Research35 min read
Running my own book
How I went from reading about systematic trading to running a six-figure systematic book on crypto perpetual futures, every idea I killed along the way, the oil trade I kept, and why I plan on a much lower Sharpe than my backtest shows.
From My own book · Live since Sep 2026
For the last month most of my evenings have gone into one thing: building and running my own systematic book on Hyperliquid, a perpetual futures exchange that lists crypto and, more recently, stock and commodity perps. It's live, it's six figures, and it trades on its own every hour.
This post is the whole arc: what I read, what I tried, what failed (most of it), and what I actually run. I'm going to show the numbers, including the ugly ones, because the ugly ones are where I learned the most.
- Early Sept 2026Copy-trading and the first systematic waveTens of thousands of leaderboard rows, millions of fills, and nothing that survived.
- Mid Sept42 pre-registered hypothesesCosts per trade, a kill rule for every idea, and a deflated Sharpe for every result. Slow trend survived.
- Sept 22 to 24Tournament, blind test, and idea miningThousands of variants, a red team, and a blind run on 2018 to 2019. The oil spread trade came out of this phase.
- Sept 25LiveThe trend core and small event sleeves, with every new idea starting in shadow mode.
The scorecard I use
No single number tells you whether a strategy is good, so every result in this post is scored on the same set of metrics. Each one answers a different question, and a strategy has to look reasonable on all of them.
| Metric | What it measures | Why I track it |
|---|---|---|
| Sharpe ratio | Annual return above the risk-free rate, divided by annual volatility (daily returns, annualized over 365 days) | Return per unit of risk; the standard way to compare strategies |
| Sortino ratio | Like Sharpe, but only counts downside volatility | Trend following has big up moves; Sharpe penalizes those, Sortino doesn't |
| CAGR | Compound annual growth rate | What the money actually does over years, after compounding |
| Annual volatility | How much returns swing in a typical year | Sets position size: I target a volatility, not a return |
| Max drawdown | Largest fall from a peak before recovering | The pain number. My whole risk budget is built on it |
| Calmar ratio | CAGR divided by max drawdown | Return per unit of worst-case pain; the one I weigh most |
| Worst month | The single worst calendar month | What a bad month really feels like |
| Win rate / hit rate | Share of trades or months that make money | Shape of returns, not quality: trend wins under half the time |
| Skew | Whether big moves are mostly up or mostly down | Trend has positive skew (rare big wins); carry has negative skew (rare big losses) |
| Turnover | How often the book trades its whole size | Drives costs, which kill most short-horizon ideas |
| Deflated Sharpe | The probability a Sharpe is real, given how many variants I tried, the sample length, and fat tails | Trying enough ideas always produces a lucky one |
Every Sharpe in this post also comes with an error bar, even when I don't print it, and it's wider than people expect. For an annualized Sharpe measured over T years, the standard error is roughly 1/√T when the true Sharpe is near zero, and the version I use, √((1 + SR²/2) / T), is wider still for a high Sharpe. Two consequences I keep in front of me:
- My crypto backtest covers about 5.7 years (2,092 days). That alone puts a standard error of about 1/√5.7 ≈ 0.42 on any Sharpe from it, and ±0.64 on the book's 1.64 with the wider formula.
- To tell a true Sharpe of 1 from 0 at two standard errors, I need 2/√T ≤ 1, so T ≥ 4 years. One good year proves nothing: a single year has a standard error of about 1.
Where the mindset came from
Before writing any code, I read. The books that shaped how I think about this:
- Robert Carver, Systematic Trading, which is the most practical book I know on sizing, forecasts, and not overfitting
- Curtis Faith, Way of the Turtle, and Michael Covel, The Trend Following Bible, for the case that simple trend rules, held with discipline, actually get paid
- Marcos López de Prado, Advances in Financial Machine Learning, and his papers with David Bailey on the deflated Sharpe ratio and the probability of backtest overfitting
- Giuseppe Paleologo ("Gappy"), Advanced Portfolio Management and The Elements of Quantitative Investing
- Moskowitz, Ooi and Pedersen on time-series momentum, and Daniel and Moskowitz on momentum crashes
Two people shaped my attitude more than any single book.
Corey Hoffstein (Newfound Research, and the Flirting with Models podcast) taught me about rebalance timing luck: if your strategy rebalances on Fridays, a big part of your backtest is just which Friday you happened to pick. His answer is to split the book into tranches that rebalance at different times, so no single arbitrary choice can make or break you. More broadly, he taught me to be suspicious of any result that depends on a choice I made without a reason.
Gappy shaped how I think about risk. The idea I took most seriously from him is that you should scale exposure with conditions, not switch it on and off, and that the job of a portfolio manager is mostly to survive long enough for a modest edge to compound. I'm risk-averse by temperament, and his writing gave that instinct a framework: decide your drawdown budget first, size everything from it, and treat a strategy you can't explain as a risk, not an opportunity.
The first wave: everything I believed was wrong
I started where a lot of people start on Hyperliquid: copy-trading. Every account's fills are public, so you can rank the "best" traders and follow them. I pulled about 45,000 leaderboard rows and over four million fills, built 14 statistical gates per wallet, and had three independent skeptics attack every candidate.
None survived. The top wallets turned out to be single lucky longs, market-making bots you can't replicate, hidden losses in other accounts, or deposits that looked like profit. A simulated follower with a one-to-two second lag kept a median of about 19% of the leader's return.
Then I tested the systematic families: momentum, mean reversion, carry, cross-sectional ranks, wallet flow. This chart is the most important thing I learned that week:
- In-sample
- Out-of-sample
Show data
| idea | In-sample | Out-of-sample |
|---|---|---|
| Funding carry | 2.33 | -0.05 |
| Cross-asset / weekend | 2.06 | -1.11 |
| Vol-regime momentum | 1.40 | -0.48 |
| Cross-sectional momentum | 1.33 | -0.53 |
| Follow top wallets | 1.23 | -1.42 |
| Time-series momentum | 1.07 | 0.36 |
| Mean reversion | 0.66 | 0.04 |
| Combined system | 0.39 | 0.32 |
The ideas that looked best on the data they were fit to did the worst afterwards. Funding carry went from an in-sample Sharpe of 2.33 to slightly negative. Smart-money flow went from 1.23 to -1.42. Only plain time-series momentum held on to anything.
The setup behind that chart: daily bars from July 2022 to September 2026, a walk-forward that trains on 365 days, tests on the next 90, then steps forward 90, with parameters chosen on each training window only, and costs of 4.5 bps taker plus 1.5 to 8 bps of slippage per side. The in-sample number is what the parameter search reported as its best, which is exactly why it's inflated: the more variants a search tries, the higher its best in-sample score climbs by luck alone. The rank correlation between in-sample and out-of-sample Sharpe across the families was negative in every combination I computed (between -0.36 and -0.64). Over seven or eight points that's a sign, not a precise magnitude, but the sign is the lesson: the better a family looked in training, the worse it did afterwards.
The second wave: 42 hypotheses
So I got more rigorous. Every idea was pre-registered with a kill rule before I looked at its results, every cost (fees, slippage by liquidity class, funding) was charged per trade, and every result was judged with a deflated Sharpe that accounts for how many things I'd tried. The second wave alone was 42 hypotheses, 54 sleeve constructions and 6,028 trials. By September 22 the whole program was about 7,800 registered backtests across 45 strategy families, on 6.7 years of hourly crypto history (355 coins, 189 of them delisted), 21 years of daily cross-asset history, and 14 months of Hyperliquid one-second candles and order books.
The cost model is deliberately simple and deliberately pessimistic. Per side, on every dollar traded:
- Fee: 4.05 bps, my taker fee at the tier I trade.
- Slippage allowance by liquidity class: 1.5 bps for BTC and ETH, 3 bps for liquid alts, 8 bps for everything else.
- Market impact on top, with the standard square-root model: daily volatility × √(order size / daily volume). Doubling an order costs about 1.4x as much per dollar, not 2x, but it never stops growing.
- Funding, charged hourly on every open position at the historical rate, long or short.
A worked example of why this kills fast ideas: a round trip on a liquid alt costs 2 × (4.05 + 3) ≈ 14 bps before impact. A rule that trades a few times a day needs a gross edge comfortably above that on every trade, and almost nothing short-horizon I tested had one.
Show data
| idea | Out-of-sample Sharpe |
|---|---|
| Slow crypto trend | 1.03 |
| Trend ensemble | 0.77 |
| Cross-sectional momentum | 0.42 |
| Weekend reversal, stock perps | 0.93 |
| Commodity seasonality | 0.18 |
| 24h reversal | -0.21 |
| Liquidation-cascade reversal | -0.32 |
| Calendar seasonality | -1.18 |
| Perp basis | -1.37 |
| Funding carry, hourly | -3.38 |
| Smart-money flow, hourly | -3.90 |
Most of it died. Funding carry, liquidation cascades, smart-money flow, seasonality (zero of 516 calendar hypotheses survived a multiple-testing correction), and every intraday rule I tried on the stock and commodity perps. My own favourite, a Bollinger-band fade, made 17 bps per trade gross and cost 14 bps to trade. Net Sharpe: zero.
The survivor was the least exciting idea on the list: slow trend on crypto, long and short, held until the signal flips. Out-of-sample Sharpe 1.03, and still 1.05 when I re-ran it with 189 delisted coins added back, so it wasn't living off survivorship.
The blind test
A backtest you tuned on 2021 to 2026 data can't tell you much about 2021 to 2026. So I froze the rules and ran them on 2018 and 2019, two years of prices I had never looked at, including one of the worst crypto bear markets on record.
- Trend book
- BTC buy and hold
Show data
| month | Trend book | BTC buy and hold |
|---|---|---|
| 2017-12-01 | 1.00x | 1.00x |
| 2018-01-01 | 1.04x | 0.75x |
| 2018-02-01 | 1.03x | 0.75x |
| 2018-03-01 | 1.06x | 0.50x |
| 2018-04-01 | 1.16x | 0.67x |
| 2018-05-01 | 1.16x | 0.55x |
| 2018-06-01 | 1.20x | 0.47x |
| 2018-07-01 | 1.17x | 0.56x |
| 2018-08-01 | 1.21x | 0.51x |
| 2018-09-01 | 1.18x | 0.48x |
| 2018-10-01 | 1.15x | 0.46x |
| 2018-11-01 | 1.23x | 0.29x |
| 2018-12-01 | 1.24x | 0.27x |
| 2019-01-01 | 1.21x | 0.25x |
| 2019-02-01 | 1.24x | 0.28x |
| 2019-03-01 | 1.28x | 0.30x |
| 2019-04-01 | 1.22x | 0.39x |
| 2019-05-01 | 1.28x | 0.62x |
| 2019-06-01 | 1.29x | 0.79x |
| 2019-07-01 | 1.30x | 0.73x |
| 2019-08-01 | 1.34x | 0.70x |
| 2019-09-01 | 1.36x | 0.60x |
| 2019-10-01 | 1.33x | 0.67x |
| 2019-11-01 | 1.35x | 0.55x |
| 2019-12-01 | 1.34x | 0.52x |
It held up on every metric, not just Sharpe:
| Blind test, 2018 to 2019 | Value |
|---|---|
| Sharpe / Sortino | 1.28 / 1.96 |
| Annual return | +15.6% |
| Annual volatility | 11.9% |
| Max drawdown | 7.5% |
| Months positive | 67% |
| Worst month | -4.4% |
| 2018 alone | +23.9% while BTC fell 73% |
That's the result that convinced me the trend core is real: not because the number is big, but because it showed up on data the rules had never seen.
It's also less precise than the table makes it look, and the details matter:
- What was frozen. Every parameter: the moving-average pairs and the volatility switch, the sizing, the universe rules, the breakout sleeve, the BTC hedge cap, the volatility target, the costs (5.55, 7.05 and 12.05 bps per side by liquidity class) and the leverage. Nothing was re-tuned after the run.
- What the data forced. Binance only launched its perps in September 2019, so the test trades spot price series as if perps had existed. Funding is charged as a flat 11.6% a year on longs only. There's no square-root impact term, because I have no Hyperliquid volume for 2018 to scale it by. The universe came from Binance's archive including delisted pairs: 190 coins, a median of 44 traded at a time.
- How wide the answer is. A 20-day block bootstrap gives a 90% interval of 0.22 to 2.27 for the Sharpe. The probability it's below zero is about 2%; below 0.9 it's about 28%.
- Stress rows. Making shorts pay funding too, or doubling taker costs, each takes about 0.12 off, to 1.16.
- The two years aren't alike. 2018 alone was a Sharpe of 1.93 (+23.9%), 2019 alone was 0.68 (+7.9%), even though BTC rose 94% that year. Two years is two draws, not a distribution.
So the blind test says "this is very unlikely to be zero", not "this is 1.28". That's why the number I plan on is lower.
The oil trade
The one trade in the book I find genuinely elegant is in oil.
Hyperliquid lists perps on WTI and Brent crude. When the CME futures market closes for its daily maintenance hour, those perps can't reference the real futures price, so each one's price oracle falls back to a moving average of its own order book. For that hour, WTI and Brent drift apart on unrelated flow, even though the real spread between them has barely moved. When the futures market reopens, both oracles snap back to the real prices.
So just before the reopen, if the spread has drifted far enough from where it was when the market closed, I fade it: long one leg, short the other, and exit once it snaps back. Trading the spread rather than one leg removes the common oil move, which is exactly what killed the simpler version.
The mechanism is written down in the venue's own oracle documentation, which is why I trust it more than a pattern. While the external market is shut, each oil perp's oracle is an exponential moving average of its own order book's impact prices (30-minute time constant, moving at most 9.5% of the gap per update), and "when external inputs resume, the oracle reverts to the externally derived price on the next tick". CME's daily halt runs 17:00 to 18:00 New York time, Monday to Thursday, plus the weekend from Friday 17:00 to Sunday 18:00.
The rule, fixed before the first backtest:
- Anchor. The log spread s = ln(WTI) − ln(Brent) at 17:00:00 (Friday 16:55 for the weekend).
- Signal. At 17:55:00, if the spread has moved at least 10 bps from the anchor, and both legs have printed in the last 10 seconds, take the opposite side: $1 of one leg against $1 of the other.
- Exit. The first second the spread is back at the anchor, or 18:15 at the latest. No stop loss.
- Costs. Taker on both legs at the measured book spread, plus funding for any position held through the 18:00 funding print.
The data is one-second candles from early March 2026 (when the Brent perp listed) to September 21, split into a development window and a holdout at June 24.
- Modelled fills at touch
- Pessimistic fills
Show data
| date | Modelled fills at touch | Pessimistic fills |
|---|---|---|
| 2026-03-09 | 172 | 124 |
| 2026-03-10 | 225 | 182 |
| 2026-03-11 | 253 | 204 |
| 2026-03-12 | 269 | 202 |
| 2026-03-15 | 487 | 289 |
| 2026-03-19 | 511 | 315 |
| 2026-03-22 | 641 | 439 |
| 2026-03-24 | 703 | 473 |
| 2026-03-26 | 715 | 479 |
| 2026-03-29 | 770 | 545 |
| 2026-03-31 | 780 | 541 |
| 2026-04-05 | 786 | 533 |
| 2026-04-06 | 790 | 531 |
| 2026-04-07 | 824 | 549 |
| 2026-04-08 | 836 | 554 |
| 2026-04-12 | 1,131 | 849 |
| 2026-04-19 | 1,194 | 918 |
| 2026-04-20 | 1,239 | 952 |
| 2026-04-26 | 1,294 | 1,006 |
| 2026-04-28 | 1,301 | 1,000 |
| 2026-04-29 | 1,414 | 1,103 |
| 2026-05-03 | 1,458 | 1,127 |
| 2026-05-06 | 1,469 | 1,127 |
| 2026-05-07 | 1,508 | 1,149 |
| 2026-05-10 | 1,525 | 1,157 |
| 2026-05-11 | 1,539 | 1,159 |
| 2026-05-18 | 1,581 | 1,192 |
| 2026-05-20 | 1,589 | 1,181 |
| 2026-05-21 | 1,623 | 1,203 |
| 2026-05-24 | 1,582 | 1,152 |
| 2026-05-25 | 1,601 | 1,136 |
| 2026-06-03 | 1,610 | 1,125 |
| 2026-06-07 | 1,648 | 1,117 |
| 2026-06-08 | 1,674 | 1,115 |
| 2026-06-11 | 1,682 | 1,107 |
| 2026-06-15 | 1,696 | 1,098 |
| 2026-07-12 | 1,749 | 1,095 |
| 2026-07-14 | 1,742 | 1,075 |
| 2026-07-15 | 1,736 | 1,056 |
| 2026-07-22 | 1,760 | 1,090 |
| 2026-07-27 | 1,771 | 1,101 |
| 2026-08-02 | 1,803 | 1,110 |
| 2026-08-10 | 1,809 | 1,104 |
| 2026-08-11 | 1,827 | 1,107 |
| 2026-08-23 | 1,839 | 1,100 |
| 2026-08-25 | 1,854 | 1,102 |
| 2026-08-30 | 1,895 | 1,126 |
| 2026-09-06 | 1,899 | 1,119 |
| 2026-09-10 | 1,969 | 1,186 |
| 2026-09-13 | 2,027 | 1,214 |
| Oil spread trade, Mar to Sep 2026 | Value |
|---|---|
| Trades | 50 |
| Win rate | 94% |
| Average net per trade | +40.5 bps |
| t-statistic | 5.0 |
| Edge per trade, Mar-Apr vs May-Sep | +67 bps vs +22 bps |
Over 50 trades, 94% were winners, at about 40 bps net each. That's what the rule earns; what convinced me the mechanism is real is where it doesn't earn:
- Six placebo clocks. The identical rule at 02:00, 05:00, 08:00, 11:00, 13:00 and 15:00, when both futures are trading and there's no oracle to snap, nets −7.8 to +2.2 bps a trade (median −2.9).
- Fridays and Saturdays at the same clock, when CME stays shut and nothing snaps at 18:00: −0.9 bps gross.
- Exit at 17:59:30, thirty seconds before the snap: −5.1 bps. The entire profit sits on the re-anchor.
Show data
| exit | Net bps per trade |
|---|---|
| 17:59:30 | -5.10 |
| 18:00:30 | 23.40 |
| 18:01 | 13.50 |
| 18:02 | 16.40 |
| 18:05 | 19.00 |
| 18:15 | 22.20 |
| 18:30 | 25.00 |
| TP at anchor | 31.40 |
The snap also isn't instantaneous, which is the other thing a skeptic should ask about the 40 bps: can I actually get those prices? The average WTI move measured from 17:59:55 is 7.0 bps at 18:00:00, 12.3 bps two seconds later and 15.1 bps after five. So I re-priced every entry and exit at the first real print after a reaction delay:
Show data
| Seconds of delay | Net bps per trade |
|---|---|
| 0 | 40.50 |
| 2 | 31.40 |
| 5 | 27.00 |
| 10 | 24.20 |
| 30 | 22.70 |
| 60 | 18.40 |
Doubling costs leaves +32.6 bps; tripling them with a two-second delay leaves +15.6. A stop-and-target bracket (reward-to-risk 1.5) does worse than no stop at all, +16.2 bps, because during the halt the spread often widens further before it snaps, and a stop books exactly that excursion.
It's also sized small, on purpose, for four reasons:
- Decay. The chart's slope flattens: net per trade falls about 1.4 bps per event (t of −2.5). It was about 67 bps per trade in March and April (21 trades) and about 22 since (29 trades). The holdout is only 14 trades.
- Concentration. The five best trades carry 46% of the net profit, four of them weekends with 100 to 300 bps spread moves in March and April.
- Capacity. Each leg prints only about $4 to 4.4 million in the 17:00 to 18:00 hour. I cap each event at 2% of that and charge square-root impact on top. Under that model the trade nets +10.5 bps of gross notional on a $25k book, +3.1 on $100k, and −1.1 on $250k (about $50k to $125k a leg), where 86% of events hit the cap. That's a few thousand dollars a year, at best.
- A fee I don't control. Both oil perps currently trade at a reduced "growth mode" fee. At the normal fee the edge is gone, so the sleeve goes to zero the day either leg leaves growth mode.
Live it runs capped at $12.5k a leg. I'm also honest that the holdout isn't clean: I picked the idea after a descriptive look at the full sample, so the only truly clean test is the live record. It's a clean mechanism that can't be a big trade, and with the edge decaying I'm comfortable describing it.
What I actually run
The final book, in structure:
- A crypto trend core across roughly 45 to 50 perps, long and short, decided hourly. A fast and slow moving-average pair, switching to a slower pair when short-term volatility spikes. Positions are sized inversely to downside volatility, and exits only happen on a signal flip. Every trailing stop I tested made it worse, because the trend premium lives in the right tail.
- A breakout sleeve on BTC and ETH, and a daily book-level volatility target with a partial BTC hedge.
- A handful of small event sleeves on the stock and commodity perps, including the oil trade, each sized to its capacity.
- Shadow sleeves that log what they would do without trading, and only get capital once their own live record earns it.
The trend core, precisely
Since "trend following" can mean almost anything, here is the exact rule.
- Universe. BTC and ETH always, plus the top 40 alts by 30-day volume. A coin enters inside rank 40 and only leaves at rank 60 or worse, so the edge of the list doesn't churn. Each one also needs a live Hyperliquid perp, at least $1m a day of volume, and 30 days of listing. Signals read Binance hourly closes, the longest clean history; orders go to Hyperliquid.
- Direction. Compare a fast and a slow exponential moving average of hourly closes. In a calm market that's 96 against 384 hours (4 and 16 days); in a volatile one, 480 against 1,920 hours (20 and 80 days). Fast above slow is long, below is short.
- Calm or volatile. A coin turns volatile when its 7-day volatility rises above 1.1x its 90-day volatility, and calm again below 0.9x. The gap between 1.1 and 0.9 is hysteresis: without it, a coin sitting near the line would flip between speeds every few hours and pay costs for nothing.
- Sizing. At entry, size = risk budget / downside volatility, so every coin starts at the same risk. I use downside volatility rather than total volatility so that a coin isn't shrunk just for moving up fast, which is the move a trend book is trying to own.
- Holding. Once in, a position is not resized while its direction holds, so a coin that doubles becomes twice the dollars. Runners didn't earn less afterwards, and they flipped less.
- Exits. Only a moving-average cross (or leaving the universe). No stop, no target, no time limit. In the exit study (eight trailing rules, paired block bootstrap), the baseline signal-exit Sharpe of 0.82 fell to −0.66 with the best of them, a 3-ATR trail (ATR, average true range, is the typical size of a bar's move). The premium is in the right tail, and a stop cuts exactly the trades that pay for everything else.
- Split. Trend and the BTC/ETH breakout sleeve (4-hour bars, 8, 16, 32 and 64-day lookbacks) share the risk 80/20.
- Book overlays, set daily. A volatility target aims the whole book at 30% a year on a 60-day estimate, capped at 2x and only moved when it's 20% off. In 2025 and 2026 that cap bound on 94% of days, so in practice the book has run well below its target (about 17% realized over 2021 to 2026). A BTC short offsets half of the book's measured BTC beta; half rather than full was the one change that also held on the blind 2018 to 2019 test.
- Cadence. The decision runs 5 seconds after each hourly close and the orders, immediate-or-cancel inside a price band, leave about 10 seconds after it. Any change smaller than 10% of a position is skipped.
Sizing is in "units": one unit is roughly a 10%-volatility core per $100k of equity, and the live core runs 1.86 units, about 0.6x equity gross today. Where 1.86 comes from is in the risk section below.
Costs and funding at that size. The core turns over about 24.5x equity a year, and 73% of that is trend flips, which no fee trick touches. The backtest charges about 3.3% of equity a year for fees, slippage and impact. Funding is charged hourly on every position. Because a trend book in a bull market is mostly long, it mostly pays funding: a live snapshot had it at about 7.9% of gross a year, against 4.9% in the research. I tested more than 99 funding-based exit rules (close a long when funding gets expensive, and so on) and none helped; for the highest-risk longs, funding is 0.3% to 8% of the gross edge, too small to be worth giving up the trend for.
Diversification, honestly. The sleeves' daily returns correlate about −0.03 with the core, which is why they help even though they're small. But that's a correlation measured on their own 5 to 14-month records, and correlations jump toward 1 in a real liquidity event. I don't count on the sleeves to save the core in a crash.
What the backtest says
- Full book
- Trend core alone
Show data
| date | Full book | Trend core alone |
|---|---|---|
| 2020-12-31 | 1.00x | 1.00x |
| 2021-01-17 | 1.13x | 1.13x |
| 2021-02-14 | 1.33x | 1.33x |
| 2021-03-14 | 1.39x | 1.39x |
| 2021-04-11 | 1.39x | 1.39x |
| 2021-05-09 | 1.51x | 1.51x |
| 2021-06-06 | 1.38x | 1.38x |
| 2021-07-04 | 1.39x | 1.39x |
| 2021-08-01 | 1.40x | 1.40x |
| 2021-08-29 | 1.63x | 1.63x |
| 2021-09-26 | 1.68x | 1.68x |
| 2021-10-24 | 1.71x | 1.71x |
| 2021-11-21 | 1.70x | 1.70x |
| 2021-12-19 | 1.78x | 1.78x |
| 2022-01-16 | 1.74x | 1.74x |
| 2022-02-13 | 1.76x | 1.76x |
| 2022-03-13 | 1.78x | 1.78x |
| 2022-04-10 | 1.83x | 1.83x |
| 2022-05-08 | 1.92x | 1.92x |
| 2022-06-05 | 2.00x | 2.00x |
| 2022-07-03 | 2.03x | 2.03x |
| 2022-07-31 | 2.00x | 2.00x |
| 2022-08-28 | 2.08x | 2.08x |
| 2022-09-25 | 2.00x | 2.00x |
| 2022-10-23 | 1.98x | 1.98x |
| 2022-11-20 | 2.08x | 2.08x |
| 2022-12-18 | 2.04x | 2.04x |
| 2023-01-15 | 2.12x | 2.12x |
| 2023-02-12 | 2.21x | 2.21x |
| 2023-03-12 | 2.25x | 2.25x |
| 2023-04-09 | 2.18x | 2.18x |
| 2023-05-07 | 2.15x | 2.15x |
| 2023-06-04 | 2.14x | 2.14x |
| 2023-07-02 | 2.21x | 2.21x |
| 2023-07-30 | 2.13x | 2.13x |
| 2023-08-27 | 2.26x | 2.26x |
| 2023-09-24 | 2.25x | 2.25x |
| 2023-10-22 | 2.24x | 2.24x |
| 2023-11-19 | 2.49x | 2.49x |
| 2023-12-17 | 2.61x | 2.61x |
| 2024-01-14 | 2.68x | 2.68x |
| 2024-02-11 | 2.63x | 2.63x |
| 2024-03-10 | 3.22x | 3.22x |
| 2024-04-07 | 3.10x | 3.10x |
| 2024-05-05 | 3.10x | 3.10x |
| 2024-06-02 | 3.07x | 3.07x |
| 2024-06-30 | 3.15x | 3.15x |
| 2024-07-28 | 3.11x | 3.11x |
| 2024-08-25 | 3.18x | 3.18x |
| 2024-09-22 | 3.11x | 3.11x |
| 2024-10-20 | 3.04x | 3.04x |
| 2024-11-17 | 3.23x | 3.23x |
| 2024-12-15 | 3.59x | 3.59x |
| 2025-01-12 | 3.31x | 3.31x |
| 2025-02-09 | 3.45x | 3.45x |
| 2025-03-09 | 3.54x | 3.54x |
| 2025-04-06 | 3.54x | 3.54x |
| 2025-05-04 | 3.58x | 3.58x |
| 2025-06-01 | 3.69x | 3.69x |
| 2025-06-29 | 3.62x | 3.62x |
| 2025-07-27 | 3.94x | 3.94x |
| 2025-08-24 | 3.76x | 3.75x |
| 2025-09-21 | 3.63x | 3.61x |
| 2025-10-19 | 3.74x | 3.70x |
| 2025-11-16 | 3.89x | 3.83x |
| 2025-12-14 | 3.90x | 3.83x |
| 2026-01-11 | 3.93x | 3.84x |
| 2026-02-08 | 4.18x | 4.06x |
| 2026-03-08 | 4.22x | 4.03x |
| 2026-04-05 | 4.21x | 3.96x |
| 2026-05-03 | 4.31x | 3.96x |
| 2026-05-31 | 4.38x | 3.94x |
| 2026-06-28 | 4.67x | 4.11x |
| 2026-07-26 | 4.60x | 3.92x |
| 2026-08-23 | 5.06x | 4.29x |
| 2026-09-20 | 5.34x | 4.37x |
| 2026-09-22 | 5.40x | 4.42x |
The full scorecard for the live book's construction, split by window. The windows matter: 2021 to 2023 is where the design was chosen, 2024 was the first real check, and 2025 to 2026 is where the most recent choices were made, so it flatters them.
| Window | Sharpe | CAGR | Volatility | Max drawdown | Calmar | Worst month |
|---|---|---|---|---|---|---|
| 2021-23 (design) | 1.73 | 35.3% | 18.5% | 9.0% | 3.93 | -4.7% |
| 2024 (first check) | 0.69 | 10.8% | 17.1% | 10.8% | 0.99 | -6.8% |
| 2025-26 | 2.20 | 33.5% | 13.6% | 6.6% | 5.08 | -4.5% |
| 2021-26 (full) | 1.64 | 30.1% | 16.9% | 12.0% | 2.52 | -6.8% |
Sharpe here is computed from daily returns and annualized over 365 days, so it won't exactly equal CAGR divided by volatility.
- CAGR
- Max drawdown
Show data
| window | CAGR | Max drawdown |
|---|---|---|
| 2021-23 (design) | 35% | 9.0% |
| 2024 (first check) | 11% | 11% |
| 2025-26 | 34% | 6.6% |
| 2021-26 (full) | 30% | 12% |
The number I keep coming back to is 2024. Sharpe 0.69, Calmar 0.99: a real but unexciting year, and the cleanest out-of-sample read on the core. That's much closer to what I expect live than the full-period figures.
Over the full period that's a Sharpe of 1.64 and a CAGR of about 30%. I plan on a Sharpe of 1.0 to 1.1 live, and I expect the drawdown ladder to fire at some point: at a Sharpe of 1 and about 17% volatility, an unconstrained four-year drawdown is more likely around 20% than 12%. The parameters were chosen on the same history, my coin exclusions were picked while looking at recent data, and haircutting for every trial behind the book takes the 1.64 to roughly 0.8 to 0.9 (the deflated Sharpe probability that the result is real still comes out at 0.97 to 0.98). It also takes something like four years of live data to reliably tell a Sharpe of 1 from 0. I'd rather be pleasantly surprised.
How much of the 1.64 is hindsight
Some of the backtest is choices I made while looking at recent data, so I added them back one at a time on the core, at the live size, and watched each window:
| Core configuration (Sharpe) | 2021-23 | 2024 | 2025-26 | 2021-26 |
|---|---|---|---|---|
| The original core, full universe | 1.77 | 1.20 | 0.94 | 1.45 |
| + 63 exclusions (no memecoins, plus names I expect to lag) | 1.74 | 0.72 | 1.17 | 1.41 |
| + spreading half the freed risk back over kept coins | 1.74 | 0.67 | 1.19 | 1.41 |
| + five priority coins at 2x risk (live) | 1.73 | 0.69 | 1.36 | 1.44 |
The overrides add +0.42 in 2025 to 2026, the window I chose them in, and cost 0.51 in 2024, the year the excluded memecoins trended. Over the whole period they're a wash, 1.44 against 1.45. So I treat the exclusions as a risk preference, not as edge. The sleeves are the same story: each one was picked on the 5 to 14-month record it's scored on, and they're what lift 2025 to 2026 from 1.36 on the core alone to 2.20 for the book. That window is not a holdout for this configuration.
Deflating for every trial
Then the multiple-testing correction. If you run N strategies with no edge at all, the best one still shows a positive Sharpe by luck. Bailey and López de Prado's expected maximum is
SR₀ ≈ σ × [(1 − γ) Φ⁻¹(1 − 1/N) + γ Φ⁻¹(1 − 1/(N·e))]
where σ is how much Sharpe varies across the trials, γ ≈ 0.577 is the Euler-Mascheroni constant, and Φ⁻¹ is the inverse normal. The deflated Sharpe is the observed Sharpe minus that luck bar, and the deflated Sharpe ratio (DSR) is the probability the true Sharpe is above the bar, after correcting for the stream's skew and fat tails.
The inputs, all from the trial ledger I keep for this:
- N. 5,422 trials that touched the core's signal, 33,492 that touched the crypto book in any way, and 1,366,675 across everything (most of those are brute-force intraday scans).
- σ, two ways. 0.20, the measured spread of Sharpe across the core's own speed and regime variants, which is the paper's recommendation. And 0.42 = 1/√5.7 years, which treats every trial as an independent coin flip: too harsh for variants of one idea, right for unrelated ones.
- The book's stream. 2,092 days, skew 0.29, kurtosis 10.8.
- Deflated, variants of one idea
- Deflated, independent trials
Show data
| n | Deflated, variants of one idea | Deflated, independent trials |
|---|---|---|
| 1 | 1.64 | 1.64 |
| 10 | 1.32 | 0.98 |
| 100 | 1.13 | 0.58 |
| 226 | 1.08 | 0.47 |
| 5,422 | 0.89 | 0.09 |
| 33,492 | 0.81 | -0.09 |
| 1.37m | 0.65 | -0.42 |
Read the way the paper intends, against the core's own variants, the 1.64 deflates to 0.89 at 5,422 trials and 0.81 at 33,492, with a DSR of 0.98 and 0.97: the edge is very probably real, and much smaller than the backtest. Read the harsh way, it stays above a DSR of 0.95 only up to about 52 truly independent ideas, and the core alone (1.44) only up to 17. And 2024's 0.69 on its own isn't distinguishable from zero even before any deflation: one year of data gives it a standard error of 1.11.
Did I overfit the speeds? I tested that directly with the probability of backtest overfitting (PBO): take a 50-cell grid of calm pair × volatile pair × switch window, split the daily history into 16 blocks, and over all 12,870 half-and-half splits ask how often the in-sample winner lands in the bottom half out of sample. The answer is 74% of the time, and the in-sample winner's out-of-sample Sharpe falls from 1.42 to 1.00, below the grid's median of 1.10. So choosing speeds by backtest Sharpe would be overfitting. The live speeds weren't chosen that way: they're the in-sample winner in only 3% of splits and sit at the 73rd percentile out of sample, a reasonable member of a flat family. What does matter is the 7-day switch window (cells with it have a median Sharpe of 1.35, the 3-day window 1.11). My default answer to any "use slightly faster averages" idea is now no.
How much is timing luck? This is Hoffstein's question, and I measured it directly. If the core decided once a day instead of every hour, the result would depend on which hour, for no reason at all:
Show data
| clock | Sharpe |
|---|---|
| Daily 00 UTC | 1.27 |
| Daily 04 | 1.16 |
| Daily 08 | 1.31 |
| Daily 12 | 1.21 |
| Daily 16 | 1.29 |
| Daily 20 | 1.27 |
| Six-clock tranche | 1.27 |
| Hourly (live) | 1.45 |
Across the six daily clocks the full-sample Sharpe runs from 1.16 to 1.31, and on the most recent holdout from −0.19 to 1.51: the same rule, a negative or an excellent year depending only on the hour it looked. Hoffstein's fix, tranching across the six clocks, averages that luck away (1.275) but doesn't beat simply deciding every hour (1.453), because an hourly clock is 24 tranches. As a placebo, I also ran the hourly version at all 60 minute offsets (bars closing at :00, :01, through :59): mean 0.95, standard deviation 0.05, range 0.87 to 1.08. So any "improvement" from a different bar size smaller than about ±0.12 is indistinguishable from moving the clock a few minutes, and none of the 59 odd bar sizes I tried beat that band.
The same idea applies to parameters. Picking one speed pair from a flat family is a timing-luck bet on that pair. The next step I'm shadowing is an ensemble of 17 speeds instead of one pair: in the backtest it beats the live pair in every window, and its overfitting score (CSCV) falls from 0.84 to 0.015. About half of that gain is just smoother execution, which slicing today's orders over two to three hours also captures, so it runs on paper for a quarter before it gets any capital.
From backtest to plan
The live plan, worked from the inputs:
- Core. The live core's 2024 to 2026 mean at 1.86 units is 15.8% a year. Take out the priority tilt's 2024 to 2026 gain as hindsight (1.9 points): 13.8%. Haircut it to 70%: 9.7%.
- Sleeves. Their own-record mean at today's size is 28.5% a year. Count a quarter of it: 7.1%.
- Volatility. Core 14.8%, sleeves 5.4%, taken as independent (measured correlation −0.03): √(14.8² + 5.4²) ≈ 15.7%.
- Sharpe. (9.7 + 7.1) / 15.7 ≈ 1.07, or about 1.4% a month.
The optimistic version (core unhaircut, half the sleeves) gives about 1.8. I plan on 1.07.
Risk comes first
The book is sized backwards from a 15% drawdown budget, the row I picked from a frontier before choosing any sizes, with 20% as the hard stop. Each row is the live core levered to the largest size whose historical max drawdown stays inside the target on both the blind 2018 to 2019 window and 2021 to 2026. The probabilities come from 10,000 bootstrapped one-year paths (60-day blocks for the core, 30-day blocks for each sleeve), with every stream demeaned and then given the haircut plan's mean, at today's account size:
| Target max drawdown | Core units | Expected month (haircut) | Worst month seen | P(drawdown beyond target within 12 months, haircut) |
|---|---|---|---|---|
| 10% | 1.18 | +1.0% | −5.1% | 23% |
| 15% (live) | 1.86 | +1.3% | −7.9% | 25% |
| 20% | 2.59 | +1.7% | −11.0% | 28% |
| 25% | 3.32 | +2.0% | −14.0% | 30% |
| 30% | 4.11 | +2.4% | −17.2% | 32% |
Three things in that table shaped the choice. From the 10% row to the 30% row, the expected month grows 2.4x but the worst month grows 3.4x (the worst months all come from the blind 2018 to 2019 window). Every row from 20% up only works if I move my own 20% hard stop, because otherwise those books end at their first 20% drawdown. And the last column is not small anywhere: even at the live row there's about a one-in-four chance, in the haircut case, of going past 15% within a year (the chance of going past 1.5x the target, about 22%, is around 3%). The budget is what I size to, not a promise, which is why there's a hard stop behind it.
Gappy's drawdown arithmetic is the intuition behind caring so much about the true Sharpe. For a strategy with Sharpe SR, the chance of ever losing b units of annual volatility from where you stand is about exp(−2 · SR · b). At my plan of a Sharpe of about 1.07 and 15.7% volatility, the 20% hard stop is 20 / 15.7 ≈ 1.3 volatility units away, so exp(−2 × 1.07 × 1.3) ≈ 6% from a fresh start. At a Sharpe of 0.5 it's exp(−2 × 0.5 × 1.3) ≈ 27%. The risk of being stopped out is exponential in the Sharpe, which is why I size on the number I can defend, not the one I can backtest.
The controls, with their triggers
| Control | Trigger | What happens |
|---|---|---|
| Daily drawdown ladder | NAV closes 10% below its peak | Exposure halves |
| 20% below its peak | Flat, until I re-arm it by hand | |
| Intraday crash breaker | NAV 8% below its 24-hour high, two reads in a row | Exposure halves; back to full within 4% |
| 15% below | Half size for 15 minutes; if it recovers, no stop | |
| Still past 15% after 15 minutes, or 20% at any time | Flatten and block entries until I clear it | |
| No top-ups inside the hour | A coin falls between decisions | It keeps its quantity; nothing buys the fall back up to target |
| Exchange backstops | Every position | A reduce-only stop resting on the exchange at half the distance to liquidation, which survives my machine dying |
| Margin gate | Margin need above 80% of the balance | New entries are cut; leverage is never raised to make room |
| Gross caps | Always | Targets at most 2x equity; one coin at most 3.2x |
| Reject breaker | More than 10 refused orders in 5 minutes | Entries pause 15 minutes (exits never pause) |
| Stale data | Market data 60 s old / 300 s old | No new entries / close everything |
| Foreign order | Any resting order the system didn't place | The runner stops (I treat it as a key compromise) |
The ladder is daily and slow on purpose: halving on an intraday wiggle would sell every dip. The crash breaker exists for the days when "daily" is too slow, and the 15-minute confirmation exists because a flash crash that recovers shouldn't lock in its low.
The account also works like a small fund, because it holds two people's money: about $130k of mine and $50k of my brother's. A NAV index started at 1.000 on go-live, and every deposit buys units at the index of that moment. So the drawdown ladder and the crash breaker read the index, not the balance: a deposit is never "profit" and a withdrawal can never look like a crash and trip a stop.
I replayed the October 10, 2025 crash through the live runner at today's size. The book fell through the 15% line intraday, the 15-minute confirmation rode it out instead of flattening at the low, and the day ended down 5.4%. A breaker without that confirmation would have sold at the bottom and ended the day flat at −14.1%.
Regimes, capacity, and what would make me stop
Regime dependence. The core is a trend book, so it earns when something trends and gives some back when nothing does. On 2021 to 2026, its Sharpe was 1.67 in bull markets and 0.98 in bear markets, and it averaged +1.62% in BTC's up months against +0.55% in its down months. The blind 2018 test leans the other way (+1.65% a month in BTC's down months, positive 78% of the time). One bear market in each sample isn't enough to say which is the rule, so I plan on the weaker one: in a long sideways market the core bleeds slowly, and that is normal, not a signal to intervene.
Capacity. Charging square-root impact at each book size, the core's backtest Sharpe goes from 1.49 with no impact to 1.43 at $100k, 1.40 at $250k and 1.31 at $1m, and it keeps about half its planned return up to roughly $16m. So capacity isn't what limits this book; the small sleeves are capacity-bound long before the core is. The live fill check below only covers small orders (a median of $189), so it says the model is conservative at my size, not that it's right at a hundred times my size.
What would make me stop or cut. I wrote these down before going live, so I don't have to decide them in a drawdown:
- The ladder. 10% down halves, 20% down is flat, and nothing restarts until I've reviewed why.
- Retiring the core itself. Each threshold sits just beyond the worst the six-year backtest ever did, so a normal bad year can't trip it: four losing months in a row (the worst on record is three), more than 700 days under water (the worst is 635), or live cost per trade above twice the model over any 200 trend legs. I also had a rolling 365-day Sharpe below −0.5 as a kill, but the stress test showed it would already have fired on the backtest's own path in December 2025 (the six-year minimum is −0.52), so it's a review flag now, not a kill. A kill rule that fires on the healthy history is just noise with authority.
- Promoting a shadow sleeve. A sleeve logging trades on paper needs 40 events or 90 days, a Sharpe above 0.3 even at twice the modelled costs, and a correlation to the core below 0.3. Then it trades at a quarter size for another 40 events or 90 days before it gets full size.
- Per-system kills. Each sleeve has its own soft and hard levels on its own live record (half size, then zero), and re-arms only after 20 days.
- A monthly ramp review. It can cut any system at once on a weak record, but raise one by at most a quarter step.
- Pre-registered dates. One stock sleeve (a short on Korean chip perps) goes to zero at the end of December if the stock and commodity sleeves together earn less than half of plan. The oil trade goes to zero the day its fee discount ends.
- For the first three to six months, the checks are that the average monthly return sits around the 1.4% to 2.3% band, the drawdown stays short of the 10% halving line, and live costs track the model. A single month outside the band isn't a verdict: at about 15.7% annual volatility, one month's standard deviation is 15.7% / √12 ≈ 4.5%, so most individual months will land outside it. A pattern of them is a verdict.
Live versus backtest
The rule I'm strictest about: live trading must reproduce the research exactly. The live engine is a pure, replayable function of time, market data, and targets, so the same code runs in research and in production. Replaying the research history through the live code reproduces 9,133 of 9,133 trend legs and none extra (a separate trade-list replay matches 7,800 of 7,800 trades), with daily returns matching to within 1.8e-15 over 2,456 days, and the same Sharpe to four decimals. The tolerance is exact for trades and under 1e-9 for the daily series, and the check runs as a test.
The detail that makes that hard is that the book is path-dependent. A 1,920-hour moving average carries state from each coin's first bar, so the live book is warmed once over the full history, saved after every hour it processes, and reloaded on restart. After any gap, the missing hours are fetched and fed in order before it trades. It never skips a bar, because a skipped bar is a different moving average, and from then on a different book.
Every 60 seconds the runner reconciles its view of orders and positions against the exchange, and if they disagree, the exchange wins and that coin freezes until I understand why.
The first live fills also let me check the cost model:
- Modelled (bps)
- Live (bps)
Show data
| cls | Modelled (bps) | Live (bps) |
|---|---|---|
| all core | 6.80 | 1.12 |
| majors (BTC, ETH) | 1.81 | 0.15 |
| liquid alts | 5.78 | 1.76 |
| other alts | 12.98 | 1.53 |
Across 608 live fills, slippage came in at about 1.1 bps against the arrival mid, versus 6.8 bps modelled. It's a small sample, but it means the backtest was charging itself conservatively, which is the direction you want to be wrong in.
The detail, because "small sample" deserves numbers. I wrote the test and its decision rule down before computing anything: the model overstates cost if the upper end of the 95% interval on realized ÷ modelled is below 1. The sample is 608 fills in 337 parent orders from 18 hourly decisions, $216.5k filled, over September 26 to 28. The interval resamples whole decision hours together, because fills in the same hour aren't independent.
- Realized: 1.12 bps, 95% interval 0.36 to 1.58. Modelled: 6.80 bps. The ratio is 0.16 (interval 0.06 to 0.22), so by the rule, the model overstates.
- Fees, measured separately: 3.89 bps per fill live, against 4.05 in the model.
- Even the flat allowance alone overstates by about 4x; alts paid 1.5 bps all in against an 8 bp allowance.
- The shape is right, the slope isn't. Cost does rise with order size, but a regression of realized cost on the model's square-root term gives a slope of about 0.29 where the model assumes 1.
The honest limit is size: the median order was $189 and the 90th percentile $1.5k. Small immediate-or-cancel orders inside a price band barely touch the book. This says nothing yet about what a $50k order costs, and I've kept the conservative model in the backtest rather than "fixing" it to the live number.
What I took from it
Most of the value of this project was in the things I stopped believing. Copy-trading the best wallets doesn't work. High in-sample Sharpe is a warning sign, not a credential. Costs decide most intraday ideas before the signal does. And the thing that survived every test I could think of was the most boring idea on the list, held with discipline and sized so that being wrong is survivable.
That's also the work I want to keep doing: careful research, honest about luck, with an execution path I can trust to do exactly what the research says.