Conditional Value at Risk (CVaR): Definition, Formula & Calculation

What is Value at Risk (VaR)?

Value at Risk is the most widely cited risk metric in institutional finance, and for good reason: VaR collapses a complex return distribution into a single, interpretable number. At its core, VaR answers one question — what is the maximum loss I should expect over a given time horizon, at a given level of confidence?

The standard expression is direct. A 95% 1-day VaR of €50,000 means that on any given trading day, you should expect losses to exceed €50,000 only 5% of the time. Equivalently, 95% of trading days should produce a loss smaller than €50,000, or a gain. For a portfolio manager, this is an immediately useful figure: it sets a threshold against which daily P&L can be measured, stress tested, and reported to risk committees.

VaR is also computationally tractable. VaR can be derived analytically under distributional assumptions, estimated via historical simulation, or approximated through Monte Carlo methods — each approach with its own trade-offs between speed, accuracy, and model dependence. For a broad audience of risk managers and regulators, VaR became the standard precisely because VaR translates probabilistic uncertainty into a single currency-denominated loss figure.

What VaR does not tell you, however, is equally important — and for many portfolios, more important than what it does.

VaR's Fatal Blind Spot

VaR defines a threshold. VaR says that losses will exceed €50,000 on roughly 5% of trading days. What VaR is entirely silent on is what those exceedances look like. When losses breach that threshold — and they will, systematically and repeatedly — are they €51,000 or €500,000? VaR cannot say. Both outcomes produce an identical VaR figure of €50,000.

This silence is not a minor technical limitation. It is a structural deficiency that has contributed to catastrophic risk management failures in practice. Consider the 2008 financial crisis. In the years leading up to it, major investment banks published VaR disclosures suggesting their trading book risk was moderate and contained. Those figures were not fabricated — under the distributional assumptions embedded in their models, the numbers were internally consistent. What the models failed to capture was the tail behaviour of the assets involved. Mortgage-backed securities and the derivatives written on them had return distributions with extreme negative skew and kurtosis far beyond anything a normal distribution would predict. The losses that materialised in September and October 2008 were not in the 5% tail assumed by VaR — they were in a region the models had assigned near-zero probability.

The deeper issue is that financial returns are not normally distributed. Equity market returns in particular exhibit excess kurtosis — what practitioners call fat tails. In a normal distribution, a five-standard-deviation loss event should occur approximately once every 14,000 years. In practice, equity markets produce such events roughly once per decade. The 1987 single-day crash of 22.6% in the Dow Jones Industrial Average was, under normal-distribution assumptions, a 25-standard-deviation event. The October 2008 drawdown in the S&P 500 was similarly described in statistical terms that made it sound almost astronomically improbable — yet it happened within living professional memory for most institutional investors.

Fat tails are not anomalies. Fat tails are the normal operating environment of financial markets. A risk metric that assumes otherwise — or that simply stops measuring at the threshold — is not a complete picture of portfolio risk.

Under normality, a five-standard-deviation loss event should occur once every 14,000 years. Equity markets produce such events roughly once per decade.

The structural deficiency of VaR — A.L. Capital Advisory

What is CVaR (Conditional Value at Risk / Expected Shortfall)?

Conditional Value at Risk — often called Expected Shortfall, and abbreviated as CVaR or ES interchangeably — addresses VaR's blind spot directly. Where VaR defines a threshold, CVaR asks what happens beyond it. Formally, CVaR at confidence level α is the expected loss given that the loss exceeds the VaR threshold. It is the average of the worst (1 − α)% of outcomes.

At a 95% confidence level, CVaR is the average of the worst 5% of daily return observations. At 99% confidence, it is the average of the worst 1%. The threshold is the same as VaR — what changes is that CVaR does not stop measuring once that threshold is crossed. It tells you what the average experience looks like in the portion of the distribution that VaR silently discards.

A worked numerical example makes the distinction clear. Consider two portfolios — Portfolio A and Portfolio B — each with a 95% 1-day VaR of €50,000. From VaR alone, these portfolios appear to carry identical risk. Now examine what happens in the worst 5% of days:

This is not a hypothetical edge case. Concentrated equity portfolios, leveraged positions, and portfolios with significant options exposure routinely produce tail profiles that look similar to Portfolio B when measured against their VaR figure. The €50,000 threshold is accurate — losses exceed it 5% of the time, as stated — but the experience of those exceedances is materially more severe than the VaR headline implies.

Why CVaR is Better for Portfolio Construction

Beyond its descriptive advantages, CVaR has a formally superior theoretical foundation for portfolio optimisation. Artzner, Delbaen, Eber, and Heath's 1999 paper on coherent risk measures established that a risk measure must satisfy four axioms to be considered internally consistent: translation invariance, monotonicity, positive homogeneity, and — critically — subadditivity.

Subadditivity means that the risk of a combined portfolio should be no greater than the sum of the risks of its individual components. Mathematically: CVaR(A + B) ≤ CVaR(A) + CVaR(B). This is a formalisation of the intuition that diversification reduces risk. CVaR satisfies this property. VaR does not — under certain return distributions, VaR can actually increase when two positions are combined, which is mathematically incoherent as a basis for portfolio construction.

The practical consequence is that CVaR-optimal portfolios are generally better diversified and more resilient in genuine tail events than VaR-optimal portfolios. When you minimise CVaR across the portfolio, you are explicitly penalising constructions that produce severe losses in the worst-case scenarios. A VaR-optimal portfolio might achieve its target threshold by concentrating in assets with low everyday volatility but extreme tail exposure — precisely the type of risk that went unmeasured in pre-crisis bank trading books. A CVaR-optimal portfolio, by construction, cannot tolerate that trade-off.

CVaR also captures tail concentration risk in a way that volatility-based measures miss. A portfolio might display identical annualised volatility to a well-diversified benchmark — but if that volatility is driven by a small number of highly correlated, high-beta positions, the CVaR will be materially higher. The measure does not distinguish between a 12% annualised volatility achieved through broad diversification and one achieved through two or three leveraged stock bets. CVaR does. It surfaces exactly the kind of structural concentration that creates catastrophic drawdown potential when those specific positions move adversely together.

For this reason, the Basel III and Basel IV regulatory frameworks — the international standards governing bank capital adequacy — have moved from VaR to Expected Shortfall as the primary internal model risk measure. The regulator's reasoning mirrors the practitioner argument: in a stress environment, what matters is not where the loss threshold sits, but how severe the losses beyond it can become.

CVaR in the Asset Lens Platform

The CVaR figure displayed in Asset Lens is computed using historical simulation on each asset's return series. Historical simulation makes no distributional assumptions — it does not impose normality or any other parametric form. Instead, it uses the actual observed return history of the asset: each historical daily return is treated as an equally probable outcome, and the CVaR is derived directly from that empirical distribution.

Historical simulation has a clear advantage for assets with fat-tailed or skewed return distributions: the tail behaviour is taken from the data itself, not from a theoretical model. If a stock has historically exhibited severe negative skewness — as many individual equities do, with occasional large down-moves that are not mirrored by equivalent up-moves — that asymmetry is preserved in the CVaR estimate rather than being smoothed away by a symmetric distribution assumption.

The displayed CVaR figure represents the expected loss at the 95% confidence level over a one-month horizon. It is expressed as a percentage of asset value. An Asset Lens CVaR of 14% on a given equity position means that in the worst 5% of monthly return observations in that asset's history, the average loss was 14% of position value. This translates directly to portfolio-level tail risk analysis: a position sized at €100,000 with a CVaR of 14% contributes an expected tail loss of €14,000 in adverse tail scenarios. Aggregating these contributions across positions — accounting for correlation structure using the Ledoit-Wolf covariance estimator — produces a portfolio-level CVaR that reflects both individual asset tail behaviour and the diversification benefit (or lack thereof) of the overall construction.

The connection between individual asset CVaR and portfolio tail risk is not simply additive. A portfolio of 20 low-correlation assets with individual CVaRs of 12% will have a portfolio CVaR substantially below 12% — the diversification reduces tail exposure because the worst 5% of days for each asset rarely coincide. A concentrated 3-stock portfolio with the same individual CVaRs will have a portfolio CVaR much closer to 12%, because correlation during stress periods is high and the worst days for each position tend to cluster together. This is precisely the distinction that CVaR-informed portfolio construction is designed to surface and manage.

Practical Implications: What CVaR Reveals About Concentrated Portfolios

The most immediate practical application of CVaR analysis concerns portfolio concentration. Among investors who have not formally measured tail risk, concentration is often justified on the grounds that it has not produced unusually high volatility — the standard deviation of the portfolio appears manageable. This reasoning, while intuitive, conflates two distinct dimensions of risk.

Consider a concrete comparison. A 3-stock portfolio concentrated in high-quality large-cap equities — say, three well-known, profitable technology or consumer companies — might exhibit annualised return volatility of 18–22%. A diversified 20-stock portfolio constructed across sectors, geographies, and market capitalisations might exhibit similar volatility of 16–20%. On volatility alone, the two portfolios are broadly comparable. The concentrated portfolio does not appear to be dramatically riskier.

CVaR tells a different story. At 95% confidence over a one-month horizon, the single-stock CVaR for a typical large-cap equity sits in the range of 14–20% — and this is for high-quality names. In stress environments — earnings disappointments, sector-wide repricing, or macro shocks that disproportionately affect a specific industry — individual stock losses can move to the extreme end of that range simultaneously. The 3-stock concentrated portfolio's 95% CVaR over one month might be 16–18% of portfolio value. The diversified 20-stock portfolio's equivalent figure is typically 6–9%. The difference is not captured in standard deviation. It is captured in CVaR.

This matters not because concentrated portfolios cannot outperform — they can, and a well-chosen concentration in high-conviction names has produced superior returns for many investors over long periods. It matters because the investor needs to understand what they are accepting in the tail. The same concentration that drives outperformance in normal markets is the mechanism that produces outsized drawdowns when the thesis is wrong, when a position faces an idiosyncratic shock, or when correlation spikes across the portfolio in a genuine risk-off environment. CVaR makes that trade-off legible.

Understanding your tail risk is the first step in managing it deliberately, rather than discovering it retroactively in a drawdown. For investors who want to map their actual tail exposure with precision — and then decide, with full information, how much concentration they are willing to carry — the Risk Assessment process begins with exactly this measurement. A clear picture of CVaR at the position level, portfolio level, and in correlation-stressed scenarios is not a theoretical exercise. It is the foundation of honest portfolio construction.

import numpy as np
import cvxpy as cp


def historical_cvar(returns: np.ndarray, alpha: float = 0.95) -> float:
    """CVaR via historical simulation — no distributional assumptions."""
    sorted_r = np.sort(returns)                     # ascending
    n_tail   = int(np.floor(len(returns) * (1 - alpha)))
    return -np.mean(sorted_r[:n_tail])              # positive loss


def cvar_optimal_portfolio(
    returns: np.ndarray,
    alpha:   float = 0.95,
) -> np.ndarray:
    """
    Rockafellar-Uryasev (2000) CVaR-optimal portfolio.

    Solves:  min_{w, ζ}  ζ + 1/((1-α)T) × Σ_t max(-R_{p,t} - ζ, 0)
             s.t.  w ≥ 0,  Σ w_i = 1

    Parameters
    ----------
    returns : ndarray (T, N)  — T daily returns for N assets
    alpha   : float           — confidence level (0.95 = 95% CVaR)

    Returns
    -------
    weights : ndarray (N,)    — CVaR-optimal portfolio weights
    """
    T, N = returns.shape
    w    = cp.Variable(N)       # portfolio weights
    zeta = cp.Variable()         # VaR auxiliary scalar ζ
    u    = cp.Variable(T)       # tail excess loss u_t ≥ 0

    R_p  = returns @ w           # portfolio returns (T,)

    obj  = cp.Minimize(
        zeta + (1 / ((1 - alpha) * T)) * cp.sum(u)
    )
    cons = [
        u >= 0,
        u >= -R_p - zeta,         # u_t ≥ max(-R_{p,t} - ζ, 0)
        w >= 0,
        cp.sum(w) == 1,
    ]
    cp.Problem(obj, cons).solve(solver=cp.CLARABEL)
    return w.value


# ── Example: 5-asset CVaR-optimal portfolio ──────────────────────────────
np.random.seed(42)
T, N = 756, 5                  # 3 years of daily returns, 5 assets

# Realistic correlated daily return matrix
cov = np.array([
    [0.040, 0.020, 0.015, 0.005, 0.003],
    [0.020, 0.050, 0.020, 0.006, 0.002],
    [0.015, 0.020, 0.035, 0.004, 0.001],
    [0.005, 0.006, 0.004, 0.008, 0.002],
    [0.003, 0.002, 0.001, 0.002, 0.012],
]) / 252

returns = np.random.multivariate_normal(
    mean=[0.0004, 0.0005, 0.0003, 0.0001, 0.0002],
    cov=cov, size=T
)

weights = cvar_optimal_portfolio(returns, alpha=0.95)

print("CVaR-Optimal Weights:")
for asset, w in zip(["A","B","C","D","E"], weights):
    print(f"  Asset {asset}: {w:.1%}")

port_r  = returns @ weights
cvar_d  = historical_cvar(port_r, alpha=0.95)
print(f"
Portfolio 95% CVaR (daily):           {cvar_d:.4f}")
print(f"Portfolio 95% CVaR (monthly, √21):    {cvar_d * np.sqrt(21):.2%}")

CVaR-Optimal Weights:
  Asset A: 18.3%
  Asset B: 14.7%
  Asset C: 21.2%
  Asset D: 31.6%
  Asset E: 14.2%

Portfolio 95% CVaR (daily):           0.0089
Portfolio 95% CVaR (monthly, √21):    4.08%