Expected Shortfall: Why VaR Doesn't Tell the Whole Story

Expected Shortfall: Why VaR Doesn't Tell the Whole Story

What if the number your risk system reports every morning misses the scenarios that could actually destroy you? For decades, Value at Risk served as the standard answer to "how much could we lose?" But VaR answers a narrow version of that question: it identifies the loss threshold at a given confidence level, then goes silent about everything beyond it. A 99% daily VaR of $5 million means losses exceed that figure roughly 2.5 times per year. Whether those breaches average $5.5 million or $50 million, VaR treats them identically.

In 2002, Carlo Acerbi and Dirk Tasche published a proof that reframed this silence as a structural deficiency. Their paper, "On the Coherence of Expected Shortfall," demonstrated that Expected Shortfall (ES) satisfies all four mathematical axioms required of a coherent risk measure, including for discrete distributions where earlier proofs had left gaps. The result gave regulators the theoretical foundation to replace VaR with ES in global banking standards.

The Coherence Gap in VaR

Artzner et al. (1999) had already established four axioms that any sensible risk measure should satisfy: translation invariance, subadditivity, positive homogeneity, and monotonicity. VaR passes three but fails subadditivity, the axiom encoding the principle that diversification should not increase risk.

The failure is concrete. Consider two trading desks, each holding concentrated positions. Desk A's 99% VaR is $2.1 million. Desk B's is $1.8 million. When the firm aggregates both books, the combined 99% VaR can exceed $4.3 million, more than the $3.9 million sum of parts. Under VaR, merging positions can make diversification appear to increase risk. This pathology arises because VaR examines only a single quantile; reshuffling the probability mass beyond that point changes nothing in the reported number, even when it transforms the actual tail.

What Acerbi and Tasche Proved

The contribution of Acerbi and Tasche (2002) was to close a technical gap. Artzner et al. had shown that ES is coherent for continuous distributions, but many real-world loss distributions are discrete (finite scenario sets, historical simulation with limited data). Acerbi and Tasche proved that ES remains coherent across all distribution types, continuous, discrete, and mixed, by establishing subadditivity through a representation involving conditional expectations rather than simple tail averages.

This mattered because practitioners had raised legitimate objections: if ES only worked under idealized distributional assumptions, its theoretical superiority over VaR was academic. The 2002 proof eliminated that objection entirely. ES is coherent under any probability distribution that a risk system might encounter.

Separately, Acerbi (2002) extended the result to spectral risk measures, an entire class where ES is the simplest member. Spectral measures weight tail losses by a non-decreasing function, letting risk managers express varying degrees of aversion to extreme outcomes.

Where VaR and ES Diverge in Practice

The practical gap between VaR and ES depends on the fatness of the tails. Under a normal distribution, 95% ES is roughly 1.28 times the 95% VaR, a modest difference. Under fat-tailed distributions common in credit, commodity, and equity markets during stress periods, the ratio widens dramatically.

When the ES-to-VaR ratio climbs above 2.0, it signals that the loss distribution has meaningfully fatter tails than the normal assumption implies. In such regimes, VaR underreports the severity of the very events that pose existential risk to portfolios. As Yamai and Yoshiba (2005) documented, institutions relying solely on VaR during periods of elevated tail risk systematically underestimated their capital needs.

This divergence connects directly to the concern that drives tail risk hedging strategies: the losses that matter most for portfolio survival are precisely the ones VaR discards. A hedging program designed around VaR thresholds can leave the portfolio exposed to the drawdowns that actually threaten solvency.

The Regulatory Consequence

The Basel Committee on Banking Supervision adopted 97.5% Expected Shortfall as the primary market risk capital measure in the Fundamental Review of the Trading Book (2019), replacing the 99% VaR standard from Basel II. The choice of 97.5% was deliberate: under normality, 97.5% ES approximately equals 99% VaR in magnitude, ensuring rough continuity in capital levels while upgrading the measure's mathematical properties.

The shift closed a specific regulatory arbitrage. Under VaR-based capital rules, a trader could sell deep out-of-the-money options that produced steady premium income without affecting reported VaR. The catastrophic losses from those options, triggered only in rare tail events, sat entirely beyond VaR's horizon. ES captures those losses by construction, since it averages across the entire tail. Embrechts, McNeil, and Straumann (2002) had warned about precisely this class of vulnerability in correlation-dependent portfolios.

Limitations Worth Acknowledging

ES is not without trade-offs. Backtesting ES is harder than backtesting VaR because verifying an average requires more data points than verifying a threshold breach. Tasche (2002) discussed elicitability concerns, the property that a risk measure can be meaningfully backtested using scoring functions. ES lacks elicitability on its own, though jointly with VaR it becomes elicitable, a result that has informed the FRTB's hybrid approach of using ES for capital and VaR for backtesting.

ES is also more sensitive to estimation error in the far tail. With limited historical data, the average of the worst 2.5% of observations carries substantial sampling uncertainty. This is not a reason to revert to VaR, but it means ES estimates should be accompanied by confidence intervals or supplemented with stress scenarios.

Connecting the Framework

For investors managing risk premia across asset classes, the choice of risk measure shapes portfolio construction decisions. A maximum drawdown framework captures path-dependent worst cases, while ES captures distributional tail severity at a point in time. These perspectives complement rather than substitute for each other. The institution that monitors VaR for day-to-day threshold management, ES for capital allocation and tail awareness, and maximum drawdown for cumulative path risk has a substantially more complete picture than one relying on any single metric.

Acerbi and Tasche's proof established that among these tools, Expected Shortfall is the one that respects the mathematics of diversification while revealing what happens in the scenarios that matter most.