CVaR vs Value-at-Risk: Measuring the Risk That Actually Hurts

VaR tells you where the losses begin; CVaR tells you how bad they get once they do. Why that gap matters, and why CVaR makes a better training objective.

Ask a risk model one question — "how much could I lose?" — and the answer you get depends entirely on how you asked. Value-at-Risk (VaR) and Conditional Value-at-Risk (CVaR) sound like variations on a theme. They are not. One tells you where the bad days begin; the other tells you how bad they actually get. In the tail, that difference is the whole game.

What VaR actually measures

VaR at the 95% level answers: "On 95% of days, losses won't exceed X." It marks a threshold — the boundary of the worst 5% of outcomes. It is intuitive, it is everywhere in regulation, and it has one fatal blind spot: it says nothing about what happens past the threshold.

Two portfolios can share an identical 95% VaR while one occasionally loses twice as much as the other on its worst day. VaR cannot tell them apart. It reports where the cliff edge is, not how far the drop goes.

What CVaR adds

CVaR — also called Expected Shortfall — answers the harder question: "Given that we're in the worst 5% of days, how much do we lose on average?" It is the mean of the tail beyond VaR.

VaR points at where this curve crosses your threshold. CVaR integrates the area underneath it. Because it averages the entire tail, it sees the difference between a portfolio that bends in a crisis and one that snaps.

The property that matters: coherence

CVaR is a coherent risk measure — most importantly, it is sub-additive: the risk of a combined portfolio never exceeds the sum of its parts, so diversification is always rewarded. VaR violates this. You can combine two positions and have VaR report more risk than the two alone, which quietly punishes diversification and breaks optimization. That single mathematical defect is why practitioners increasingly treat VaR as a reporting convention and CVaR as the thing you actually optimize against.

Why CVaR belongs in the objective, not the report

The deepest use of CVaR isn't measurement — it's training signal. An agent rewarded purely on return learns to hide catastrophic tails behind smooth averages. Put CVaR inside the reward and the agent becomes explicitly afraid of its worst days, steering away from allocations that look great on average and ruin you in the tail. That is the principle behind our risk-aware portfolio work.

The takeaway

VaR tells you the bad days exist. CVaR tells you how much they cost — and, because it's coherent, it's the one you can safely optimize. If you only track one number for tail risk, track CVaR.