Portfolio risk: investment committee memo
The recommendation. Move the $100M mandate from 60/40 to the proposed five-asset mix: 45% US equities, 25% 10-year Treasuries, 12% gold, 13% real estate, and 5% Bitcoin. Over a full cycle it is expected to earn 8.1% a year against 7.2% for 60/40, a Sharpe ratio of 0.37 against 0.32, and a median of $145.5M after five years against $139.7M. It takes about one point more risk on every measure, and in a liquidity crisis its worst years run 4.1 points deeper than 60/40's. We recommend adopting it with two guardrails for that case. Every figure here is live on the Stage, where any allocation can be tried.
What the mix earns, and the risk it takes
Each figure comes from the same 10,000 simulated five-year paths, over the full cycle:
| Proposed mix | 60/40 | Change | |
|---|---|---|---|
| Expected return | 8.1% | 7.2% | +0.9 pts |
| Volatility | 11.1% | 10.1% | +1.0 pts |
| 1-year VaR, 95% | 10.5% | 9.6% | +0.9 pts |
| 1-year CVaR, 95% | 14.9% | 13.9% | +1.0 pts |
| Expected max drawdown, 5 years | 14.9% | 13.8% | +1.1 pts |
| Sharpe ratio | 0.37 | 0.32 | +0.05 |
| Median value after 5 years | $145.5M | $139.7M | +$5.8M |
Nine paths in ten end between $96M and $218M after five years. The extra return is paid for, not free: the mix adds return and risk in about the same measure, which is why its Sharpe ratio rises only a little.
How it behaves in each regime
A single covariance matrix hides what matters most to this committee: correlations change with the market. In expansions stocks and bonds move against each other (a correlation of −0.25), and Treasuries cushion an equity fall. In an inflationary rate shock that correlation turns to +0.60, and both fall together. In a liquidity crisis nearly everything sells off at once.
We measure tail risk as 1-year CVaR at 95%, the average loss in the worst year in twenty:
$$ \text{CVaR}{95%} = \mathbb{E}\left[,L \mid L \ge \text{VaR}{95%},\right] $$
| Regime | Proposed mix | 60/40 | Gap |
|---|---|---|---|
| Full cycle | 14.9% | 13.9% | +1.0 pts |
| Expansion | 9.8% | 8.9% | +0.9 pts |
| Rate shock | 30.4% | 30.9% | −0.5 pts |
| Liquidity crisis | 45.1% | 41.0% | +4.1 pts |
The mix holds up slightly better than 60/40 in a rate shock, where gold does not follow stocks and bonds down: its expected return stays at 1.1% a year against −0.1%, and its median after five years is $98.9M against $93.3M. In a liquidity crisis it does worse. Real estate and Bitcoin fall with equities, and its median after five years is $62.0M against $65.5M.
Where the risk comes from
Equities are 45% of the money and 60% of the risk, measured as each asset's share of the portfolio's variance. Bitcoin is 5% of the money and 16% of the risk; real estate is 13% and 17%. Treasuries and gold, 37% of the money between them, carry 7% of the risk.
The record
Over the ten years of daily returns from June 2016 to January 2026, $100 in the mix grew to $204, against $177 in 60/40, and its deepest fall was 21.5%, against 22.7%. It lost more than 60/40 in the first liquidity crisis (−4.3% against −2.9%, February to April 2020), and less in the 2023 rate shock (−17.1% against −19.3%).
Decisions for the committee
- Adopt the 45/25/12/13/5 mix in place of 60/40.
- Cap Bitcoin at 5%, rebalancing back whenever it drifts above, since it carries three times its weight in risk.
- De-risk in a crisis: move to the all-weather weights (30/35/20/15/0) when stocks and bonds start to fall together. Their 1-year CVaR in a liquidity crisis is 33.7%, against 45.1% for the proposed mix.
How the figures are made
Three scripts in the Model tab build every figure. The first assembles the daily return history across normal, rate-shock, and crisis regimes. The second estimates each regime's volatilities, correlations, and covariance matrix. The third simulates 10,000 monthly paths over five years with fat-tailed (Student-t, six degrees of freedom) shocks through each regime's covariance. The Stage and the deck run the same simulation, from the second script's matrices, for whatever allocation the reader chooses, so their figures can differ from the third script's tables by a few tenths of a point of sampling noise.