Measuring and Reporting Value at Risk

Value at Risk sounds precise, which is why people misuse it. VaR is not your maximum possible loss. It is not a promise that losses will stay inside a box. It is a loss threshold at a chosen probability level over a chosen horizon. If your 1-day 99% VaR is $5 million, that means 1 day out of 100 you expect to lose more than $5 million. The whole job of risk reporting is to make that sentence impossible to misunderstand.

What VaR Actually Measures

A correct VaR statement looks like this:

At the 99% confidence level, the portfolio is expected to lose no more than $X on 99 out of 100 trading days, assuming the model assumptions hold.

That sounds subtle, but the distinction matters.

VaR answers:

• What loss threshold corresponds to a chosen percentile?
• How much downside are we carrying under the model's assumptions?
• How does today's risk compare with yesterday's or the desk limit?

VaR does not answer:

• What is the worst loss we can suffer?
• What happens beyond the cutoff?
• What happens if liquidity disappears or correlations break?

The point is: VaR is a useful dashboard number. It is a bad standalone risk philosophy.

The Inputs That Drive the Number

Every VaR report embeds a set of choices:

If you do not disclose these choices, the number is not comparable across desks, dates, or firms.

The Three Main Calculation Approaches

Parametric VaR

This is the fast, closed-form approach. It typically assumes returns are approximately normal and risk can be summarized by volatility and correlation.

For a simple linear portfolio:

VaR = Portfolio Value x Volatility x Z-score x Square Root of Time

Advantages:

• Fast
• Easy to update daily
• Good for broad directional books

Weaknesses:

• Thin-tail assumptions can understate real stress
• Poor fit for options or path-dependent books
• Correlation behavior can look stable right until it is not

Historical simulation VaR

This approach replays actual historical return moves against today's portfolio.

Advantages:

• No normal-distribution assumption
• Captures real historical joint moves
• Easier to explain to committees

Weaknesses:

• The future may not look like the sample
• Window selection matters a lot
• May miss shocks that have not occurred in the chosen history

Monte Carlo VaR

This uses simulated scenarios based on chosen distributions and dependencies.

Advantages:

• Flexible
• Works better for complex derivatives
• Can incorporate richer factor dynamics

Weaknesses:

• Model risk increases
• Computationally heavier
• Easier to build something that looks sophisticated but is poorly specified

Worked Example: A Simple 1-Day Parametric VaR

Assume:

• Portfolio value: $100 million
• Daily volatility: 1.2%
• Confidence level: 99%
• Z-score for 99%: about 2.33

Then:

VaR = $100,000,000 x 1.2% x 2.33 = about $2.80 million

That means the model expects:

• On roughly 99 out of 100 trading days, losses should be less than about $2.80 million
• On roughly 1 out of 100 trading days, losses may exceed that amount

That last sentence is the part people forget.

Where Expected Shortfall Fits

Expected Shortfall (ES), also called Conditional VaR, asks the better tail question:

If we are already in the worst 1% of outcomes, what is the average loss there?

If your 99% VaR is $2.80 million, the 99% ES might be $4.10 million. VaR tells you where the cliff starts. ES tells you how deep the ravine may be once you fall off it.

This matters operationally and regulatorily:

• Many investment firms still report VaR internally because it is familiar and limit-friendly
• Under the modern Basel market-risk framework, the internal-models capital framework shifted from VaR to Expected Shortfall under stress

Why this matters: a risk function that still treats VaR as the only serious number is behind the framework.

VaR for Option Books: What Goes Wrong Fast

Options make bad VaR setups obvious.

A delta-only view can miss:

• Gamma acceleration
• Vega shocks
• Skew moves
• Correlation breaks across underlyings

Suppose you run a short-index-options book that looks calm most days. A linearized VaR can look benign right up until a gap move, volatility spike, and liquidity deterioration happen together. That is why serious derivatives reporting pairs VaR with:

• Stress tests
• Greek limits
• Gap-risk scenarios
• Liquidity overlays

Backtesting: The Minimum Credibility Test

If you publish VaR, you must compare predicted losses with realized outcomes.

This is the basic backtesting logic:

• If you run a 99% 1-day VaR
• Over 250 trading days
• You would expect about 2 to 3 exceedances

An exceedance means actual loss was worse than VaR predicted.

Example exceedance table

Eight exceedances is a problem. It tells you one or more of the following:

• Volatility estimate was too low
• Correlations were too stable
• Non-linearity was undercaptured
• The book changed faster than the model
• The market regime shifted

The durable lesson: a VaR model is not "validated" because it exists. It earns trust only by surviving comparison with reality.

Reporting VaR So People Do Not Misread It

A usable VaR report should include more than one number.

Good daily report structure

Then add plain-English commentary:

• What changed?
• Which desk or factor drove it?
• Was the move from positions, markets, or model inputs?
• Did any limit, watch level, or escalation trigger fire?

That commentary is not optional. It is the difference between risk reporting and number-dumping.

The Most Common VaR Mistakes

Calling VaR a "maximum expected loss"

This is the classic error. It overstates precision and hides tail risk.

Scaling blindly with square-root-of-time

That can be a useful approximation in stable conditions. It is not a law of nature.

Treating correlations as fixed

Correlations are lowest when you least need the comfort and highest when you most need the hedge.

Ignoring liquidity

A mark-to-model VaR can look tight while the real exit price in stress is much worse.

Reporting one confidence level only

Showing 95% VaR, 99% VaR, and Expected Shortfall gives a much better picture of tail shape.

Using VaR without scenario analysis

VaR is backward-looking or model-looking. Scenarios let you ask, "What if the next stress is unlike the calibration window?"

A Practical Risk Stack

For most derivatives books, a solid stack looks like this:

1. VaR for day-to-day comparability and limit management
2. Expected Shortfall for tail severity
3. Stress testing for regime breaks
4. Sensitivity limits for position-level control
5. Backtesting for model accountability

If one of those layers is missing, the stack gets fragile.

Checklist Before You Distribute a VaR Report

• State confidence level and time horizon explicitly
• Disclose methodology: parametric, historical, or Monte Carlo
• Show major assumptions, including lookback window
• Pair VaR with Expected Shortfall or stress metrics
• Highlight any exceedances and explain them
• Add commentary on the main position and factor drivers
• Avoid language that implies VaR is a hard loss cap

The bottom line: VaR is still useful, but only if you describe it honestly. Use it as a threshold statistic, not as a substitute for judgment, tail analysis, or stress testing.