the sandbox above ships a market selector: switch presets and watch the same model re-skinned to each venue, so you can see that the technique is portable and the institutions are not.

Is market making still profitable in 2026?

Yes, but the edge has bifurcated. In lit equities and listed futures, market making is a commoditised, scale-and-latency utility dominated by a few firms on razor-thin margins. The open, still-profitable frontier for an independent is crypto and prediction markets, where books are thinner and competition shallower. The structure of the problem is permanent; the easy money is not.

The honest take, dated to 2026: providing liquidity is structural (order-driven venues always need it) so market making does not "die" the way a single signal does (see alpha decay). But on the most-contested venues it has become an infrastructure business: you compete on speed, fee tiers, fair-value quality and capital, not on knowing the formula. The brand-level answer is at is HFT still profitable in 2026.

What AI changes: machine-learned fair value and toxicity classifiers sharpen your inputs (a better microprice, an earlier read on toxic flow) but the trilemma (spread vs adverse selection vs inventory) is invariant, and so are the closed-form intuitions on these pages. AI moves the operating point; it does not remove the triangle (see what AI changes for HFT). For an independent in 2026, the realistic path runs through the open venues: a clean book, a sound inventory model, a fair-value edge, and honest cost accounting: exactly the ladder these pages build, ending in the datasets and tools you would need to run it.

Worked example

A simplified market-making round on a synthetic instrument, as of a 2026 worked snapshot: mid = 100.00, tick = 0.01, no fees. You quote bid 99.99 / ask 100.01, a one-tick half-spread each side. A seller hits your bid: you buy 100 at 99.99, leaving inventory $q = +100$.

Balanced case (the clean ideal). Moments later a buyer lifts your ask: you sell 100 at 100.01, back to $q = 0$. The round-trip P&L is $(100.01 - 99.99)\times 100 = +2.00$: the full captured spread, with no price risk taken.

Inventory case. The second fill never comes, and the mid drifts to 99.96. You are long 100 at 99.99, now marked at 99.96, an unrealised $-3.00$ on inventory, which swamps the $+1.00$ half-spread you earned on the buy. The price moved against your position and the spread did not cover it. Skewing your quotes down (per the reservation price) would have made your ask more likely to lift and flattened you sooner.

Adverse-selection case. The seller who hit your bid was informed; the mid gaps to 99.90 right after. Your fill was good news for them and bad news for you: you bought at 99.99 something now worth 99.90, a $-9.00$ mark before you can react. This is why a fill is ambiguous and why the break-even spread must exceed expected adverse-selection cost (worked fully on spread vs adverse selection).

The per-fill arithmetic of the trade-off: with half-spread $\delta$ and a fill intensity that decays in $\delta$ (you fill less the wider you quote) expected spread P&L per unit time is $\delta\,\lambda(\delta)$ on each side, maximised at a finite $\delta$, not at $\delta \to \infty$. The Avellaneda–Stoikov model derives that optimum; tune it yourself in the A–S simulator.

The numbers are illustrative and synthetic; real spreads, ticks, fees and fill rates vary by venue and instrument, so check the venue spec, as of 2026. Educational only, not investment advice; no P&L is promised.