The worst day isn't in the average: VaR and CVaR explained
4 min read by Opthest
On July 13, 2026 the VIX jumped to 17.40, a move of over 13% in a single session, as tensions in the Strait of Hormuz pushed oil higher and the Nasdaq and S&P 500 fell 1.5% and 0.8% respectively (source: eciks.org, July 2026). The next day oil climbed further after a new 20% “protection tax” was announced on tanker traffic through the Strait. But the interesting number isn’t the VIX level itself — it’s that the ratio between Nasdaq implied volatility (VXN, 27.34) and S&P 500 implied volatility (VIX, 16.73) hit 1.64: fear was concentrating in tech and chip names, not spreading evenly across the market. A standard deviation computed on the whole portfolio would never have told you that story.
What doesn’t standard deviation tell you?
It doesn’t tell you where the risk is hiding, only how much on average the portfolio swings. Standard deviation treats a 5% gain and a 5% loss as the exact same amount of “risk” — but no investor experiences them the same way. Worse, it weighs every trading day equally, calm days included, which dilutes the signal from the few days that actually matter — the rare, violent tail days. If your portfolio’s risk is concentrated in one corner (a sector, a shared factor, a single geopolitical trigger), an average taken across the whole distribution can stay low and reassuring right up until the day it isn’t.
What is VaR, and why isn’t it enough on its own?
Value at Risk (VaR) tries to answer a sharper question: “with 95% (or 99%) confidence, what’s the most I could lose over a given period?” If a portfolio’s 95% VaR is -3%, it means that historically, in 19 days out of 20, the loss stayed within that bound. That’s already an improvement on standard deviation, because it looks directly at the tail of the distribution rather than its average width. But it has a built-in blind spot: it says nothing about how bad the remaining 5% (or 1%) of cases actually get. A 95% VaR of -3% is equally compatible with a worst day of -3.5% or a worst day of -25%. The number is identical either way — the real risk is not.
What is CVaR, and why is it the tougher measure?
Conditional Value at Risk (CVaR), also called Expected Shortfall — formalized by Rockafellar and Uryasev in 2000 — answers exactly the question VaR leaves open: on average, how much do you lose when you actually breach the VaR threshold? It isn’t the threshold itself, but the average of the losses beyond it — the real weight of that worst 5% of scenarios, not just the point where they begin. That’s why CVaR is always equal to or worse than VaR at the same confidence level: it isn’t satisfied knowing where the cliff starts, it measures how far down you go on average once you’re over the edge. It’s the difference between “the water crosses the levee” and “how much water actually floods the house when it does.”
Are optimizing for variance and optimizing for CVaR the same thing?
No, and the difference isn’t just terminology. Mean-variance optimization (the efficient frontier) minimizes overall portfolio wobble, treating an upside swing and a downside swing identically — it’s blind to the shape of the distribution’s left tail. CVaR-based optimization instead directly minimizes the expected loss in the worst-case scenarios, focusing on exactly the part of the distribution an investor cares about most: the one where things actually go wrong. Two portfolios with the same variance can have very different tails — one with rare but enormous extreme losses, the other with more frequent but contained ones — and only a CVaR-oriented optimization tells them apart correctly.
What does Opthest actually compute?
Every tracked portfolio’s risk panel shows 95% VaR and CVaR alongside the other metrics (Sharpe, Sortino, Calmar), all computed on the same historical return series — never mixing windows, for the same reason a backtest has to be read together with its window. And among the optimization methods available — alongside Markowitz, Black-Litterman, risk parity and the evolutionary solver — there’s a method that directly minimizes portfolio CVaR (a linear program, the Rockafellar-Uryasev formulation) instead of variance alone — useful when the goal isn’t “swing as little as possible on average” but “cap how much you lose in the worst case.” As always, these remain algorithmic simulations on historical data: educational/informational content, not personalized financial advice, with no guarantee about future results.
Standard deviation measures how much the floor shakes. VaR tells you where the hole starts. CVaR is the only one of the three that tells you how deep it is — which is exactly what matters on the day you fall in.