Value-at-Risk earned its place measuring equity and rates portfolios. Commodity books are a different animal - and the measure you choose changes what your desk sees.
The distribution is the problem
VaR assumes, implicitly, that the tail beyond the quantile is somebody else's problem. In equity and rates markets that approximation is defensible. In commodities it is not. Physical supply constraints, seasonal demand, storage optionality and geopolitical disruption produce price distributions that are fat-tailed and markedly skewed - negative price events (WTI settling below zero in April 2020), seasonal squeezes and regime breaks are structural features, not anomalies. A 99% VaR computed on such a distribution answers a narrow question: "how bad is a fairly normal bad day?" It is silent on the loss you actually eat when the tail arrives.
Expected Shortfall: the better default, with caveats
Expected Shortfall answers the question VaR dodges - the average loss in the worst cases - and it is coherent: sub-additive, so diversification can never increase measured risk. That property matters operationally, because desks aggregate books, and a risk measure that punishes aggregation distorts hedging decisions. The caveats are estimation and validation: ES deep in the tail is estimated from few observations, and it is not elicitable, so backtesting requires the newer generation of tests - ridge, regression-based, exceedance-based - rather than the simple traffic-light counts VaR enjoyed.
Spectral risk measures: tuning the tail to your appetite
ES treats every loss beyond the quantile equally. Spectral risk measures generalise it: they weight the tail with a deliberately chosen risk-aversion function, producing a family of coherent measures that interpolate between VaR-like and maximally tail-punishing views. For a firm, this is the bridge between regulation and appetite - a spectral measure can encode "we weight the worst 1% five times as heavily as the 95th percentile" in a single, coherent, aggregable number. Robust estimation of these measures on commodity data is well studied (Homescu, 2014), and scenario-based evaluation methods (Wang & Ziegel, 2021) supply the validation machinery.
What we recommend in practice
- One simulation stack, many measures. VaR, ES and spectral measures should be read off the same Monte Carlo paths - differences of measure, not of data or engine - so the numbers reconcile by construction.
- Stress the estimation, not just the measure. On commodity books the estimation window matters as much as the confidence level; calibrate to stressed periods and test stability across windows.
- Adopt modern ES backtests before the supervisor asks. The tools exist; deploying them early turns model validation from a scramble into a routine.