Every market has a fastest layer, and the firms that own it are paid in microseconds. Below it sits a second layer, defined not by speed but by the willingness to hold risk across thousands of milliseconds. Binance altcoin perpetuals are a clean specimen of that second layer: over 45 days to 19 November 2026, twelve USD-M contracts carrying between 4.6 and 304 million dollars of daily turnover moved 10 to 19 basis points in a typical minute, quoted roughly two basis points of spread, and cost five basis points of taker fee on each side.
The consequence is arithmetic before it is strategy. A round trip on this venue costs:
| Entry | Exit | Round trip |
|---|---|---|
| maker | maker | 4.0 bps |
| maker | taker | 7.0 bps |
| taker | taker | 10.0 bps |
Everything below is measured against that band. The data is the Binance public archive plus the public futures data endpoints: thirteen USD-M perpetuals, one-minute klines, five-minute metrics and funding history, 6 October to 19 November 2026, data cut-off 20 November 2026 07:45 UTC. Sample sizes are printed next to every number because a median without an n is a rumour.
1. The liquidation vacuum is priced by severity, not frequency
A liquidation is not a decision. It is a market order that has to happen, sent by an engine that does not care about price. On a book that is one increment wide, an order larger than five levels of resting size does not move the price, it removes the book. The sellers who would normally refill it are the same ones who step away when inventory risk spikes, which is exactly why the vacuum persists and exactly why its closing is the reversion.
For each symbol, take one-minute returns, compute the standard deviation, and mark every minute whose return is below minus three (or five) sigma with turnover at least five times the symbol's median minute. Then measure the return from the close of that minute forward. No filters, no discretion, no fitting.
| Shock | Events | +3 min | +5 min | +15 min | Win 5m |
|---|---|---|---|---|---|
| -3 sigma | 5,235 | +4.0 | +7.6 | +8.7 | 55.4% |
| -5 sigma | 1,007 | +9.9 | +21.2 | +26.4 | 60.5% |
At three sigma the median recovery is 4.0 basis points, which is a maker round trip and nothing else. At five sigma it is 9.9 at three minutes, 21.2 at five. The same phenomenon, measured at two thresholds, is either break-even or a trade, and the difference is severity.
Per contract, at five sigma:
| Contract | Sigma (bps) | Events | +1 min | +3 min | +5 min | +15 min | Win 5m |
|---|---|---|---|---|---|---|---|
| 1000PEPE | 16.4 | 105 | +18.5 | +19.9 | +33.0 | +26.8 | 65% |
| ENA | 17.0 | 58 | +13.3 | +19.7 | +32.8 | +34.7 | 64% |
| TIA | 18.9 | 73 | +7.5 | +14.9 | +23.5 | +40.1 | 58% |
| SUI | 11.6 | 76 | +14.8 | +12.1 | +23.4 | +30.9 | 66% |
| APT | 14.3 | 87 | +16.4 | +11.7 | +23.1 | +40.4 | 63% |
| 1000FLOKI | 11.2 | 89 | +8.8 | +7.9 | +20.6 | +20.5 | 60% |
| WIF | 12.1 | 81 | +6.0 | +13.6 | +19.9 | +29.7 | 60% |
| ORDI | 11.4 | 93 | +5.5 | +13.0 | +19.3 | +22.2 | 57% |
| ARB | 14.0 | 68 | +6.5 | +9.7 | +17.5 | +23.9 | 63% |
| DOGE | 9.0 | 111 | +2.8 | +5.9 | +16.0 | +20.8 | 59% |
| OP | 14.5 | 70 | -0.4 | +3.8 | +15.1 | +30.8 | 54% |
| SEI | 11.8 | 69 | +8.0 | -4.0 | +8.1 | +15.7 | 57% |
Two details are worth more than the averages. The recovery scales with the symbol's own volatility rather than its liquidity: DOGE, the most liquid alt in the set, recovers 16 basis points while PEPE recovers 33. And the one-minute column is erratic, negative for OP and barely positive for DOGE and ORDI, so the reversion is a process of several minutes, not a bounce to scalp.
The distribution is the strategy. Pooled at five minutes, the 10th percentile is minus 124 basis points, the 25th is minus 45, the median plus 21, the 75th plus 65, the 90th plus 107, and the mean plus 4.4. A mean that is a fifth of the median means a small number of events are not reversions at all but the first leg of something larger. A price stop placed 30 basis points below entry is hit by about a third of the winners. A time stop at fifteen minutes cuts the tail without cutting the winners.
The trigger, as a program sees it:
import collections, statistics WIN = 7 * 24 * 60 # trailing sigma window, minutes BUF = collections.deque(maxlen=WIN + 1) def on_minute(close, quote_volume, hist_vols): BUF.append(close) if len(BUF) < 240: return None rets = [(BUF[i] / BUF[i - 1] - 1) for i in range(1, len(BUF))] sd = statistics.pstdev(rets) r = rets[-1] med_vol = statistics.median(hist_vols) if sd == 0: return None z = r / sd surge = quote_volume / med_vol if med_vol else 0 if z <= -5.0 and surge >= 5.0: return {"z": round(z, 2), "surge": round(surge, 1), "sd_bps": round(sd * 1e4, 1)} return None Note what changed from the measurement to the code: sigma is trailing, not computed over the whole window. Using the full sample to set the threshold is a look-ahead bias and it is not available live, which is the first place a backtest stops being a backtest.
2. Bitcoin's second leg is a filter, not a lag
When Bitcoin moves, alts follow inside the same minute. Fifteen-minute buckets containing a 0.75 per cent Bitcoin move show the alt moving 120 to 220 basis points in the same bucket, so the folk version of the trade, buying the alt that has not moved yet, is a race nobody on a home connection wins. What is measurable is the next bucket, and it is conditioned by two public numbers.
Thirty-eight Bitcoin moves of 0.75 per cent or more in 45 days produced 452 alt observations. Open interest counts as contracting below minus 0.4 per cent over the bucket, expanding above plus 0.4. Taker flow counts as agreeing when the buy/sell ratio is at least 1.15 on an up move or at most 0.87 on a down move.
| Condition | N | Same bucket | Next 15 min | Positive |
|---|---|---|---|---|
| OI down, flow with move | 53 | +213 bps | +36 bps | 70% |
| OI down, flow against | 114 | +224 bps | -41 bps | 41% |
| OI flat, flow with | 90 | +122 bps | -10 bps | 40% |
| OI flat, flow against | 121 | +127 bps | -8 bps | 43% |
| OI up, flow with move | 26 | +190 bps | +0 bps | 50% |
| OI up, flow against | 48 | +162 bps | -6 bps | 44% |
The result inverts the usual story about confirmation. The cell a momentum trader would call confirmed, open interest expanding with the move, produced a median of minus 6 basis points and 44 per cent positive. The cell where open interest fell while the alt moved with Bitcoin and taker flow agreed produced plus 36 basis points and 70 per cent positive.
The mechanism is mechanical. A move with falling open interest is a move in which somebody was forced out, so the supply above the market has been consumed. A move with rising open interest leaves a crowd of positions opened on the other side, now underwater, waiting to be let out at the price they entered. The trade is the move paid for by people closing positions; the trap is the move paid for by people opening them.
Tightening the threshold to one per cent sharpens it to a median plus 70 basis points with 79 per cent of 34 observations positive, but 34 observations drawn from 18 underlying events are not 34 independent facts, and a filter that only speaks once every three days is a different product from one you can schedule.
3. The spread on most alt books is a specification, not a market
Across 525 trading USDT spot pairs the median minimum price increment is worth 6.6 basis points of price. Fifty-seven per cent have an increment above 5 basis points and 36 per cent above 10. PEPE quotes in increments worth 27.8 basis points.
Inside the liquid band it gets stranger: of 28 books carrying 10 to 150 million dollars of daily turnover, 82.1 per cent quoted exactly one increment wide, the median spread was 1.62 basis points, and the median top-five-level depth was 90,176 dollars. On those symbols a maker cannot improve on a price the exchange has defined, take-profit and stop distances quantise to the same lattice, and the only decision left is where on that lattice to sit.
That single fact decides whether a symbol can be traded as a market or only as an event. On a 27.8 basis point grid, a two-basis-point edge does not exist and a liquidation vacuum is not a reversion, it is a single step in a staircase.
4. Funding is a clock, not an income
Across 4,692 settlements the mean funding rate was 0.33 basis points per print, 3.6 per cent annualised. The median was 0.5, the 95th percentile 1.0, the largest single print in the sample was minus 21.6 basis points on 1000PEPE, 78.7 per cent of prints were positive, and no print reached 30 basis points.
The popular trade is the pre-settlement skew, and it did not appear. Over 1,233 settlements the three-minute move before the print averaged minus 1.67 basis points with a t-statistic of minus 1.74, and the median absolute pre-print move of 10.9 basis points sits inside the 10.5 to 15.0 range of an ordinary three-minute window in the same symbols. In the top decile of funding magnitude, 1.70 basis points per print, the aligned pre-print drift was 2.26 basis points with a t-statistic of 1.31 and 52.8 per cent of prints followed by a positive move.
What survives is a decision rule rather than a story:
- tradable window: at least 7.0 basis points per print with two consecutive prints above it;
- carry candidate: annualised rate above twice the spot financing cost;
- otherwise: the clock is only a timestamp for one of the other playbooks.
5. Execution is the strategy
Four of the five structures above are decided by how the order is sent, not by when. The same signal is a 10 basis point trade as a taker and a 4 basis point trade as a maker, and on a tick grid coarser than the fee, five of the six decisive variables are execution variables: maker or taker entry, quote lifetime, what happens on a partial fill, stop against timeout, and how much of the order-rate budget is left.
That budget is finite and published: 100 orders per ten seconds, 200,000 per day per account, 200 open orders per symbol. A strategy that re-quotes four symbols every 500 milliseconds spends the daily allowance in about seven hours. The stop that ends most medium-frequency attempts on this venue is not the alpha, it is the rate limiter.
What this leaves
Five structures, one unit of account, and a cost band that decides which of them exists on any given day. Expand each into its trigger, its entry, its exit and its failure modes and you have a book rather than a blog post.
The full version is 31 pages: nine chapters, three appendices, seven figures, working Python for each trigger, and the failure modes listed in the chapter where they bite. It is public-data only, no keys, no order path: The Sub-Institutional Edge - Binance Altcoin Playbook, 49 USD, instant PDF.
Free, and better read before the paid thing: the lead-lag paradox tear sheet and the 73-hour gold gap tear sheet, both three pages, and venue-latency-probe, a public probe with no keys and no order path.
Measured numbers describe a window that has closed. The structures they describe, the increment, the funding clock, the liquidation engine and the fee, change far more slowly. If one of these numbers does not reproduce on your own capture, send the capture and the date and I will look at it; that is the only review process this work has, and it is a better one than a comment score.
![UI Practice - Duolingo [1]](https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2F8uawvr6a0e07ilktpkk7.png)

