Industry: Banking and Financial Services Product: ModelRisk Application: Risk budgeting, marginal risk contributions, and stress-test oversight
The risk-oversight function at a $5B institutional mandate inherited an allocation from the portfolio-management team that, by capital weight, looked balanced — 32% US Equity, 18% International, 10% EM, 30% Fixed Income, 10% Alternatives. The same allocation, viewed through risk contributions instead of capital weights, was anything but balanced: US Equity contributed roughly 47% of the portfolio's total risk, the three equity buckets together contributed about 92%, and the entire fixed-income allocation contributed barely 2%. Risk oversight is the function that converts that observation into a constraint the PM team can actually trade against, and the Monte Carlo simulation is the engine that makes the constraint quantifiable rather than rhetorical.
Portfolio risk does not aggregate the way capital does. With weights \(w\) and covariance \(\Sigma\), the portfolio volatility is \(\sigma_p = \sqrt{w^\top \Sigma w}\), and the marginal contribution to total risk from asset \(i\) is \((\Sigma w)_i / \sigma_p\). The asset's risk contribution is \(w_i \cdot (\Sigma w)_i / \sigma_p\), and the sum of risk contributions across assets equals total portfolio vol — a tidy decomposition the analyst can read line by line.
For the inherited allocation, with annualized return / vol assumptions on the six assets and a realistic post-2010 correlation matrix, the portfolio vol comes out near 10.6%/year. The risk-budget chart shows where that 10.6% actually comes from.
The pattern is structural to any equity-heavy allocation: each percentage point of equity exposure carries far more risk than each percentage point of bond exposure, and the cross-asset correlations within the equity bucket amplify it. The 30% of capital in fixed income contributes essentially nothing — both because its standalone vol is low and because its negative-to-near-zero correlation with equities makes it a net hedge inside the portfolio.
The risk-budget decomposition is the micro view. The macro view is the annual P&L distribution at the AUM level — what the board sees.
At $5B AUM, the simulation shows a mean annual P&L of roughly +$256M (the 5.1% expected real return on the portfolio), a VaR 95% loss of roughly $620M, a VaR 99% loss of roughly $980M, and an Expected Shortfall 99% near $1.15B. The probability of an annual loss exceeding $500M (the board's 10%-of-AUM appetite threshold) sits around 8% — small in isolation, but a one-in-twelve event the appetite statement had never quantified. The Expected Shortfall conditional on a tail event runs roughly 16% above VaR99, the structural feature of an equity-heavy portfolio whose tail is dominated by correlated drawdowns rather than independent ones.
The board's risk-appetite document had been written in capital-loss terms ("no scenario exceeds 10% of AUM"). The simulation translates that into a quantifiable probability — and the Expected Shortfall conditional on a breach gives the committee a defensible answer to "how bad is bad." Both numbers are now reported quarterly.
VaR and ES are good for the central distribution. Historical stress scenarios are how the risk function tests the portfolio against named events the board recognizes by name.
A 2008 GFC repeat would cost the portfolio roughly $1.23B (25% of AUM) — well in excess of even the ES99 number. A 2020 COVID Q1 repeat costs about $650M (13% of AUM); the 2022 rates shock is worse still at roughly $800M (16%) — the most informative scenario because it is the most recent and the one for which the post-event PM-team narrative ("nothing to hedge against") needed direct contradiction. The 1998 LTCM analog ($530M) and the milder 1994 bond crisis and 2011 EU sovereign analogs (each around $120M) cost less, but each highlights a different correlation that the average-case model under-prices. The risk function now publishes the stress panel alongside VaR99 in every quarterly report; the contrast between the ~$980M VaR99 and the $1.23B GFC analog is the single most-asked-about number in the committee meetings.
The risk function's leverage comes from showing the PM team not just where the risk sits but what it would cost to move it. A tornado on VaR99, with each driver corresponding to a 5-percentage-point trim of one asset into the US Govt bucket, makes the trade-offs concrete.
Trimming 5pp of EM Equity or US Equity into US Govt cuts VaR99 by the largest amounts — roughly $110M and $108M respectively, a near dead heat. EM Equity edges ahead per percentage point because of its outsized standalone vol, while US Equity matches it on the strength of its much larger starting weight. Trimming International Equity is next at about $60M. Trimming US IG Bonds returns only about $27M of VaR99 — and at a measurable cost in expected return — so it would be the worst defensive trade the team could make.
The parametric alternative — variance-covariance VaR computed from the same mu/cov inputs — is cheaper but is wrong in two specific ways. First, it assumes joint Normality, which the historical stress panel makes visibly false: the 2008 GFC return vector is not a Normal draw from the calibrated covariance, it is a tail event whose magnitudes the parametric formula has no representation for. Second, it cannot directly produce conditional metrics like Expected Shortfall in any form the board can audit. Simulation produces both naturally and produces the same numbers each rerun.
The risk-oversight rebuild had three operational consequences:
The boundary between portfolio management and risk management gets crossed every time the question changes from "what should we hold?" to "how should we measure what we hold?" Monte Carlo provides the answer in a form both sides of that boundary can read off the same chart.