The Market Maker's Real Enemy: Gueant, Lehalle and Fernandez-Tapia on Inventory Risk

A market maker's job sounds like a licence to print money. You post a price to buy at 99 and a price to sell at 101. Somebody sells to you at 99, somebody else buys from you at 101, and you have made 2 for doing nothing but standing in the middle.

The catch is that those two trades do not arrive politely, one after the other. Sometimes twenty people sell to you in a row and nobody buys. Now you own a large pile of the asset, you did not want it, and if the price falls you lose far more than every spread you have ever earned. That pile is called inventory, and it is the thing that kills market makers.

Olivier Gueant, Charles-Albert Lehalle and Joaquin Fernandez-Tapia's paper is about how to price when you are afraid of your own inventory.

The problem: the spread is small and the inventory is big

The economics of market making are brutally asymmetric. Your profit per trade is tiny, a fraction of the spread. Your risk from holding inventory is not tiny at all, because a stock can move several percent in a day, which is worth hundreds of spreads.

So the entire game is: earn the spread as often as possible while never being caught holding a large position.

Avellaneda and Stoikov had already set this up beautifully as a control problem: you choose your bid and ask quotes, orders arrive at a rate that depends on how aggressive your quotes are, and you want to maximise your utility at the end of the day. It was the right framework. The difficulty was that the resulting equations were nasty, and getting from the framework to an actual number you could type into a quoting engine required approximations and numerical work.

This paper's contribution is largely about making the thing solvable, and about what becomes clear once it is.

The key idea via analogy: the market stall with a hot oven

Picture a baker running a stall who both buys bread from a supplier and sells it to the public. Bread goes stale fast. Every unsold loaf on the shelf at the end of the day is a loss.

The baker sets two prices: what she pays the supplier, and what she charges the public. Now watch how her behaviour changes with her stock level.

• Shelves empty. She is relaxed. She will pay a decent price to the supplier because she needs stock, and she will charge a full price to the public because she is in no hurry to sell. Her quotes are centred and she is happy to trade either way.
• Shelves groaning with unsold loaves. Everything shifts. She becomes stingy with suppliers, offering a low price so that fewer of them bother selling to her. And she cuts her price to the public, aggressively, because she desperately needs to move stock before it goes stale. She has skewed both her prices downward.

That skew is the whole insight. A market maker with too much inventory does not simply widen their spread. They shift the entire quote pair in the direction that unwinds the position. Long too much? Lower both bid and ask, so buyers find you attractive and sellers do not. The market maker's quotes stop being a symmetric fence around the fair price and become a steering wheel.

What this paper delivers is a set of closed-form, computable formulas for exactly where those quotes should sit as a function of your current inventory, your risk aversion, the volatility of the asset, how frequently orders arrive, and how much time is left. Instead of solving a nasty differential equation on every tick, you evaluate an expression.

The paper also establishes something practically valuable: as the trading horizon gets long, the optimal quotes settle down to a stable, time-independent rule that depends only on your inventory. That means a real quoting engine does not have to constantly recompute a time-varying policy. It can use a simple lookup: this much inventory implies these two quotes. That is the sort of result that gets a paper adopted by practitioners.

Why it mattered

• It closed the gap between theory and a working quoting engine. Avellaneda and Stoikov gave the field its framework. This paper gave the field usable formulas, plus asymptotic approximations that are accurate and fast. That is a large part of why the Gueant-Lehalle-Fernandez-Tapia model is a name that practitioners actually say out loud.
• It made inventory the star of the show. The clear message is that the market maker's central decision variable is not the width of the spread but the skew, the asymmetry driven by what you are already holding. That reframing is now standard.
• It handled inventory limits properly. Real market makers have hard position limits imposed by risk managers. The framework accommodates constrained inventory, which matters because unconstrained models cheerfully recommend positions no risk desk would ever allow.
• It generalised beyond one narrow setup. The authors handle general order arrival intensities rather than a single convenient functional form, which makes the model far easier to fit to real order flow.

The honest limitations

• The traders on the other side are dumb. In the model, orders arrive randomly at a rate that depends only on your quoted distance from the fair price. Nobody is trading against you because they know something you do not. Real market makers get run over by informed traders constantly, and this model has nothing to say about that. Glosten and Milgrom's adverse selection problem is simply absent.
• The mid-price wanders aimlessly. The reference price is a random walk with no memory, no momentum, and no reaction to order flow. In reality, a surge of buying tells you the price is likely to go up, which is precisely the information a market maker needs to avoid being picked off.
• You are alone. There are no competing market makers in the model, no queue you have to stand in, no other liquidity provider undercutting you by a tick. Real electronic markets are queue-priority races.
• Risk aversion is a knob nobody knows how to set. The formulas depend on a risk-aversion parameter that governs how nervously you unwind inventory. There is no principled way to pick it, and it materially changes the quotes.
• The clean formulas rely on approximations. The genuinely closed-form results come from asymptotic limits. They are excellent approximations in the regimes studied, but they are approximations.

The one-line takeaway

Gueant, Lehalle and Fernandez-Tapia showed that a market maker's optimal behaviour is not to sit symmetrically around the fair price but to skew both quotes in whatever direction unwinds their inventory, and they turned that principle into fast closed-form formulas that a real quoting engine can evaluate on every tick.