Research · market microstructure
Published · SSRNAdverse Selection Consumes the Touch
A true-aggressor-signed maker-P&L decomposition across 13 CME futures
A market maker posts a bid and an ask and waits. It earns the spread for standing ready, and it pays adverse selection to the informed traders who pick off its quotes just before the price moves. This paper measures, fill by fill, on futures data that carries the exchange's true buyer-or-seller tag on every trade, how large that cost is against the spread actually captured. The answer is that adverse selection consumes almost the entire captured spread, and the pieces below let you rebuild the mechanism in the browser.
Data partner
The measurement rests on true-aggressor-signed CME futures order books from AlgoSeek, accessed under a sanctioned developer relationship. The ground-truth sign on every one of 12.4 million trades is the ingredient that lets the study measure adverse selection directly instead of guessing the trade side.
View on SSRN
ssrn.com/abstract=7022599
Replication code & engine
github.com/DaruFinance/quant-mm-simulator
The result in three lines
13 roots · 353,387 fills
Adverse selection eats the touch
The half-spread captured at the touch is +0.654 bp; the post-fill adverse drift is −0.661 bp. Net maker P&L after the fee is −0.022 bp, statistically indistinguishable from zero, and every root lands within ±0.13 bp.
Proposition 1
A ceiling, not a point
Passive fills are charged no slippage, so the accounting is the friendliest possible to the maker. Every real friction left out pushes the number more negative. The honest claim is a one-sided bound: net maker P&L is at most about zero.
two corollary nulls
No quote, no signal, beats it
An Avellaneda–Stoikov inventory-skew quoting family is null net of the fee (0 of 234 cells significant). A cross-root order-flow signal validates in-sample at R² = 0.545 but adds nothing out-of-sample and turns negative at full cost.
Overview
Glosten and Milgrom argued forty years ago that the spread a market maker earns is compensation for the adverse selection it pays. The prediction is textbook; the direct, large-scale, true-sign measurement of how the two sizes compare was missing. Most futures microstructure work has to guess who the aggressor was on each trade, because the exchange sign is not in the ordinary feed. This study supplies the measurement on data that carries the real sign, so adverse selection is measured rather than inferred.
One event-driven simulator replays the real book and trade tape, quotes at the touch, and marks each passive fill out at 10 and 60 seconds. Splitting each fill into the half-spread captured and the drift the counterparty takes back gives a fill-by-fill decomposition across 13 contract families and 234 contract-days. The captured spread and the adverse drift move together, root by root, exactly as the theory says: they are the same information showing up twice, once as the price of liquidity and once as its cost.
The measurement
How a passive fill is scored
Every result flows from one decomposition. When a resting order fills, the maker captures the distance between the mid and its limit price, which is the half-spread and is always non-negative. Over the next seconds the mid drifts, and because the counterparty tended to know something, it drifts against the maker. Marking the position out at a fixed horizon turns that drift into a number. Net P&L is the captured spread plus the adverse drift, less the fee.
Per fill: ε is +1 for a passive buy and −1 for a passive sell, m is the mid, p the limit price, and τ the markout horizon. Capture is fixed at the fill; adverse drift accrues over τ.
Net P&L nets the fee f (0.015 bp on CME, no rebate). The whole paper is the observation that capture ≈ −adverse, so net rests at zero.
There is no look-ahead in the machinery: a fill at a given instant can only see data at or before that instant. The fill rule is deliberately optimistic, filling the maker as soon as the queue ahead clears, which can only add fills and therefore add adverse selection, never invent profit. The fast kernel is verified bit-for-bit against a slow reference on every root-day.
One fill, two measurements
The maker earns the gap between the mid and its quote the instant it fills, then watches the mid walk away over the next 10 to 60 seconds. The first gap is the pay; the second is the cost.
Interactive · the markout laboratory
The captured half-spread is fixed the moment you fill. The adverse drift grows with the horizon and with how informed the flow is. Drag the adverse intensity and watch the captured spread get eaten. Real futures sit near 1×, where the two halves nearly cancel.
near the balance: capture and adverse nearly cancel, net rests at zero
capture
+0.651
bp
adverse 10s
-0.656
bp
net 10s
-0.020
bp
net 60s
-0.149
bp
x: markout horizon 0 to 60s · y: bp per fill
The two halves cancel
Wide-spread roots capture the most and pay the most: natural gas earns +2.701 bp and gives back −2.808 bp, and the two nearly annihilate. The residual is the whole story, and it is tiny. The net after the fee ranges from +0.083 bp on the 10-year note to −0.122 bp on natural gas, with the fill-weighted average at −0.022 bp. That co-movement of capture and adverse is the Glosten–Milgrom mechanism, measured at scale.
The reading
A competitive floor, and one derived condition
Write the per-fill net as capture minus adverse minus fee. A rational maker only quotes while the spread it can capture clears the adverse drift plus the fee. Free entry then bids the spread down until the two meet and profit vanishes. The measured near-zero is what that zero-profit condition of a competitive liquidity business looks like from the outside. The paper is careful to hold this as consistent with competition, not as a proof of it.
Free entry pins the equilibrium half-spread s* at the adverse drift a plus the fee f, so per-fill net is driven to zero.
One condition is genuinely derived rather than described. A book cannot quote a half-spread tighter than half a tick, so a coarse tick puts a hard floor under the captured spread. If that floor already clears the adverse drift plus the fee, the maker is guaranteed to net a profit. Across the 13 roots the guarantee holds only for the 10-year note and silver. Other roots turn a profit anyway by capturing more than the bare floor, which is why the frontier is descriptive: the tick floor is sufficient, not necessary.
The sufficient survival condition: if half a tick δ already clears adverse plus fee, net is non-negative regardless of anything else.
Interactive · the tick-size floor
Because a maker cannot quote inside one tick, half a tick is the least it can capture. Set a tick width and an adverse drift, or load a real root, and see whether the floor alone clears the cost. Only ZN and SI pass this sufficient test, though several others net positive anyway.
0.707 ≥ 0.647
floor clears the cost: net guaranteed ≥ 0
measured net +0.083 bp
Only coarse ticks survive
The survivors are all coarse-tick: the five-year and ten-year notes, crude, silver, and the bond. The fine-tick equity indices, where the book can quote almost continuously, have no structural floor to protect the spread and go negative. The 0.74 bp crossover is a description of where the roots happen to sort, not a law, and natural gas proves it: the widest tick on the board, and still the largest loss.
Across regimes
Does it hold in a crisis?
The headline sample is one calm 2023 window, with July carrying most of the fills, so the cancellation is established across contracts within a single volatility regime. Whether it survives a crisis needs data from other years, and that is where a subtle trap sits. The obvious move, comparing the calm depth feed against a legacy per-year feed, is confounded: the two reconstruct the book differently. On the identical July-2023 sessions they read 99.5% and 82.1%, a 17.4 point gap from the feed alone, with dates and roots held fixed.
The paper resolves it by measuring every regime on one nanosecond true-signed tape that spans all the years under a single reconstruction, verified to agree with the depth feed to 0.3 points on matched calm sessions. On that feed the cancellation does break under stress. Adverse selection over-consumes the captured spread in the SVB, FOMC, and COVID weeks, and net deepens with volatility. A continuous 2016 to 2026 sweep of 33,956 root-days then sharpens it: the volatility sensitivity is carried by natural gas, copper, crude, and FX, and vanishes when those are dropped. So this is matched-feed evidence of regime dependence, not a universal law.
Interactive · regime and the feed confound
Pick a regime and read the adverse-to-captured ratio on the matched feed; above 100% adverse selection over-consumes the spread. The right panel is why the comparison must hold the feed fixed: the same calm dates read very differently on the two reconstructions.
adverse ÷ captured
109.8%
deepest measured loss
net after fee -0.139 bp
same dates, two feeds
Identical July-2023 sessions, roots and regime held fixed. The 17.4 point gap is the reconstruction, not the market, which is why every regime is measured on one feed.
Stress tips the balance
In calm markets the ratio hugs 100% and the maker breaks even. In the crisis weeks it runs to 105, 110, and 112%, the cost overruns the pay, and net turns clearly negative. The effect is broad across roots in each stressed week, but the way the loss scales with volatility is concentrated in a handful of energy, metal, and FX contracts rather than being a property of every market.
The foundation
Why the true sign matters
The whole measurement rests on knowing who was the aggressor on each trade. Because AlgoSeek carries the exchange's real sign, the study can grade the classifiers everyone else relies on. The standard tick rule is weak on coarse-tick futures, from 0.607 on the bond to 0.834 on the Nasdaq, far below the roughly 0.90 that holds in equities. The quote rule is strong, between 0.917 and 0.993. Any futures microstructure result built on tick-rule signs inherits a 17 to 39% sign error on the worst roots, which is exactly the error the ground-truth sign removes here.
Two nulls follow from the balance
If the captured spread and the adverse drift are co-determined, no cleverness can separate them, so any strategy that tries should come up empty. Both do. An Avellaneda–Stoikov inventory-skew quoting family, swept into a grid of more than a thousand cells across the 13 roots, is null net of the fee: the inventory-skew move clears zero of 234 cells even before any multiplicity correction, and the probability of backtest overfitting sits at 0.543, the coin-flip value of a true null. Standard inventory-aware quoting does not beat quoting passively at the touch.
The second null asks whether one market's order flow leads another's. A cross-root integrated order-flow-imbalance signal validates its pipeline in-sample, with own contemporaneous flow explaining returns at R² = 0.545, which reproduces the known result and proves the code is right. But the predictive part fails: the out-of-sample incremental R² is −0.005, positive in only 1 of 17 windows, and cost buries whatever thin content remains. At zero cost the signal earns +0.091 bp; at full spread cost it is −1.30 bp. The lagged cross-market signal does not survive trading costs.
The same cost, in DeFi clothes
The adverse drift a futures maker pays is, at the role level, the cost the automated-market-maker literature calls loss-versus-rebalancing: an informed counterparty picking off a passive liquidity supplier. The paper puts the two on one axis only as a consistency check. They do not coincide in level, and both scaling with volatility is equally consistent with one shared cost or two unrelated ones, so the shared σ-scaling is reported as consistent-with, not as proof.
The AMM primitive: loss-versus-rebalancing scales as the variance rate, the same volatility dependence the maker's adverse drift shows.
Method
- One event-driven simulator replays the real depth-10 book and true-signed trade tape, quotes at the touch, and tracks queue position so a fill only lands once the orders ahead are consumed.
- Each passive fill is marked out at 10 and 60 seconds and split into captured half-spread and post-fill adverse drift, with the maker charged the 0.015 bp CME fee and no slippage, which makes the net a one-sided bound.
- Significance is led by the probability of backtest overfitting via combinatorially-symmetric cross-validation, with permutation nulls and pre-registered hypotheses hash-stamped before any number was computed.
- The cross-regime work holds the book reconstruction fixed on one nanosecond true-signed feed, after a same-dates control showed a 17.4 point gap between feeds on identical sessions.
- The fast kernel is verified bit-for-bit against a slow reference; a fill can only see data at or before its own instant, so there is no look-ahead in the machinery.
Reproducibility
The simulator, cost model, and significance toolkit are collected in a companion repository, quant-mm-simulator, which ships correctness tests on small bundled fixtures in the provider's own format plus the committed per-run result files the figures render from. The licensed tapes are not redistributable, so the same engines rebuild everything from those fixtures and results. The interactive figures on this page read the paper's reported tables directly.
Cite
See also
The companion study on decentralized-exchange liquidity, No Edge Without Information, carries the same adverse-selection cost into on-chain markets, and the narrative note The edge is in the process develops the selection discipline that runs through the nulls here.