A quote is posted. An aggressive trade hits it. The price continues in the same direction. The market maker is left holding inventory just as the market moves away.
That is adverse selection. It is not just a bad fill. It is a bad fill followed by evidence that the other side knew more than the liquidity provider at that moment.
The findings here are based on our SSRN working paper, Predicting Adverse Selection in High-Frequency Cryptocurrency Markets Using Gradient Boosting, which studied BTC/USDT perpetual futures on Bybit using 31,081,463 second-level observations from February 1, 2025 to February 16, 2026, with 360 out-of-sample evaluation days in a strict walk-forward framework.
"The practical edge is not in predicting every tick. It is in knowing when normal quoting behavior has become unsafe."
The dangerous seconds were defined, not guessed.
A second was treated as toxic only when strong directional order flow was followed by price continuation over the next five seconds. In simple terms: buyers hit, price kept rising — or sellers hit, price kept falling.
The label also used adaptive rolling thresholds over a one-hour lookback window. That mattered because toxicity was not stable across the sample. Monthly toxicity rates ranged from 0.081% to 0.795%, nearly a tenfold variation. A fixed threshold would have mixed up volatility regime changes with real toxicity.
Probability alone was not enough.
The model did not only ask whether a second was likely to be toxic. It also asked how large the move could be if toxicity occurred.
That is why the framework used a composite TailScore:
TailScore = predicted toxicity probability × predicted 99th-percentile absolute price move
For a market maker, this distinction is important. A small toxic move is annoying. A large toxic move is what damages the book.
The worst seconds were concentrated far better than random.
At the tightest gate — the top 0.1% of seconds by TailScore — the model reduced CVaR99 by 2.33%. A random gate of the same size would reduce CVaR99 by only 0.090%.
That means the TailScore was 25.85× more efficient than random at identifying the seconds that mattered most for tail risk.
This is the key interpretation: the system was not trying to predict everything. It was trying to isolate the short windows in which passive liquidity becomes structurally dangerous.
The large moves were not captured by classification alone.
The paper also compared the TailScore with two weaker alternatives: a pure classifier and a VPIN-style proxy.
The difference was large. At the 0.1% gate, the pure classifier achieved 2.11× efficiency, while the TailScore achieved 25.85×. The VPIN proxy reached only 0.30×, below the random benchmark.
The practical takeaway is simple: for a market maker, knowing that a second may be toxic is not enough. What matters is whether the toxic second is large enough to justify changing trading posture.
- Reduce quote size during short toxic windows
- Slow replenishment instead of immediately refreshing stale quotes
- Widen spreads within allowed bands when tail risk concentrates
- Protect inventory when the next few seconds are likely to be structurally adverse
The edge held across all 13 months.
The chart below shows monthly CVaR99 efficiency for TailScore on BTC/USDT. Every month remained well above the random benchmark. Even the weakest month stayed above 12× efficiency, while the strongest crossed 53×.
This is a posture problem, not a prediction obsession.
Market makers do not need a warning on every second. They need a warning when the next few seconds are unusually dangerous for passive liquidity.
That is where a microstructure risk overlay becomes useful. The job is not to replace the strategy. The job is to keep the strategy from behaving normally during abnormal seconds.
Adverse selection is not evenly distributed. It clusters in rare seconds where order flow, volatility, book state and continuation align. For market makers, those are the moments when spread capture disappears and inventory risk expands. The practical edge is not in predicting every tick. It is in monitoring the few seconds that matter most.
This note is based on our SSRN working paper: Predicting Adverse Selection in High-Frequency Cryptocurrency Markets Using Gradient Boosting by Suresh Rajendran and Divya Singaravelu. Read the paper here: https://papers.ssrn.com/sol3/papers.cfm?abstract_id=6344338