Confidence Intervals: 95% vs. 99% VaR Explained

The confidence level in a VaR calculation is the probability that actual losses won't exceed the VaR estimate. A 95% confidence VaR says: there's a 95% chance portfolio losses stay within the threshold. Conversely, a 5% chance they exceed it. A 99% VaR is more conservative: a 99% chance of staying within threshold, a 1% chance of exceeding it.

The choice between 95%, 99%, and other confidence levels is not academic—it drives portfolio position sizing, margin requirements, and risk governance. A retail trader might use 95% for routine monitoring but 99% for stress scenarios. A bank might use 99% for internal risk monitoring but report 99.9% to regulators. This article explains how confidence levels work, how to interpret them, and which to use for different purposes.

Quick definition: The confidence level is the probability that actual portfolio losses won't exceed the VaR estimate, expressed as a percentage from 0 to 100.

Key takeaways

• Higher confidence levels (99%) mean larger expected losses compared to lower levels (95%)
• 95% confidence is appropriate for routine risk monitoring and position sizing
• 99% confidence is appropriate for stress scenarios and regulatory reporting
• Confidence levels are not about portfolio probability—they're about the tail of the loss distribution
• Exceeding a VaR threshold should happen regularly, not be treated as a crisis

The Core Relationship: Confidence and VaR Magnitude

For the same portfolio and time horizon, higher confidence always produces larger VaR numbers. This relationship is not because the portfolio is riskier—it's because you're asking a more extreme question.

Example: A $100,000 portfolio with 1.5% daily volatility and 0.5% expected daily return.

Using parametric VaR:

95% confidence: $100,000 × 1.645 × 0.015 = $2,468
99% confidence: $100,000 × 2.326 × 0.015 = $3,489
99.9% confidence: $100,000 × 3.090 × 0.015 = $4,635

The same portfolio, measured at different confidence levels, shows dramatically different loss estimates. The 99% figure is 41% larger than the 95% figure. This is correct. At 95% confidence, you're comfortable with a 5% chance of exceeding the loss. At 99% confidence, you're only comfortable with a 1% chance. That additional conservatism costs you position size.

The Percentile Interpretation

Confidence levels map directly to percentiles in the loss distribution. A 95% confidence VaR is the 5th percentile of returns (or the 95th percentile if you rank from worst to best). A 99% confidence VaR is the 1st percentile.

In historical VaR terms: If you have 252 trading days of historical returns:

• 95% confidence: 252 × 0.05 = 12.6 ≈ 13th worst day
• 99% confidence: 252 × 0.01 = 2.52 ≈ 3rd worst day
• 99.9% confidence: 252 × 0.001 = 0.252 ≈ not even once in 252 days

This illustrates a critical limitation: calculating 99.9% confidence VaR from 252 days of history is unreliable because your lookback period hasn't captured that tail. You need 1,000+ trading days (four years) to reliably estimate 99.9% tail outcomes. This is why banks requiring 99.9% VaR use much longer lookback periods or simulation methods.

The Exceedance Frequency: When VaR Gets Broken

By definition, the actual loss should exceed your VaR threshold with a frequency matching the complement of your confidence level:

• 95% VaR: Exceeded about 5% of the time (roughly 12 trading days per year)
• 99% VaR: Exceeded about 1% of the time (roughly 2.5 trading days per year)
• 99.9% VaR: Exceeded about 0.1% of the time (roughly once per four years)

This is not a failure of the VaR model. This is the model working as intended. If you track your 95% VaR every trading day for a year and see that actual losses exceeded it only 3 times out of 252 days, your model is overly conservative—you're not using enough leverage for your risk tolerance. If you see 20 exceedances, your model is too optimistic—you need to reduce leverage or recalibrate.

A professional risk team backtests daily: they plot actual returns against the VaR estimates from the previous day and count exceedances. Too many signals a model problem (stale volatility, wrong distribution assumption, parameter error). Too few signals the risk limit is unnecessarily tight.

95% Confidence: The Day-to-Day Standard

95% confidence VaR is the most common level for routine portfolio monitoring, position sizing, and intraday risk management. Why 95%?

Practical frequency: At 95% confidence, you exceed VaR about once a month. This is frequent enough that it doesn't feel like a rare catastrophe, yet rare enough that it commands attention. Traders and portfolio managers expect to occasionally breach their 95% VaR and view it as a signal to investigate, not panic.

Statistical stability: 95% confidence falls within the bulk of the distribution. You need fewer observations to estimate it reliably. A 252-day lookback provides about 13 samples of the 95th percentile, which is adequate.

Regulatory acceptance: Most securities regulators and brokers use 95% confidence for internal position limits and margin requirements. It's the market standard.

Professional norm: Trading desks report 95% VaR to senior management as their primary risk metric. Compliance teams review it daily.

Example scenario: Your $250,000 portfolio's 95%, 1-day VaR is $5,200. Today you lose $5,800—exceeding VaR by $600. Should you panic? No. Your model says this happens about once per month. The excess is small, suggesting no fundamental model error. You investigate whether volatility increased, but you don't overhaul your risk framework.

99% Confidence: The Stress Threshold

99% confidence VaR is appropriate for stress scenarios, larger portfolio reviews, and regulatory reporting. Banks report 99% VaR to regulators because it's more conservative; it ensures banks hold more capital.

99% confidence has a very different psychological meaning. Exceeding a 99% VaR is much rarer—roughly once per year—so it feels more like a true "crisis" event. This is useful for contingency planning: if your 99% VaR is exceeded, you ask: "Why? Is this the 1-in-100 normal event, or did something structural break?"

Example scenario: Your $250,000 portfolio's 99%, 1-day VaR is $8,100. One day you lose $7,500, staying within VaR. The next day you lose $9,200, exceeding VaR by $1,100. You now have two consecutive large losses—roughly a once-per-year event. You immediately review correlation changes, check if volatility has spiked, and consider whether a market regime has shifted. This two-day pattern suggests the market environment may have changed, warranting a deeper analysis.

Regulatory Requirements

Different regulatory contexts require different VaR confidence levels:

Securities and Exchange Commission (SEC): Large hedge funds must report Value-at-Risk using a 95% confidence level for public disclosures.

Basel III Banking Regulations: Banks must hold capital based on a 99% confidence level, 10-day VaR, plus an additional "stress test" multiplier (typically 1.4-3.0x). This ensures banks hold more capital than the 99% VaR alone would suggest, protecting against tail risks and regime changes.

Dodd-Frank Act: Derivatives dealers report Value-at-Risk using standardized confidence levels, typically 99% for regulatory purposes.

Financial Industry Regulatory Authority (FINRA): Brokers use VaR internally for margin requirements, typically 95% confidence for daily monitoring.

These regulatory frameworks assume that 99% confidence provides sufficient buffers. In the 2008 crisis, many institutions found even 99% VaR was insufficient because market behavior broke the models' assumptions. This led to additional stress-testing requirements and the introduction of Expected Shortfall (CVaR) as a supplement.

99.9% Confidence: The Regulatory Reserve

99.9% confidence VaR is very rare in practical use because it requires an enormous amount of historical data or simulation. If you have 252 trading days, the 99.9th percentile is outside your data. You'd need 1,000 days or more.

Central banks and prudential regulators sometimes use 99.9% confidence internally to size capital reserves for systemic risk. The Federal Reserve, Bank for International Settlements, and European Banking Authority all reference 99.9% confidence in their guidance, but they typically use simulation or complex statistical models rather than direct empirical percentiles.

For a retail trader or fund manager, 99.9% VaR is overkill. Stick to 95% and 99%.

Comparing Methods Across Confidence Levels

Different calculation methods produce different relationships between confidence levels.

Parametric VaR shows a smooth, proportional relationship. Double the Z-score, and VaR doubles. A 99% parametric VaR is always 1.41x the 95% VaR (the ratio of Z-scores: 2.326 / 1.645 ≈ 1.41).

Historical VaR shows a jagged relationship because you're reading from discrete historical outcomes. If your 13th worst day was -2.0% and your 3rd worst day was -4.5%, your 99% VaR is 2.25x your 95% VaR. If the 3rd worst day had been -3.9% instead, the ratio would be 1.95x. The proportionality is not fixed.

Monte Carlo VaR resembles parametric (smooth relationships) because the simulation generates a large sample that converges to a smooth distribution. With 10,000 simulations, the 99% and 95% relationship will be approximately 1.4-1.5x in ratio.

The divergences between methods are useful. If parametric and historical VaR give very different confidence-level relationships, that signals something non-normal in the distribution (fat tails, skewness, regime breaks).

Real Example: A Personal Trading Account

You manage a $150,000 account trading stocks and ETFs. Your risk policy specifies:

• Position limit: No single position exceeds 5% of portfolio value = $7,500
• Daily stop-loss: Any position with greater than 8% loss is exited
• Portfolio limit: 95% confidence daily VaR < $4,500 (3% of portfolio)
• Stress limit: 99% confidence daily VaR < $7,500 (5% of portfolio)

You calculate VaR at end of day:

• 95% VaR: $4,200 (within limit)
• 99% VaR: $6,800 (within limit)

You have adequate room for tomorrow's trading. Next day, markets open sharply down. By midday:

• Intraday 95% VaR: $4,800 (exceeding limit)

You reduce position size by 15% to bring VaR back to $4,100, respecting your 95% limit. You keep the 99% limit at $6,600 (also respecting the 5% stress level).

A week later, a corporate earnings announcement creates a market regime shift. Volatility spikes:

• 95% VaR: $5,900 (exceeding limit)
• 99% VaR: $8,200 (exceeding limit)

Both limits are violated. You immediately reduce portfolio size by 25%. Your risk limits exist to force this kind of discipline: when volatility spikes, you shrink exposure proportionally.

Choosing Between 95% and 99% for Different Purposes

Daily position sizing: Use 95%. Most trading firms set intraday position limits based on 95% VaR. This allows sufficient leverage for profitability while triggering position reductions when volatility rises.

Overnight limits: Use 95%. Your broker sets overnight position limits using 95% VaR, assuming you can't rebalance until market open.

Monthly risk review: Use both. Compare 95% to understand routine volatility. Check 99% to stress-test against bad days.

Regulatory reporting: Check your regulator's requirement. SEC requires 95%; Basel III requires 99%.

Stress scenarios: Use 99% or higher. When you ask "how much could I lose in a crisis day?" you want a more conservative number.

Portfolio insurance: Use 99%. When deciding how much capital to reserve for insurance or hedging, use the higher confidence level.

Common Misunderstandings

Confusing confidence level with probability of ruin: A 95% VaR is not the probability the portfolio will be negative. It's the probability that losses won't exceed a specific amount. A portfolio with 95% confidence VaR of $10,000 might still have a 0.01% probability of losing $50,000 on a very bad day.

Treating VaR exceedance as a model failure: If your 95% VaR is exceeded three times in a month, your model isn't broken. You expect roughly two exceedances per month (5% of 252 days / 12 months ≈ 1 exceedance per month, roughly). Three is within noise.

Using the wrong time horizon for confidence level: 95% and 99% are confidence levels, not time horizons. A 95%, 1-day VaR is different from a 95%, 10-day VaR. Always specify both.

Assuming higher confidence is always better: A 99% VaR is more conservative but forces you to size positions smaller. The right confidence level depends on your risk tolerance and liquidity. A high-frequency trader uses 95%; a pension fund uses 99%.

FAQ

What's the difference between one-sided and two-sided confidence intervals?

VaR is always one-sided (downside losses). It doesn't measure upside. A two-sided 95% confidence interval would split the risk: 2.5% loss and 2.5% gain. VaR focuses only on losses. If you want upside/downside symmetry, look at "Value-at-Return" or expected return ranges, not VaR.

Can I interpolate between 95% and 99% confidence?

Yes, in historical VaR, you can calculate intermediate percentiles. For example, 97% confidence would be the 3rd percentile. In parametric VaR, look up the Z-score for 97% (about 1.88) and use it. In Monte Carlo, just read the appropriate percentile from your ranked simulations. However, most practitioners stick to 95% and 99% for consistency.

What confidence level should I use for options portfolios?

Start with 99% for options. Options have nonlinear payoffs and fatter tails than stocks. A 95% confidence might be optimistic for a portfolio with significant option exposure. Use 99% as the minimum or pair 95% with detailed scenario analysis.

How does confidence level affect time-scaling?

The square-root-of-time rule (used in parametric VaR) applies independently of confidence level. If your 95%, 1-day VaR is $2,000, your 95%, 10-day VaR is approximately $2,000 × sqrt(10) ≈ $6,325. The same scaling applies at 99% confidence. The confidence level doesn't affect how volatility grows with time.

Should I use a mix of 95% and 99% VaR?

Yes, professional risk teams do. They report 95% as the primary metric for daily monitoring and position sizing, then separately track 99% to flag stress conditions. Some firms set 95% as a soft limit (warning) and 99% as a hard limit (action required).

If I exceed my VaR, am I necessarily taking too much risk?

Not necessarily. Exceeding your VaR at the right statistical frequency is normal. But if you exceed it far more frequently than expected, your model is wrong. If you exceed it less frequently, you're being too conservative. The statistical frequency of exceedances is how you calibrate whether your confidence level is appropriate.

What if I want to be 99.5% confident?

In parametric VaR, the Z-score for 99.5% confidence is about 2.576. Use it just like 2.326 (for 99%). In historical VaR, with 252 days, the 99.5th percentile is approximately the 1.26th worst day—you're outside your sample, so reliability is low. With 500 days, it's about the 2.5th worst day. In Monte Carlo, generate 1,000+ simulations and read the appropriate percentile.

How do I explain confidence levels to non-technical stakeholders?

"95% confidence VaR of $5,000 means: 95 out of 100 trading days, we don't lose more than $5,000. Five days per year, we might lose more. 99% confidence VaR of $8,000 means: we expect to lose more than that about once per year. The higher number accounts for bad-day scenarios."

Related concepts

• What Is Value-at-Risk?
• Parametric VaR: The Variance-Covariance Method
• Historical VaR: Using Past Returns
• What Is a Black Swan?

Summary

The confidence level is the probability that actual losses won't exceed your VaR estimate. Higher confidence levels produce larger VaR numbers because they protect against worse outcomes. 95% confidence is the standard for routine risk monitoring and position sizing; it expects exceedances roughly once per month. 99% confidence is the standard for stress scenarios and regulatory reporting; it expects exceedances roughly once per year. The choice between them depends on your context: active traders use 95%, large institutions and regulators use 99%. The key insight is that exceeding VaR is not a failure—it happens regularly, by design. The real test is whether the frequency of exceedances matches the statistical expectation of your chosen confidence level.

Next

→ VaR vs. CVaR (Expected Shortfall)