- Every trading strategy pitch eventually meets the question "how big can this get?" For stablecoin dislocation trading, no systematic public answer existed. We built one from public tick data: 1,136 dislocation events on Binance USDC/USDT over 24 months.
- Capacity is a distribution, not a number: the dollar volume flowing through a dislocation spans three orders of magnitude — from under $100K in a quiet-day wobble to $1.6 billion in the October 2025 depeg.
- The counter-intuitive finding: the valuable events are slow. Capacity-weighted, roughly 99% of dislocation volume flows through episodes lasting minutes to hours — not the millisecond wicks that dominate trading folklore.
Ask an allocator what kills niche strategies and you will hear the same word every time: capacity. A strategy earning 40% on one million dollars and 4% on thirty million is not one strategy — it is two, and only one of them is investable at size. For stablecoin dislocation trading (buying the transient discount when USDC/USDT breaks from parity, earning the reversion), the capacity question has a special edge: the opportunity only exists for minutes at a time, so the relevant measure isn't average daily volume — it is how much money flows through the market during the dislocation itself. We searched the academic literature, the data-vendor research desks and the market-maker blogs for a systematic measurement of that number. We did not find one. So we measured it.
Where the literature stops: the Bank of England's post-mortem of the 2016 sterling flash crash comes closest in method — mapping observed order flow to price impact. The NBER's 2025 study of stablecoin runs quantifies the primary-market arbitrage channel (finding that Tether redemptions concentrate in roughly six entities per month). Kaiko has published valuable depth snapshots of individual depeg days. And at least one practitioner essay has described the strategy in narrative form. None of them answer the operative question: event by event, how much capital could the dislocation have absorbed?
1Define the event before counting it
A dislocation is a mechanical definition, not a headline: price breaking ±2bp away from a slow, volatility-aware reference level, gaps under 10 minutes merged.
Show the working
The raw material is public: exchange-published tick-level trade data for the deepest USDC/USDT order book, August 2024 through August 2026. The reference level is a slow, volatility-aware measure of where the pair has recently traded — constructed to sit still through a dislocation while tracking the pair's genuine drift. An event begins when trades print beyond ±2 basis points of that reference and ends when no such print has occurred for 10 minutes. We deliberately leave the reference's exact construction unspecified — it is part of our working toolkit — but the census is robust to it: any reasonable slow anchor reproduces the counts within a few percent, and moving the threshold from 2bp to 3bp drops the count from 1,136 to 295 without changing any conclusion below.
Two basis points sounds absurdly small to anyone from equities. On this pair it is not: the modal spread is a fraction of a basis point, fees on the pair are zero, and the reference level barely moves — a 2bp print is roughly a six-sigma excursion against quiet-market volatility. It is also, empirically, the threshold at which reversion becomes economically meaningful.
2The census: dislocations are not rare events
577 events per year — one every 0.6 days. This is not a "wait years for the crisis" trade; it is a permanent weather system with occasional hurricanes.
Show the working
The headline counts, across 24 months of tick data:
- 1,136 events at the ≥2bp threshold — 577 per annum, a median gap of well under a day even through the quietest summer months.
- Duration: median 113 seconds; 64% of events last a minute or longer. The distribution is heavy-tailed: the largest events run for hours.
- Depth (at the ≥3bp census tier): median peak excursion 3.6bp, 90th percentile 6.0bp — and a maximum of 207bp. Depth, too, is a fat-tailed distribution wearing a calm median as a disguise.
- Direction: broadly symmetric over the full sample (161 upward / 134 downward at the 3bp tier), but strongly regime-dependent — recent quarters have skewed heavily to upward spikes.
A strategy that triggers 577 times a year generates enough independent observations to be graded statistically within months, not decades. Contrast the classic tail-risk fund problem — a strategy that pays off twice a decade can't be distinguished from luck inside a career. Frequency is what makes this niche auditable.
3Capacity: measure the flood, not the book
The capacity yardstick is the dollar volume that actually traded while price was dislocated — the "flood" — and a participation discipline of 10–20% of it.
Show the working
Resting order-book depth is the wrong yardstick for event capacity: during a dislocation the book is being torn up and rebuilt every second, and what a passive participant can capture is bounded not by any snapshot but by the total flow that transacts at dislocated prices. We call this the event's flood. Absorbing more than a small fraction of the flood means you are the market — at which point the dislocation shallows out and the measured opportunity no longer exists. We apply a 10–20% participation ceiling, consistent with standard execution practice.
The flood distribution is the article's central exhibit:
| Event tier | Flood (volume through the dislocation) | Deployable at ~15% participation |
|---|---|---|
| P25 — quiet-day wobble | ~$80K | ~$12K |
| Median event | ~$800K | ~$120K |
| P75 — a real dislocation | ~$6.7M | ~$1M |
| P95 | ~$35M | ~$5M |
| October 2025 depeg (the maximum) | ~$1.6B | $160–325M |
Three orders of magnitude separate the quartiles from the tail. This is why any single "strategy capacity: $X" claim for an event-driven book should be treated as a red flag in due diligence: capacity here is a per-event distribution, and capital sized to the tail sits mostly idle through the body. The only coherent capital structure is tiered — a small always-on layer sized to the median event, and standby capital that only meets its match in the top quartile.
4The finding: capacity and depth are different animals
The deepest events are not the biggest, and the biggest are not fast. The money is in slow dislocations; millisecond wicks are folklore, not flow.
Show the working
Rank the 295 larger events by flood and a pattern appears that we did not expect when we started counting. The top fifteen floods — $37M to $1.6B each — are without exception slow events: durations from 98 seconds to over three hours, mostly at moderate depths of 4–17bp. Meanwhile the deepest event in the census (a 207bp spike in May 2026) ranks 86th by flood: spectacular on a chart, nearly empty of volume.
At the extreme short end, we cross-checked the census against our own millisecond-resolution order-book recording of an August 2026 spike: a +5.7bp excursion whose deepest tradeable window lasted 2.8 milliseconds. No reaction-based system on earth participates in that window — and it doesn't matter, because the volume transacted inside such windows rounds to nothing next to the slow floods.
Weight events by flood rather than counting them, and roughly 99% of all dislocation dollars flow through episodes lasting a minute or longer — timescales at which a well-built system reacting in under a second, or even a human with an alert, is not late. The engineering folklore of this niche — colocation, microwave links, nanosecond obsession — optimises for the 1% of the money. The unglamorous problems (being funded and present at 3am, borrowing capacity at the moment of stress, automated capital mobilisation) govern the 99%.
One regime caveat belongs here: an April 2026 change in the venue's price-increment granularity visibly reshaped the microstructure — recent dislocations run deeper and thinner than the 2024–25 vintage. The census's older events are a depth-distribution playbook, not an execution blueprint; whether billion-dollar floods return in the next active season is an open measurement, not an assumption.
5What the distribution does to capital and leverage
Leverage is a tail amplifier, not a general accelerator: in the body of the distribution the participation ceiling binds first, and extra leverage adds nothing but risk.
Show the working
Push a hypothetical capital stack through all 1,136 events, respecting the participation ceiling event by event, and a clean result falls out. For the bottom three quartiles of events, modest capital already saturates the 15% ceiling — doubling the leverage on those events changes returns by zero, because the constraint is the flood, not the balance sheet. Only in the fat tail — floods above a few million dollars — does additional balance-sheet capacity keep buying additional participation all the way up.
The design consequence is the inverse of the usual retail instinct. Instead of "maximum leverage always, and hope", the distribution rewards: conservative leverage as the standing posture, with pre-arranged capacity to scale into the rare event that can actually absorb it — governed by hard, mechanical risk lines (in our own testbed work: a cross-venue index co-movement check that distinguishes a genuine multi-venue depeg, where liquidation indices actually move, from a single-venue wick, where they don't). The stress-test discipline for the standing posture came from replaying the strategy through the worst historical episode at each leverage tier and requiring the worst observed adverse excursion to clear the liquidation buffer with margin to spare — the same worst-case-replay logic we described in the backtest validation case study.
6Honest limits — what a census cannot tell you
The flood is an upper-bound proxy, self-impact is unmeasured by construction, and the only final arbiter is live execution.
Show the working
Three limits are worth stating as plainly as the findings:
- The flood counts everyone. It is total transacted volume, including the panicked flow you would be trading against and the competitors you would be queueing with. The 10–20% ceiling is a discipline, not a guarantee — actual attainable share depends on queue position and quote quality.
- Self-impact is invisible to replay. No historical dataset can measure how the dislocation would have changed because you were there. Every capacity study, including this one, measures the market without you in it.
- Fill models need calibration, not faith. In our queue-position replay work on full order-book (L2) data, real fill counts landed well below the optimistic "front-of-queue" model — a reminder that uncalibrated backtest fills overstate event participation by a wide margin. The calibration factor itself, like the reference-level construction, stays in the toolkit. We treat replay as a ranking tool and live execution as the measurement of record.
What this changes in practice: "how big can this get?" now has a structured answer for at least one event-driven niche — a reproducible event definition, a per-event capacity distribution, and a capacity-weighted decomposition showing where the money actually flows. The same census method transfers to any episodic-liquidity strategy: define the event mechanically, measure the flood per event, apply a participation ceiling, and read capacity as a distribution.
The reusable checklist
Before accepting a capacity claim for any event-driven strategy:
- Is the event defined mechanically and reproducibly — or does "opportunity" mean whatever the pitch deck needed it to mean?
- Is capacity quoted as one number? Event-driven capacity is a per-event distribution; ask for the quartiles, not the average.
- Is the yardstick resting depth or transacted flow? Snapshots of a book being torn up measure the wrong thing.
- What participation share of event volume does the claim assume — and would that share leave the opportunity intact?
- Does leverage in the model add return in the body of the distribution (a red flag — the constraint should bind first) or only in the tail?
- Are backtest fills calibrated against any ground truth — and by how much did calibration shrink them?
Where this fits
This census is part of the same discipline MOA applies to enterprise claims under Independent Verification & Validation: when a load-bearing number has never been measured, measure it before capital relies on it. The analysis uses public exchange data throughout; derived event tables and replay tooling are being prepared for open release. We review methodology and evidence; we don't sell or manage strategies, and nothing here is investment advice or a recommendation to trade any instrument.
- Noss et al., The October 2016 Sterling Flash Episode, Bank of England Staff Working Paper (2017) — the closest methodological relative: order flow mapped to price impact during a flash event.
- Ma, Zeng & Zhang, Stablecoin Runs and the Centralization of Arbitrage, NBER WP 33882 (2025) — the primary-market side: depeg arbitrage via issuer redemption concentrates in a handful of entities.
- Kaiko Research, Defining Depegs (2023) — depeg severity metrics from OHLCV data; depth snapshots of individual depeg days in related notes.
- BIS Markets Committee, The Sterling Flash Event of 7 October 2016 (2017).
- A practitioner narrative of the underlying strategy appeared at Papers With Backtest (2026) — engaging as description, but containing no event census, capacity measurement or calibrated fill model; the gap between narrative and measurement is precisely what this article addresses.