A disclosed, risk-budgeted paper book built like a multi-manager platform: three systematic desks and seven sector long/short desks run under one risk allocator, a 10% vol target, and an annual drawdown budget. The exhibit is the risk discipline, not a secret edge.
The Risk Museum: an autonomous AI pipeline tests ideas for trading strategies the way a quant desk would, and publishes the ones that fail its tests. New here? Start at the entrance →
- How to read it. Inception 2026-07-25; everything before is a transparent backtest (shaded and dashed on the curve below), and everything after accrues nightly, out-of-sample. Last valued 2026-09-04; backtest reconstructed from 2006-06-27.
Equity curve
Paper-book NAV, linear scale, starts at 100. Dashed blue is the transparent pre-inception backtest (2006 to 2026); the solid live track begins at the inception line on the right and accrues nightly, so the forward record is only just beginning.
Desk blotter
Risk first. Where the book sits against its own limits today.
The overlay runs a drawdown budget the way a platform runs a pod. Inside a calendar year the book cuts to half weight once it is 5.0% below the year’s high, and goes flat at 7.5%. The budget resets each January and the book trades again. So the within-year drawdown stays near 7.5%, while the all-time figure (-18.2%) can run past it across a long decline, when two annual budgets stack through the same bear.
A hard stop measured from the all-time peak would be a trap. A book taken to cash cannot climb back to an old high on its own, so it sits out the whole recovery. Harvey and co-authors document exactly this in “Drawdowns” (2020). Budgeting the limit by year is how a real desk holds the floor without killing the book.
What the overlay costs, plainly. Holding the traded max drawdown to -18.2% against the unmanaged -28.5% meant giving up return and time in market. The book sat in cash on about 17% of all days (~851), the price of the stop. Month by month the book keeps about 79% of the unmanaged book’s up months and takes only 71% of its down months, so the overlay sheds more downside than upside.
Diversification
Why ten desks under one roof run at lower risk than any desk alone. Add up each sleeve’s own volatility, weighted by how much of the book it carries, and you get 18.0%. The combined book realizes just 7.7%, because the desks do not move together. That gap is the diversification a multi-manager platform is built to harvest, and the whole case for running many uncorrelated sleeves instead of one big bet.
Pairwise correlation of the ten desks’ daily returns. Red = move together, blue = move opposite, stronger colour = stronger link. Low-vol / BAB is -0.66 to equity beta, a built-in hedge that fires exactly when beta hurts. The seven sector desks share one momentum factor and cluster in low-to-mid red, which is why the allocator budgets them as a single group and never lets the cluster swamp the book. Uncorrelated sleeves are the engine; the 2.35× ratio is what they buy.
The Holy Grail, measured
Ray Dalio’s best known chart says fifteen to twenty uncorrelated return streams cut portfolio risk by about 80% without giving up expected return. The arithmetic is real, and the dashed blue line below is that arithmetic: at zero correlation, risk falls as one over the square root of the stream count, touching the 80% mark near 25 streams. The solid line is this book, computed from the measured matrix above. The ten desks average +0.21 pairwise, so ten desks deliver the diversification of 3.5 truly uncorrelated ones and cut risk 46% from a single desk.
The dotted continuation is the trap: adding more desks at the same average correlation flattens into the correlation floor at 46% of one desk’s risk and never reaches the ideal curve’s 20%. The x axis of Dalio’s chart is the hard part: uncorrelated streams are the scarce input, not the count. Idealized curve after Dalio, Principles (the Holy Grail of investing); measured curve from the desk correlation matrix above, recomputed nightly.
Risk dashboard
Computed at render from the book’s daily returns and the Fama-French factor data. These are the tilts the book means to carry, and the tail it accepts.
| Desk | Risk weight | Std. Sharpe | Contribution |
|---|---|---|---|
| Systematic desks | |||
| Equity beta (SPY) | 20.2% | +0.64 | +63.1% |
| Low-vol / BAB | 9.6% | -0.69 | -41.1% |
| Trend (TSMOM) | 20.2% | +0.78 | +59.4% |
| Systematic subtotal | 50.0% | +81.4% | |
| Sector long/short desks | |||
| TMT L/S | 2.0% | +0.04 | +4.2% |
| Financials L/S | 6.0% | -0.23 | +3.4% |
| Healthcare L/S | 8.0% | -0.40 | -10.9% |
| Industrials L/S | 7.1% | -0.13 | +4.5% |
| Energy & Materials L/S | 6.9% | +0.00 | +8.1% |
| Consumer L/S | 8.5% | -0.07 | +4.7% |
| Utilities & Real Estate L/S | 11.6% | -0.19 | +4.7% |
| Sector subtotal | 50.0% | +18.7% | |
| Commodities | roadmap | ||
| Global macro (rates, FX) | roadmap | ||
| Credit | roadmap | ||
| Book (ex-ante, diversified) | 100.0% | +0.40 | 100.0% |
The platform splits risk 50.0% to the systematic desks and 50.0% to the sector long/short desks, then runs inverse-vol (risk-parity) weighting within each group, so no desk can swamp the book. Construction, stated in the terms a portfolio manager would ask for: cross-sectional long/short sleeves built from point-in-time signals and rebalanced monthly, equal-weight within leg so no single name carries the sleeve; sleeves combined by inverse-vol on trailing 60-day realised volatility; a 10% annualised vol target; a 2.0x gross cap; and a pod-style drawdown ladder that halves exposure at a declared threshold rather than at whatever felt right on the day. Every weight is decided on data through day t and applied to day t+1. Contribution is each desk’s share of the book’s return and sums to 100%. The directional and trend desks carry the return, while the sector long/short desks earn little on their own but decorrelate the book, which is why its SPY correlation is only 0.32. Standalone Sharpes are backtest figures on a survivorship-biased universe, so they are upper bounds. Commodities, global macro, and credit are real platform desks that need data outside this US-equity universe, shown as roadmap rather than faked.
| Confidence | VaR | Exp. shortfall |
|---|---|---|
| 1-day 99% | -1.22% | -1.60% |
| 1-day 95% | -0.73% | -1.06% |
Empirical (non-parametric) 1-day figures over 5,079 daily book returns; a loss, not a forecast. Expected shortfall is the average loss in the tail beyond VaR.
| Factor | Beta | t-stat |
|---|---|---|
| Mkt-RF | +0.15 | +39.9 |
| SMB | -0.04 | -5.5 |
| HML | +0.03 | +4.5 |
| RMW | -0.02 | -1.6 |
| CMA | -0.06 | -4.7 |
| Mom | +0.24 | +53.6 |
| Alpha (ann.) | -1.2% | -1.1 |
| R² | 0.47 |
Excess daily book return regressed on the Fama-French 5 factors plus momentum (Ken French daily data), 5,054 overlapping days, 2006-06-28 to 2026-07-31. The French factors publish on a lag, so this regression ends 2026-07-31 while everything else on the floor is as-of 2026-09-04. The book is what it says it is: a long market beta (Mkt-RF +0.15) and a trend/momentum tilt (Mom +0.24), with a large-cap tilt (negative SMB, since this is a big-cap universe) the only other significant loading; value, profitability and investment betas are statistical zeros. R² 0.47 leaves most of the variance idiosyncratic. SPY correlation 0.32 corroborates the market beta. The annualized alpha is negative and insignificant (t -1.1). On a survivorship-biased universe, treat it as an upper bound.