Was the published return honest
Partly. 11 percent of its legs traded coins the venue had not listed yet, and
they earned 34 percent of its gross profit.
inflated
Twenty one registered attempts to raise the profit of a live book, and the one that held
Report one put the crypto book at 13.7 percent a year. Measured again on every coin the venue ever listed, trading each one only while the venue listed it, the same book earns 12.2 percent at a Sharpe ratio of 1.29, once costs are charged the way the live executor actually pays them.
Twenty one changes were then registered in writing and tested against that honest baseline. One survives, in a weak sense. Admitting a coin after sixty days of price history rather than a year raises the Sharpe ratio to 1.74 and the annual return at one times gross to 20.7 percent. It is better in both halves of the sample and fails the significance test it was registered under. Most of what follows is about why it was adopted anyway, where its profit appears to come from, and how much leverage it can carry.
An experiment that compares a change against a baseline inherits every error in the baseline. Before any of the twenty one changes ran, the baseline itself moved three times, twice down and once up, and it ended close to where it started for reasons that have nothing to do with each other.
Signals come from spot prices on one exchange and trades go to perpetual futures on another. The backtest in report one assumed that any coin with a year of spot history could be traded. The venue's own record says otherwise. Its linear perpetuals launched on 22 March 2022, five days before the test sample begins, with seventeen contracts, and the rest arrived over the following months and years. A listing calendar was rebuilt from the venue's daily candles, taking the first and last day on which each contract actually traded. The volume filter matters, because a delisted contract keeps printing flat candles with zero volume, which would otherwise make it look alive forever.
Under that calendar, 8,839 of the deployed book's 77,236 legs, 11.4 percent, sat in coins the venue did not list at the time. Most fell in 2022 and 2023. They earned 34 percent of the book's gross profit, roughly three times their share of the legs, which is the signature of a selection effect rather than noise. Restricting legs to listed coins gives the baseline called B0, at 11.8 percent a year and a Sharpe ratio of 1.24.
The coin list behind report one was built from the contracts listed on the day it was built, with older history attached. Coins the venue had listed and then dropped were missing, which is survivorship in its most familiar form.1 Report two found the same mechanism erasing an equity strategy completely. Here the damage is milder. Adding the 107 coins the venue once listed and later removed lowers the Sharpe ratio from 1.24 to 1.05 at the original cost model.
The book runs four sleeves, each split into daily tranches that overlap for one or two weeks. The backtest charged a round trip on every one of those virtual legs. The live executor does something cheaper. Each day it adds up every tranche's target per coin and sends only the change in the net position, so a coin that one tranche sells while another buys costs nothing. Charging costs on that net trade instead gives 2.4 basis points a week at one times gross, against 6.7 per leg, and the difference is identical in both halves of the sample because it is arithmetic rather than luck.
Put together, the honest baseline used for everything below is every coin the venue ever listed, traded only while listed, with netted costs. It returns 12.2 percent a year at a Sharpe ratio of 1.29, with a deepest drawdown of 12.2 percent.
The backtest prices funding from the spot exchange's sister futures market, because that history is long. The venue's own funding is only published for about the last twelve months, so an archive of it now runs daily. Over the 15,467 legs that overlap the venue's history, from October 2025 to August 2026, the two venues' funding per leg correlates at 0.96 for longs and 0.97 for shorts.
The carry sleeve is the exception worth recording. It received 35.5 basis points per leg on the venue against 43.7 in the backtest, about a fifth less. Eleven months are too few to correct the backtest with, so this stands as a known optimism in the carry sleeve's figures.
Every change was written down with its parameters before it ran, so the rules could not drift towards whatever happened to work. Each arm is compared with its baseline after being rescaled to the same volatility, because a change that only adds risk is just leverage, and leverage is decided separately.
An arm passed if three things held. The weekly gain at equal volatility needed a t statistic above two, with a Newey and West variance to respect the overlap between weeks.2 The gain had to be positive in both halves of the sample. And the arm's annual return at the largest leverage inside the risk budget had to beat the baseline's at its own largest leverage. Counting the roughly forty configurations of report one, the final book is the best of 61 trials, which is the number its deflated Sharpe ratio is computed against.3,4
The two sides of each sleeve were scaled so that the book carries no net beta to the alt market, in the spirit of Frazzini and Pedersen.5 On the old universe this was the strongest arm of the whole programme, raising the netted Sharpe ratio from 1.48 to 1.88 and better in both halves.
Gain 7.2 bp a week with netted costs, t 0.81. On the wider universe it adds nothing, 1.72 against 1.76 without it.Two ways of tilting each side towards its best or calmest names.
−1.4 and +0.9 bp a week, t −0.40 and 0.23. Equal weights stay.Volatility management raises Sharpe ratios in many equity factors.6 Here the sleeves' volatility is already similar and changes slowly, so there is little to manage.
−1.5 bp a week, t −0.61.The four sleeve scores averaged into a single weekly book.
Higher raw return and a drawdown of 23.5 percent, so it carries far less leverage. t 0.88.The alt longs replaced by bitcoin and ether sized to each short's beta, which asks whether the long side earns anything of its own.
Positive in the first half and negative in the second once the hedge pays its funding. t 0.52.K2 deserves a sentence more, because its disappearance is informative. Two arms that each help alone and do nothing together are usually buying the same thing. The wider universe changes which coins land on each side, and whatever beta imbalance K2 was correcting in the old universe appears to be corrected by that change as well. The data cannot separate the two explanations, so the simpler book, without K2, was kept.
Posting the closing order at the touch saves a taker fee when it fills. The fills are not random, though. A resting sell fills when buyers arrive and the price is about to move away from it, which is the adverse selection that Glosten and Milgrom formalised.7 Unfilled exits then cross two hours later, after the price has drifted.
−0.4 bp a week, t −1.87, negative in both halves. The cleanest failure in the set.Fresher models sound strictly better. They are not, because a tree ensemble refitted on a few more weeks of overlapping labels mostly re-estimates noise and turns its ranking over without cause.
−4.7 bp a week, t −1.52, negative in both halves.Two changes to which rows the models learn from.
−0.8 and +0.8 bp a week, each with halves of opposite sign.Halve the shorts when alts have beaten bitcoin by twenty points in a month and funding is in its top decile, a crude guard against the crashes that hit momentum books.8
The registered rule was active on one day out of 2,069. It changes nothing because it almost never fires.The gap between two venues' funding rates and the shift in open interest from one to the other, as ranked features. Funding on perpetual futures is a price of leverage demand,9 so a venue where longs pay more might flag a crowded coin.
+0.3 bp a week, t 0.14, negative in the first half. Report one's positioning features failed the same way.Time series momentum, the most replicated trend signal there is,10 as a fifth sleeve and a hedge for the regime that hurts a market-neutral alt book.
Sharpe 0.29 alone and correlated at 0.06 with the book. Too weak to pay for its capital. t 0.15.The weakest sleeve, kept in report one because it hedges the other three.
Return rises and the drawdown deepens from 10.5 to 17.2 percent. The hedge is real. t −0.45.The common thread in the failures is that each one re-arranges information the model already has. Weights, refits, overlays and execution tricks act on the same ranked list of the same coins. None of them changes what the ranking can see, and the ranking was evidently close to as good as those levers can make it.
The book ranks coins by liquidity first and trades the thirty most liquid. Every earlier version also required a full year of history before a coin could be ranked at all. Recent listings carry a large share of crypto's trading volume, and coins under a year old now make up a sixth of the legs v2 trades. The old rule kept every one of them out.
The history requirement was loosened in three registered steps on the honest universe, from 365 days to 180, then 90, then 60. The Sharpe ratio rose at every step. A monotone response to a dose is weaker evidence than significance and stronger evidence than a single good number, since noise has no reason to line up in order.
| Minimum history | Return | Sharpe | Max drawdown | Weekly mean |
|---|---|---|---|---|
| 365 days, the old rule | 9.7% | 1.05 | −13.6% | 18.7 bp |
| 180 days | 14.8% | 1.45 | −16.5% | 27.5 bp |
| 90 days | 16.6% | 1.51 | −13.3% | 30.7 bp |
| 60 days | 17.8% | 1.56 | −14.6% | 32.6 bp |
| 60 days, with the 26 current listings | 18.1% | 1.54 | −14.6% | 33.2 bp |
The last row is v2. The 26 coins added in October 2026 enter the backtest only where the venue listed them, so all of their legs fall in 2025 and 2026. With netted costs v2 returns 20.7 percent a year at a Sharpe ratio of 1.74, a deepest drawdown of 13.3 percent and a worst week of −3.8 percent.
A natural first look splits v2's profit by the age of the coin when each leg opened and by the side of the leg. Taken raw, that split credits almost everything to the short side. It is misleading, because the alt market fell over much of the sample and young coins are shorted far more often than they are bought, so a short leg earns the market's fall whether or not the ranking was right. Charging every leg for its beta to the equal weighted alt market over the same days removes that, and the picture changes shape.
The amber point is worth about 4.6 basis points a week after adjustment, roughly a fifth of v2's market-adjusted return, from 4,467 short legs in coins less than six months old. Crypto returns are known to load on size and momentum,11 and neither explains why the youngest coins should be the profitable shorts. Equity markets offer a closer analogy. Newly listed firms underperform comparable ones for years after their offering,12,13 and prices dip when lock-up periods end and insiders become free to sell.14 A token's early months combine both, with heavy attention at listing and a supply schedule that keeps releasing locked coins to early holders.15 Nothing in the model knows about supply, so whatever it captures here it captures through price and volume alone.
A one year history rule removed exactly the part of the market where the short side had the most work to do.
Against the honest baseline, v2's gain at equal volatility is 7.9 basis points a week at a t statistic of 1.12. The registered rule asks for two. The deflated Sharpe ratio, which estimates the probability that the true Sharpe ratio is above zero after choosing the best of 61 trials, is 0.89 for v2 against 0.60 for the baseline.4 The evidence that v2 is better than the baseline is weak. The evidence that it is a real book at all is reasonably strong.
Three things tipped the decision. The dose-response in the table above is orderly. The gain is positive in both halves. And the cost of being wrong is limited by construction. v2's weekly returns correlate at 0.79 with the baseline's, so if the young-coin effect is noise, v2 degrades towards the baseline rather than towards zero. Under a mandate that explicitly trades more risk for more profit, a change with a bounded downside and a plausible mechanism is worth running in parallel, where live data can confirm or retire it.
A market-neutral book with a Sharpe ratio near 1.7 at one times gross is a candidate for leverage almost by definition. How much depends less on the backtest's average than on how much of that average survives contact with live markets, which nobody knows in advance.
Three year paths were drawn by block bootstrap from v2's 233 weekly returns, in blocks of eight weeks to keep the clustering of bad weeks, and 5,000 paths per setting.16 Four assumptions were run about the edge, from the full backtest mean down to none of it, by shifting the weekly returns before resampling. The risk budget, fixed before any of this, is at most a ten percent chance of a drawdown deeper than half the capital within three years, at half the backtest edge.
The venue's retail margin schedule across its 204 perpetual contracts asks for an initial margin between 1 and 20 percent of notional, mostly 5 or 10, with maintenance at half of that. At two times gross the whole account would have to lose somewhere between 80 and 99 percent of its capital before positions were liquidated. Every risk limit in the book acts long before that.
Four stress episodes were replayed hour by hour on the venue's own candles, with v2's actual open positions at two times. A pessimistic bound marks every long at its hourly low and every short at its hourly high at the same moment, which cannot happen but cannot be exceeded either.
| Episode | Priced | Worst drawdown | End of episode | Every wick at once |
|---|---|---|---|---|
| Liquidation cascade, October 2025 | 99% | −3.8% | +4.1% | −24.4% |
| Alt rally, November and December 2023 | 86% | −1.7% | +2.4% | −1.3% |
| Alt top, March 2024 | 92% | −2.5% | +0.6% | −6.8% |
| Carry unwind, August 2024 | 97% | −1.9% | +4.0% | −2.1% |
One finding here was operational rather than statistical. The book runs as a short process each day, and every process started with fresh risk state. The sleeve limits and the kill switch therefore only ever saw the losses of a single run, which is to say they could never trip. The state is now rebuilt from the trade log when a process starts.
That fix made the limits real, and the limits written for one times gross turned out to be wrong for two. A single sleeve swings far more than the book, with weekly volatility between 2.4 and 3.2 percent against 1.6 for the four together, because the sleeves partly offset each other. A limit on each sleeve throws that diversification away.
| Limit | Leverage | Full edge | Half the edge |
|---|---|---|---|
| Retire a sleeve at −30% | 1× | 21 to 41% | 42 to 60% |
| Retire a sleeve at −30% | 2× | 81 to 95% | 92 to 98% |
| Kill switch at −35% | 2× | 8.2% | 31.9% |
| Kill switch at −50% | 2× | 0.8% | 6.1% |
A limit that fires on a working book nine times in ten is a scheduled shutdown. Drawdown control belongs at the level of total capital, where the theory of optimal drawdown-constrained investment puts it,18 and where the diversification between sleeves has already been counted.
v2 runs at two times gross. Sleeve retirement is effectively switched off in both books. v2's kill switch sits at −50 percent of capital, and the original book keeps −35. The daily loss limit of ten percent per sleeve is unchanged.
The plan for real money starts at one times for a few weeks to check fills against the backtest, then moves to two. Going beyond two would need live evidence that more than half the edge is real.
v2 trades in paper beside the original book, with the same capital, its own state and kill switch, and its own weekly report. Whichever holds up better on live data goes to real money, and the target is to do that before the end of 2026.
Four mechanical changes went to both books with it. The risk state survives restarts, as above. The protective stop resting at the venue now triggers on the mark price rather than the last trade, which a single wick can reach, and at the same distance the backtest models. Before, a short was stopped at a fifty percent rise where the model holds it to sixty five. The live process now loads only the price history a decision uses, and its decisions match those from the full history to within rounding. And a weekly job adds the venue's new listings to v2's universe with enough history to rank them. Without it the young-coin effect could not reach the live book at all, since a coin that never enters the universe can never be sixty days old in it.
The central result is not significant at the level the programme set for itself, and the deflated Sharpe ratio rests on an assumed spread of Sharpe ratios across trials, since that spread was not recorded for report one's configurations. The probability of backtest overfitting was not computed.19 Published anomalies tend to weaken after discovery,20 and an effect found in one sample of four and a half years can do the same.
v2's advantage over the baseline is concentrated in 2023 and 2025, years with large waves of new listings. v2's own two year walk forward, which the weekly report uses to set expectations, shows far higher returns than the full sample, and should be read as an upper bound. The carry sleeve earns about a fifth less funding on the venue than the backtest assumes. Coins the venue listed without a spot market on the signal exchange are absent from every universe here, a selection this report cannot measure. And the attribution in Figure 4 is approximate, since it rests on a sixty-day beta.
Part three suggests the young-coin profit lives on the short side of new listings. The model sees only prices, volumes, funding and positioning, so it can only infer that from price behaviour. The obvious next experiment gives it the cause directly.
Two kinds of data would do it. The first is a history of each coin's circulating supply against its total supply, which says how much of a coin is still locked away. The second is the schedule of unlocks, the dates on which locked tokens are released to founders, investors and treasuries. Both are sold by specialist data providers, and one month of access to each is enough to download the history.
The hypotheses would be registered before the data arrives, in the same form as the twenty arms. A large scheduled unlock in the next thirty days predicts underperformance over the following weeks, in the way lock-up expiries do for equities.14 A low circulating share predicts negative cross-sectional returns, through the steady supply overhang that token models describe.15 Both would enter as ranked features and as a filter on the long side, and both would count as new trials against the deflated Sharpe ratio.
One risk has to be checked first, and report two shows why. Supply and unlock histories are probably best for coins that still exist. If the data covers delisted coins poorly, a supply feature would end up trading survivors while the price features trade everything, which reintroduces survivorship through a side door. Coverage of the 107 delisted coins is therefore the first thing to measure, and the experiment runs only if it is close to complete.
Three smaller items are already in motion. The hourly bid and ask archive will allow the execution hour to be chosen on cost. A year of the venue's own funding will make it usable as the carry signal. And the live weeks of both books will update the estimate of how much of the edge is real, which is the number every leverage decision above depends on.
Hourly spot bars from the signal exchange, from 2020 for most coins, for 299 coins that the trading venue has listed as linear perpetuals at some point. Funding and positioning come from the signal exchange's futures archive. The venue's daily and hourly candles, funding prints and instrument snapshots give the listing calendar, the margin schedule, the stress replays and the funding comparison. A second venue's public funding and open interest history serves the cross-venue arm only.
Nine walk-forward windows of 180 days, with models refitted at the start of each window on everything before it, give 233 weekly returns from March 2022 to September 2026. Four sleeves at equal capital, a weekly gradient boosted ranker, a two-week blend, two-week momentum and weekly carry, each hold the top and bottom fifth of the thirty most liquid eligible coins, split into daily tranches. Entries post at the touch and cross what has not filled. A stop closes a leg after a log move of 0.5 against it, checked against hourly extremes.
Each arm was evaluated through a table holding one row per leg, from which construction, cost and overlay arms only reweight. That table reproduces the library backtest exactly, to a largest weekly difference of 5×10−17, and v2's library backtest matches it in the same way, so the paper book trades what was tested.
Differences between arms are tested after rescaling the arm to the baseline's volatility, with a Newey and West variance and a Bartlett kernel.2 The deflated Sharpe ratio uses 61 trials and an assumed cross-trial dispersion of 0.5 in annual Sharpe ratio.4 Annual Sharpe ratios are read with the usual caution about their sampling error, which for four and a half years of weekly data is between 0.6 and 0.75 at the levels reported here.21 Bootstrap paths are circular blocks of eight weeks, 5,000 per setting, with the edge scaled by subtracting a fraction of the mean return from every week before resampling.
Every figure is generated by a script from the stored weekly series and leg tables, and the derived numbers in the text come from one file computed from the same store, so no figure can drift from the result it shows. The registered arms, their parameters and the pass rule were committed before the first run.
Research publications of Ciya Labs. Not investment advice; backtested and simulated results do not predict future returns. Imprint · Privacy