What Value at Risk Is

A trading desk might hold ten thousand positions — stocks, bonds, swaps, options, a few things nobody fully remembers buying. Each has its own volatility, its own correlations, its own ways of going wrong. The board does not want ten thousand numbers. The regulator does not want ten thousand numbers. They want one: how much could we lose? Value at Risk is the industry’s answer to that demand — a single dollar figure that claims to summarise the danger of an entire book. It is imperfect, occasionally dangerous, and utterly everywhere. So let’s learn exactly what it does say, and — just as important — what it quietly refuses to.

A quick gut-check before we define anything.

From a thousand risks to one number

Analogy. A weather service could hand you barometric pressure, humidity, wind shear, and dew point for every square kilometre. Instead it says: “70% chance of rain.” One number, lossy on purpose, actionable. VaR is that “chance of rain” for a portfolio’s losses.

Why does finance crave a single figure? Because risk has to be compared and aggregated. A desk head wants to know whether the equities book is riskier than the rates book; a CRO wants to add up risk across desks; a regulator wants every bank measured on the same yardstick. Volatility alone won’t do it — a 2% daily vol on equities and a 0.3% daily vol on Treasuries aren’t directly comparable, and you can’t sensibly “add up” ten thousand standard deviations. VaR collapses all of it — every position, every correlation, every fat or thin tail — into one dollar amount on one scale, so it can be compared, summed, and limited.

That universality is exactly why it became the lingua franca of risk. Under the Basel market-risk framework, banks have for decades sized regulatory capital off VaR-style measures: the more your VaR, the more capital you must hold against trading losses. When a number determines how much capital a global bank parks on the sidelines, everyone learns to speak it. (Basel has since pushed the regulatory measure toward Expected Shortfall — the very thing VaR omits, and the subject of a later lesson — but VaR remains the everyday vocabulary on the desk.)

The VaR sentence: three ingredients

A VaR figure is meaningless until you attach three things to it. Quote a VaR without all three and you’ve said nothing.

1. Horizon — over what period? One day for a liquid trading book you could exit fast; ten days for the Basel regulatory measure; sometimes a month for slower portfolios. Risk grows with time, so the horizon is half the number.
2. Confidence level $c$ — how sure? 95% and 99% are the classics. It sets how deep into the bad tail you’re looking. Higher confidence → a rarer, larger loss.
3. The loss amount — how much? The actual output, quoted in currency ($4.2M) or as a percent of the portfolio (3.1%).

Stitch them together and you can read a VaR aloud. Take “1-day 99% VaR = $4.2M”:

“Over 1 day, we are 99% confident the loss will not exceed $4.2M. On the worst ~1 day in 100, we expect to lose more than that.”

The confidence level has a blunt operational meaning: it’s how often you breach the line. The complement $1 - c$ is the expected fraction of periods the loss is worse than VaR.

Worked example. Your book runs a 1-day 95% VaR of $1M. Over a 250-trading-day year, you expect to breach $1M on about $0.05 \times 250 = 12.5$ days — roughly one a month. If you breached it 40 times last year, your model is badly underestimating risk; if you breached it twice, it’s far too conservative (or you got lucky). That counting exercise — comparing realised breaches to the expected $1-c$ rate — is the seed of backtesting, which we’ll meet later.

VaR is a quantile of the loss distribution

Here’s the picture that makes everything click. Imagine the distribution of your portfolio’s profit-and-loss over the horizon — a bell-ish curve centred near zero (or a hair positive), gains to the right, losses to the left. Most days you land near the middle. The scary days live in the left tail.

VaR is simply a cutoff in that left tail. Slide a vertical line leftward until exactly $(1-c)$ of the probability mass sits beyond it. The loss at that line is your VaR.

Precise definition. The 1-day $c$-confidence VaR is the loss $L$ such that $\Pr(\text{loss} > L) = 1 - c$ — equivalently, VaR is the $(1-c)$ quantile of the loss distribution (the $c$ quantile of losses, the $(1-c)$ percentile of returns). It’s a percentile, nothing more exotic.

Drag the slider below: as you raise the confidence, the line marches deeper into the loss tail and the reported VaR grows.

Worked example (the parametric shortcut). Suppose daily P&L is roughly normal with mean ≈ 0 and a daily volatility $\sigma$ of $1M. The 99% quantile of a normal sits about 2.33 standard deviations below the mean. So:

$\text{VaR}_{99\%} \approx 2.33 \times \sigma,$

which with $\sigma$ of $1M gives $2.33 \times 1 = 2.33$, i.e. $2.33M. At 95% the multiplier is about 1.65, giving $1.65 \times 1 = 1.65$, i.e. $1.65M. Same portfolio, two confidence levels, two numbers — both just quantiles of one curve. (This normal shortcut is the parametric method; it’s clean but assumes the curve really is normal, which the tails love to violate.)

What VaR deliberately does NOT tell you

VaR is honest about exactly one thing — where the line is — and silent about almost everything else. The classic blow-ups come from forgetting these gaps.

• It says nothing about how bad the breach is. VaR tells you the loss is “more than $4.2M” on a bad day — $4.3M? $40M? $400M? VaR shrugs. It marks the door to the tail, not the depth of it. The number that does measure average severity beyond the line is Expected Shortfall (CVaR) — the entire next topic, and the shaded region in the chart above.
• It is not the worst case / maximum loss. A 99% VaR is breached ~1 day in 100 by design. Treating VaR as “the most we can lose” is the single most expensive misreading in risk management. Your true maximum loss can dwarf VaR.
• Garbage assumptions → garbage number. VaR inherits every flaw of the distribution you feed it: assume normal returns and you’ll systematically under-state tail risk; estimate volatility from a calm period and your VaR will be serene right up until the storm. The number is only as good as its inputs.
• One calm figure can hide fat tails. Two portfolios can post the same VaR while one has a thin, well-behaved tail and the other a monster lurking just past the cutoff. VaR can’t see the difference — it only reads the position of the line, not the mass piled up behind it.

Dollar vs percent, absolute vs relative

Two quick framing choices change how a VaR number reads.

• Currency vs percent. VaR can be quoted as dollars ($4.2M) or as a percent of portfolio value (3.1% of a $135M book). Same risk, different unit — percent travels better across portfolios of different sizes; dollars land harder in a board meeting.
• Absolute vs relative. Absolute VaR measures the loss from zero (today’s value). Relative VaR measures the loss relative to the expected outcome — the mean. Over a single day the mean drift is tiny, so the two nearly coincide; over longer horizons, where expected return matters, relative VaR (loss below the mean) can sit meaningfully below absolute VaR. When someone quotes a number, it’s worth knowing which baseline they measured from.

Putting it together

Value at Risk takes a portfolio of countless moving parts and reports one number on one scale — the loss you won’t exceed over a given horizon with a given confidence, quoted in dollars or percent. It’s a quantile of the P&L distribution: slide a line into the left tail until $(1-c)$ of the mass is beyond it. That single comparable figure is why it became the language of bank risk and Basel capital. But it marks the door to the tail, never the depth: it’s silent on how bad a breach gets, it is not a maximum loss, and it’s only as trustworthy as the distribution behind it.

We’ve defined the line. Now the obvious question: where do we actually get the loss distribution to draw it on? The next lesson starts with the most assumption-light answer — historical simulation — letting the past few hundred days be the distribution, no bell curve required.