Historical VaR: Using Past Returns to Estimate Future Losses

Historical VaR is perhaps the most intuitive Value-at-Risk calculation method. Rather than assuming returns follow a bell curve or running thousands of simulations, it simply looks backward. It takes your actual historical returns over the past 250 or 500 trading days, ranks them from worst to best, and extracts a loss threshold directly from the data. If you want a 99% confidence VaR, you find the loss level that occurred only 1% of the time in your history. What happened before, the logic goes, might happen again.

This method requires no distributional assumptions. It doesn't assume normal returns, constant volatility, or stable correlations. It works with whatever shape the historical data actually has. For traders and portfolio managers who distrust parametric assumptions, historical VaR offers transparency: you can see the exact days that produced the loss threshold.

Quick definition: Historical VaR ranks actual past portfolio returns, finds the loss percentile matching your confidence level, and uses that as the forward-looking maximum loss estimate.

Key takeaways

• Historical VaR makes no assumptions about return distributions—it uses empirical data directly
• It ranks historical losses and finds the threshold that matches your confidence level
• Longer lookback periods (250-500 days) capture more market regimes but become stale
• Shorter lookback periods (63 days) respond to current volatility but miss rare events
• Historical VaR fails when unprecedented events occur or market regimes shift dramatically

The Method: Ranking Loss Days

The historical VaR calculation is straightforward in principle:

1. Gather historical daily returns for your portfolio over a lookback period (typically 252 trading days, one year).
2. Rank these returns from worst to best.
3. Identify the return at the percentile matching your confidence level.

Example: Suppose you have 252 daily returns for your portfolio. You want a 95% confidence VaR.

95% confidence means you're looking for the 5th percentile of the return distribution—the loss that was exceeded only 5% of the time. With 252 observations, the 5th percentile is approximately the 13th worst daily return (252 × 0.05 ≈ 12.6, round to 13).

If your 13th worst historical day produced a -2.1% return on a $100,000 portfolio, your 95% VaR is:

$100,000 × -2.1% = -$2,100

For a 99% confidence VaR (the 1st percentile), you'd look at roughly the 2.5th worst day (252 × 0.01 ≈ 2.5). If that day's return was -4.5%, your 99% VaR is:

$100,000 × -4.5% = -$4,500

No volatility estimate needed. No correlation matrix. No assumption of normality. You're using the actual distribution shape from your historical data.

Lookback Period: The Trade-Off

The choice of lookback period—how many days of history to examine—is critical and involves a fundamental trade-off.

Longer lookback periods (252-500 days): Capture multiple market regimes, rare events, and a broader range of outcomes. They're more statistically stable; a few unlucky days won't dramatically change the VaR estimate. However, they include very old data that may no longer be relevant. If the market regime has shifted—volatility has changed, new correlations have formed—old data contaminates the calculation.

Shorter lookback periods (63 days): Respond quickly to current volatility and market conditions. If volatility spikes, a 63-day lookback immediately reflects higher historical losses. However, with only 63 observations, rare events might not appear at all. If you're calculating 99% VaR (the 1st percentile), you need about 100 observations to reliably estimate the tail. With 63 days, you're extrapolating slightly beyond your data.

Practical approach: Most professional traders use a 252-day lookback as baseline, then supplement it with a 63-day lookback. If the two numbers diverge significantly—the 63-day VaR is much higher—that signals rising volatility and prompts position reduction or hedging.

Comparing Historical VaR Across Lookback Windows

Consider a portfolio with 252 days of returns:

• 252-day historical 99% VaR: -3.2% (from the 2nd worst observed day)
• 63-day historical 99% VaR: -4.1% (from the worst observed day in the past two months)

The 63-day figure is higher, indicating recent volatility has been worse than average. This is valuable information. If only the 252-day figure is used, you might be underestimating current risk.

Conversely, suppose the past two months have been calm:

• 252-day historical 95% VaR: -2.5%
• 63-day historical 95% VaR: -1.6%

The shorter window shows lower risk. A risk manager might question whether a single calm month should override a year's worth of historical evidence. This is judgment, not pure mathematics.

Real-World Example: Calculating Historical VaR Step by Step

Suppose you manage a three-asset portfolio: SPY, QQQ, and TLT (bonds). Your position weights are 50%, 35%, and 15%. Portfolio value is $100,000.

You gather 252 daily closing prices and calculate daily returns for each ETF. Then you calculate portfolio returns each day as:

Daily Portfolio Return = (0.50 × SPY Return) + (0.35 × QQQ Return) + (0.15 × TLT Return)

You now have 252 daily portfolio returns. Let's say they range from -5.8% (worst day) to +4.2% (best day).

You sort them from worst to best:

Rank 1: -5.8%
Rank 2: -4.1%
Rank 3: -3.9%
Rank 4: -3.2%
...
Rank 252: +4.2%

For 95% confidence (5th percentile), you want rank 252 × 0.05 ≈ rank 12.6, which you round to rank 13.

Rank 13's return was -2.4%. Your portfolio's 95% historical VaR is:

$100,000 × -2.4% = -$2,400

For 99% confidence (1st percentile), rank 252 × 0.01 ≈ rank 2.5, which you round to rank 3.

Rank 3's return was -3.9%. Your 99% historical VaR is:

$100,000 × -3.9% = -$3,900

Notice that you didn't calculate volatility or correlation. You didn't assume normality. You're reading the loss threshold directly from historical experience.

Handling Gaps and Missing Data

What if the stock exchange was closed on one of your 252 days due to a holiday? Or a security was delisted? The standard approach is to skip that day and include day 253 instead, so you always have exactly 252 observations. Some practitioners use a 251-day or 250-day window rather than forcing exactly 252. The exact number matters less than consistency; always use the same methodology to ensure comparability across days.

Advantages of Historical VaR

No distributional assumptions: You don't assume returns are normally distributed, independent, or anything else. The empirical data is your foundation.

Captures fat tails: If your history includes crashes, those crashes are reflected in the tail of the return distribution. Historical VaR "remembers" extreme days, unlike parametric VaR which averages them into volatility.

Intuitive: You can name the specific day in history that produced the VaR threshold. "Our 99% VaR is based on the August 5, 2011 flash crash." This transparency builds credibility.

Handles non-linear payoffs: Unlike parametric VaR, historical VaR works for options portfolios without modification. You simply calculate the historical P&L distribution, which already accounts for nonlinearity.

No correlation estimates needed: Correlations are baked into the historical returns. You don't need to estimate or update a covariance matrix.

Critical Limitations: Recency Bias and Structural Breaks

Historical VaR's greatest weakness is that it assumes the future will resemble the past. When market structure changes dramatically, historical VaR becomes dangerously optimistic or pessimistic.

Pre-crisis periods: Before the 2008 financial crisis, many portfolios' 252-day historical VaR was very low because the preceding year was relatively calm. This meant risk models said everything was safe even as systemic stress was building. Actual losses far exceeded historical VaR.

Structural breaks: If a market regime fundamentally shifts—volatility changes, new correlations form, or leverage constraints tighten—the old historical data no longer represents future risk. A 252-day lookback that includes calm periods becomes stale.

Rare events not yet observed: If your history doesn't include a -10% market move, your historical VaR might cap out at -8% (the worst you've seen). If a -10% move then occurs, you have no historical precedent for its probability.

The "2007 VaR problem": Hedge funds that relied on historical VaR in 2007 used data that included the 1990s and early 2000s—generally calm periods. When 2008 hit with leverage constraints and forced selling, outcomes were far worse than any day in the historical window. Historical VaR simply couldn't predict what it had never seen.

When Historical VaR Works Well

Despite its limitations, historical VaR is appropriate in several contexts:

Liquid asset classes without regime changes: Mature equity and bond markets with deep history are good candidates. U.S. Treasury bonds, large-cap stocks, and major currency pairs have long histories without severe structural breaks.

Supplementary risk measurement: Used alongside parametric VaR, historical VaR provides a reality check. If the two methods diverge significantly, that's a signal to investigate.

Stress scenario analysis: Historical VaR can identify past days with specific characteristics (high volatility, market gaps, geopolitical shocks) and use those days to stress-test the portfolio.

Options and derivatives portfolios: For nonlinear instruments, historical simulation is often more reliable than parametric VaR.

Real-World Divergence: Parametric vs. Historical

Consider a scenario during the COVID-19 market crash in March 2020:

On March 15, 2020:

• A parametric VaR model using 252-day volatility estimated losses wouldn't exceed 6% at 99% confidence.
• The same day saw actual stock market declines of 7-12% in many sectors.

Why the miss? The 252-day lookback included the previous year's relatively calm period. Parametric VaR's volatility estimate was too low because the build-up of leverage and systemic risk weren't visible in daily volatility numbers until the crisis itself unfolded.

On March 16, 2020:

• A 63-day historical VaR, recalculated after the 16% S&P 500 decline on March 16, would show much higher risk because the crash day was now in the lookback window.

Historical VaR adapts faster to regime changes than parametric VaR, but it still can't predict what it hasn't seen.

Adjusting for Non-Trading Gaps

When markets are closed (weekends, holidays), the historical VaR measured over 252 trading days spans about one calendar year. However, if you hold a position over a weekend, you face three calendar days of gap risk with only two trading days of history embedded in your lookback.

Some practitioners stretch the lookback to 378 calendar days to capture the same range of calendar time. Others use "business-day adjusted" volatility. The standard, however, is to stick with trading days; it's simpler and most brokers do the same.

Real Example: A Personal Portfolio

Suppose you hold $50,000: $30,000 in VOO (S&P 500), $15,000 in VGV (Growth ETF), and $5,000 in BND (Bonds). Over the past 252 trading days:

Your worst daily loss was -3.2% (March 16, 2020 crash). Your 13th worst day (95th percentile) saw -1.8%. Your 3rd worst day (99th percentile) saw -2.9%.

Your 95% historical VaR: $50,000 × -1.8% = -$900 Your 99% historical VaR: $50,000 × -2.9% = -$1,450

If the 63-day lookback shows:

Your worst day: -2.1% 13th worst day: -1.5% 3rd worst day: -2.0%

Your 63-day 95% VaR: $50,000 × -1.5% = -$750 Your 63-day 99% VaR: $50,000 × -2.0% = -$1,000

The divergence is modest here, suggesting volatility hasn't changed dramatically. But if the 63-day window included a recent crash, the 63-day VaR would be much higher, signaling increased risk.

Common Mistakes with Historical VaR

Using insufficient history: A 30-day lookback for 99% confidence (which needs about the 1st percentile) is unreliable. Stick to 252+ days for 99% confidence VaR.

Ignoring regime changes: If markets have fundamentally changed (new Fed policy, new correlations, deleveraging), old history is misleading. Recalibrate your lookback period or shift to a shorter window.

Assuming past extremes bound future ones: Just because the worst day in your history was -5% doesn't mean losses can't exceed -5%. Markets can produce unprecedented moves.

Not backtesting: Track your historical VaR estimate each day and check how often actual losses exceed it. If exceedances are far more frequent than predicted, your method needs adjustment.

FAQ

How do I choose between 252 and 63 days?

Use both. A 252-day lookback is your baseline. A 63-day lookback flags rising volatility. If they diverge materially, investigate. Many professionals set a rule: if the 63-day VaR exceeds the 252-day VaR by more than 25%, raise alert level and reduce position size.

What if I'm starting a new portfolio with no history?

For the first 252 days, use the most relevant comparable benchmark. If your portfolio is 80% stocks and 20% bonds, use the 80/20 historical returns of a broad market proxy until you have actual data.

Can historical VaR be zero?

Technically yes—if your portfolio never declined in the historical period. In practice, this is rare and indicates your history is too short or your holdings are genuinely uncorrelated with historical movements.

How do I handle corporate actions like stock splits?

Adjust the historical prices backward in time so the splits don't distort returns. Most financial data providers do this automatically, so daily returns are already split-adjusted.

Should I use log returns or simple returns?

For daily or weekly VaR, the difference is negligible. Use simple returns (percentage change) for clarity. Use log returns if you're doing multi-period analysis to avoid compounding errors.

What if my historical period includes a one-time crisis that won't repeat?

You could exclude it, but that's dangerous. If you'd excluded the 2008 crisis from pre-2009 VaR calculations, you'd have missed a real precedent. Better to use a dual-window approach: publish the 252-day VaR but footnote the impact of the crisis period. Or use 63-day rolling windows to naturally downweight old data over time.

Can I use historical VaR for intraday trading?

Yes, if you use intraday returns. Calculate intraday (5-minute or 15-minute) returns over 252 days, then apply the historical VaR method. Your threshold will be much tighter because intraday volatility is lower than daily.

Related concepts

• What Is Value-at-Risk?
• Parametric VaR: The Variance-Covariance Method
• Monte Carlo VaR: Simulating the Future
• Defining Investment Risk

Summary

Historical VaR offers a distribution-free approach to risk measurement: it ranks actual historical losses and extracts the percentile directly. This method requires no assumptions about normality, correlation stability, or volatility constancy. It works well for liquid assets with long, stable history and is particularly useful for nonlinear portfolios with options. However, historical VaR assumes the past resembles the future—a dangerous assumption when market regimes shift or unprecedented crises occur. The 2008 financial crisis and 2020 COVID crash both produced losses exceeding historical VaR estimates because market structure broke in ways prior data couldn't predict. Professional risk teams use historical VaR as a supplement to parametric VaR, comparing the two methods to detect regime changes and stress-test against scenarios not yet observed.

Next

→ Monte Carlo VaR: Simulating the Future