Adverse Selection in Market Making: Every Fill Is Bad News, and We Measured How Bad

Picture a money machine. Rest a bid $4 below the fair price and an ask $4 above it, and every time someone sells to you and someone else buys from you, you pocket the gap. No view on where Bitcoin is going, no forecast, just patience.

Ours ran on Coinbase BTC-USD on 3 July 2026: 0.01 BTC a quote, re-quoted every 50 order-book events, inventory capped at five fills in one direction, a small skew leaning against whatever we held, 3 ms to the exchange, and fees switched off. Every fill is worth four cents on paper, and the machine took 7,211 of them, which comes to +$288 of paper profit for the day.

Good news first: the spread came in, and we captured $305.43, a little more than the paper promised (partly because capture is measured against the mid just before each fill, and the skew moves the quotes).

The account, meanwhile, closed at −$331.37.

There were no fees to blame. There was no bug either: our ledger matches the engine to the cent, and we have checked that 181 times across 1–5 July 2026 with a largest disagreement of zero. The machine did exactly what it was built to do, and $619 disappeared anyway.

Every dollar of PnL lands in one of three boxes. Let $m_i$ be the fair price, taken as the mid, at fill $i$; $p_i$ the price we traded at; $q_i$ the size; $s_i$ the side, $+1$ for a buy and $-1$ for a sell; and $m_T$ the fair price at the close:

Spread capture is how far from the fair price you traded, in your favour, and it is the number people brag about. Inventory holding is what the price did to each fill's position afterwards, up to the close; traders call it the markout, summed over fills it is adverse selection in dollars, and it is the one you leave off the slide. The residual marks whatever is still open at the close. We ended flat, so it is zero.

This part is plain arithmetic (it holds for any strategy on any engine), so the only interesting question is what lands in each box.

Spread capture: +$305.43. Inventory holding: −$636.80. Residual: $0.00. Total: −$331.37. The box most dashboards leave out is twice the size of the one they show.

The obvious excuse is that Bitcoin trended and we were on the wrong side. It did rise on 3 July. But a trend pays one side of the book and punishes the other, and our long and short positions lost about equally, so the loss tracked the fills rather than the day.

Take the average fill and slow it down. You rest a bid. Something big enough to walk through every price level in front of you arrives, and you are filled.

At $t = 0$ you have earned 4.24 cents. One second later, 7.95 cents are gone. After ten seconds it is 8.60, after five minutes 8.85, and at the close 8.83, so you earned 4.24 and lost 8.83. Each of those is the day's total divided by 7,211, so it is an average, measured against the mid 10 ms before the fill.

Across the day, −$573.24 of the −$636.80 was gone one second after the fills. 90% of the damage, in the first second. Some of that move was already under way in the 10 ms before the fill. After five minutes the price does essentially nothing, which is why we don't read this as drift; drift builds up slowly, and this arrives all at once.

Fills on one side come in runs, because the traders hitting you are reacting to the same thing. Each fill pushes your inventory further in the direction the price is about to go against it. We sat at the five-fill limit for 6% of the day, about 83 minutes, and lost more than half the holding loss there. Per bitcoin held, a minute at the limit cost about twenty times what a minute holding a single fill did.

So the machine got mugged, again and again, in the same alley.

Stand further away, charge more per fill, and the dangerous traders will go bother someone else. We re-ran the same day at twelve widths, from $1 to $20 either side of the mid, and at first it looks like it works.

From $1 to $6, capture per fill rose six-fold, while the daily loss shrank from $1,545 at $1 to $95 at $6. That is a sixteen-times smaller loss from turning one knob.

Keep turning the knob, and at $10 the day finishes +$6.49. At $14 it finishes +$13.07.

A market maker that makes money, and all it took was one parameter. The knob works.

At $14 the strategy traded 251 times all day; at $10, 552. The uncertainty on those days is wider than the profit itself, and at $20 the day is back in the red. Call that an edge and you would have to call a few hundred coin flips an edge too.

Now look per fill.

Across $1 to $6, where there are thousands of fills behind every point, the holding loss per fill grew right alongside the capture, and the net per fill barely moved: between −3.6 and −4.6 cents at every width.

The whole sixteen-fold improvement came from fewer fills (43,322 at $1, 2,436 at $6). Wider quotes get you the same counterparties, just fewer of them. The leak is still there; you are simply running less water through it.

If you keep one section of this article, keep this one.

At $4, 78.7% of our quotes sat beyond the fifth price level, which itself sat a median $2.42 behind the best price. Routine flow never gets that far. A trade that reaches you out there has swept everything in front of you, and a trade that big is rarely noise. It may be someone who has decided the price is wrong, or a forced seller clearing a position. Either way, the fact that it reached you is information about the next second, and the next second goes against you.

The further out you stand, the stronger the filter. Our strategy placed 4.4 orders per fill at $1, 26.5 at $4 and 171 at $8, and every extra dollar of width selects your fills more tightly from the violent end of the tape.

Which is adverse selection in one line. Whoever trades against you picks your fills for you, so they are never a random sample of the market, and the pick has a direction.

One more uncomfortable fact. The capture figures above take the mid 10 ms before each fill. Take it at the fill timestamp and capture drops to 2.71 cents, because the trade that filled you has already torn up the book. Take it a second earlier and it rises to 4.89 cents. The net is −4.60 cents at every reference point we tried.

So if someone shows you a spread-capture figure without saying where they took the mid, they have shown you a choice, and probably a flattering one. The same goes for any "adverse selection costs N times your spread" rule of thumb: holding over capture is 4.3 at a $1 half-spread and 2.1 at $4. We had such a figure in an internal note and withdrew it, because the ratio moves with the quote width you picked and tells you nothing fixed about the market.

So much for knobs. This is the conclusion we stand behind.

Put adverse selection in the model before you calibrate anything. The plain Avellaneda–Stoikov setup most people start from assumes every fill earns exactly the half-spread and the people trading with you know nothing. Here the holding term is twice the capture at $4. An optimum found inside that assumption belongs to a different market, and refitting $k$ will not move it.

Use width to set exposure. It changes how many fills you take and leaves their quality alone, and you pay for it with an order-to-trade ratio that rose from 4.4 at $1 to 171 at $8.

The real work is deciding whether you are in the book when the informed trade arrives: skew against inventory before the book fills up, gate quotes on toxicity, pull a quote, and hold queue position early enough that your fills aren't only the ones that swept past everyone else. We work on these in strategy design and risk controls.

The spread is priced fine; ours earned every cent of it. It loses because being filled is itself a signal, the price answers within a second, and the standard model assumes that signal isn't there. Fix the model first, and keep the width for what it is, an exposure dial.

Full method, data and what would falsify the decomposition: Market-Maker PnL: Capture and Holding, Separated.