It settles interest three times a day, trades straight through the weekend, and has never once rung a closing bell. Most of what is known about markets quietly assumes one.
Edgeworth Research tests what survives contact with this market — and publishes what fails.
Every pair of the 140-instrument universe, ordered by sector so clusters form blocks on the diagonal. The blocks are real — sectors do move together. The problem is everything off them. Cross-sector correlation sits near 0.36, and majors carry 0.58 against the whole universe. A single factor dominates.
This is why the cross-sectional sleeve failed. Ranking instruments against each other only pays when the spread between them is wide enough to cover turnover. When one factor explains most of the variance, that spread is not there.
A universe of 140 instruments holding roughly three independent bets. What disciplined sizing of three bets returns, net, is the next table.
| Walk-forward | Live | |
|---|---|---|
| Sharpe | 1.74 | 1.61 |
| Sortino | 2.38 | 2.11 |
| CAGR | 9.9% | 8.4% |
| Volatility | 5.7% | 5.2% |
| Max drawdown | −4.0% | −2.6% |
| Skew | +0.31 | +0.19 |
| Deflated Sharpe | 1.42 | — |
| Trials adjusted for | 61 | — |
Monthly net returns, %. Composite includes all discretionary-free accounts managed to the systematic book. Past performance is not indicative of future results.
None of these figures existed until they survived a set of constraints. The constraints come first — they are below.
Every limit below is enforced in the execution layer, not in a policy document. An order that would breach one is rejected at submission — the book cannot be talked into an exception on a bad afternoon.
The drawdown governor is mechanical. At −4% peak-to-trough the book halves risk; at −7% it goes flat and does not re-enter until the research committee signs off in writing. It has triggered twice since 2019.
| Event | Window | BTC peak-to-trough | Book P&L | Book drawdown |
|---|---|---|---|---|
| COVID liquidation cascade | 12–13 Mar 2020 | −49.7% | −1.8% | −2.2% |
| Leverage flush | 19 May 2021 | −30.2% | −0.9% | −1.4% |
| Terra / UST collapse | 9–13 May 2022 | −22.4% | +0.6% | −0.8% |
| FTX insolvency | 8–11 Nov 2022 | −25.8% | −1.1% | −1.9% |
| Yen carry unwind | 5 Aug 2024 | −17.6% | −0.4% | −0.7% |
Replayed on point-in-time data with the venue outages, funding spikes and widened spreads that actually occurred, including the venues that halted withdrawals. A stress test run on today's clean tape is not a stress test.
What clears the limits gets written up in full — the wins, and more unusually, the losses. The most recent write-up is a loss.
We allocated 15% of the book to cross-sectional momentum on the strength of a walk-forward study. Seven months of live forward testing returned −15.7% on that sleeve against −0.4% for the rest of the book.
This paper sets out what the backtest showed, what the forward period showed, why we think the two disagree, and what we changed.
Read the paper →Every signal in these papers comes out of one small model family. Small is a deliberate choice, argued next.
Twenty-two features per instrument per bar — returns across three horizons, realised volatility, funding and its z-score, open interest change, orderbook imbalance and depth, taker flow, liquidations on both sides, perp–spot basis, and cross-sectional rank.
Three heads rather than one: expected return μ, its uncertainty σ, and the probability of sign. Position size is a function of all three, so a confident small edge can outrank a noisy large one.
Deliberately small. A network with more parameters than the data can support will fit the noise, and the walk-forward will not catch it if the refit cadence is wrong. Every architecture change is treated as a new trial and counted in N.
A model is an opinion about the market. The battery below is how an opinion earns the right to hold positions.
A reported Sharpe ratio is inflated by the number of configurations tried before it. We publish the deflated statistic, which corrects the observed maximum for the trial count N and for the non-normality of the return series.
Bailey & López de Prado (2014). γ₃, γ₄ are the skew and kurtosis of realised returns; SR0 is the expected maximum Sharpe under the null across N independent trials. Reported alongside the probability of backtest overfitting.
Most backtests run on the instruments that exist today. We rebuild the tradeable universe as it stood on each date, including the listings that later failed — the ones carrying the losses.
Funding settles three times a day and is routinely larger than the edge. It is applied from realised rates rather than an average, and every result is re-run at double the slippage assumption.
Each run writes its configuration, commit hash and trade log to an immutable artefact. Any published figure can be regenerated from the record years later.
| Test | What it catches | Pass criterion |
|---|---|---|
| Poison-future sentinel | Look-ahead leakage | History identical under truncation |
| Purged & embargoed CV | Label-overlap leakage | Embargo ≥ label horizon |
| Shuffled-date calibration | Pipeline false-positive rate | ≈5% on randomised dates |
| Deflated Sharpe | Selection under multiple testing | DSR > 0 at 95% |
| Probability of backtest overfitting | Configuration overfit | PBO < 0.30 |
| Stationary block bootstrap | Autocorrelated, overlapping samples | 95% CI excludes zero |
| Newey–West HAC errors | Serial correlation in t-statistics | lag = ⌈4(T/100)2/9⌉ |
| Benjamini–Hochberg | Multiple comparisons across the set | FDR ≤ 0.10 |
| Trade-sequence Monte Carlo | Path dependence, drawdown luck | 5th pct drawdown within mandate |
| Cost sensitivity | Fragile execution assumptions | Survives 2× slippage |
| Regime partition | Single-era dependence | Positive in ≥3 of 4 regimes |
| Parameter surface | Search artefacts | Plateau, not spike |
A result that fails any row does not reach the book. It is written up instead — the failures are the part of the record most firms discard, and the part a reader can actually learn from.
Every idea that reaches the battery is registered before it runs — hypothesis, universe, parameter ranges — and the trial is counted whether it survives or not. The count is N in the deflated Sharpe: the more we try, the higher the bar for what we keep.
Of 214 registered trials, 13 hold positions today. The other 201 are not deleted; they are the archive's most useful material, and several papers below are their post-mortems.
Twelve gates and 201 rejections presume one thing: a tape that can be trusted. That is built, not bought.
Vendor bars are revised, back-filled and silently corrected. We capture our own from venue websockets, store every message with the timestamp it arrived rather than the timestamp it claims, and never overwrite history — a correction is a new record, not an edit. Any figure in any paper can be recomputed against the tape exactly as it stood on the day.
Nothing reaches the live book without clearing the paper book first, on the same code path. The research cluster and the execution stack share one implementation of every feature — a signal cannot mean one thing in a backtest and another at 3am.
That is the machine. What follows is the kind of question it was built to answer — try this one by hand first.
A strategy wins 55% of the time, loses 45%, and pays even money. Two traders each start with $10,000 and run it for 1,000 independent rounds. Ana stakes 20% of her bankroll every round; Ben stakes 10%.
Ana has the higher expected wealth after 1,000 rounds. Ben is far more likely to finish ahead of her. Explain how both statements are true, and find the staking fraction that maximises the median outcome.
Solutions to problems@edgeworthresearch.com — we publish the best one, with attribution.
Everyone here has sat with problems like this one — first at desks where being wrong was expensive.
Research is carried out by people who built and ran systematic strategies inside banks and market-making firms before doing it here.
We do not publish individual biographies. The work is meant to stand on its method and its data, both of which are released with every paper. Subscribers and counterparties are introduced to the researchers directly.
Monthly NAV struck and published independently of the manager.
Statements audited each year; the composite is separately examined.
Client assets held away from the manager and away from trading venues.
No venue holds more than 40% of posted margin at any time.
Each mandate ring-fenced; no cross-liability between books.
Holdings, exposures and P&L attribution delivered on a fixed calendar.
Counterparty names, service agreements and the operational due-diligence questionnaire are released under NDA to allocators at the diligence stage.
Institutional subscriptions include the full archive, the underlying data and specification for every published result, and the configurations that were not selected. A fixed number of individual seats is released each year.
No newsletter, no sales sequence. You will receive the papers.