Why VaR Fails During Market Crises
Why VaR Fails During Market Crises
Value at risk is a useful daily risk metric, but it becomes dangerously incomplete during market crashes. VaR tells you the maximum loss you might face on a normal day with 95% or 99% confidence, but it says nothing about the 1% or 5% of worst-case scenarios when markets seize up. This silence about catastrophic losses is VaR's most critical flaw. During the 2008 financial crisis, the 2020 COVID-19 market shock, and many other crises, portfolios lost far more than their daily VaR estimates suggested was possible. Understanding VaR's limitations is essential for anyone using it to make real investment decisions.
Quick definition: VaR limitations are the blind spots and mathematical assumptions that cause VaR to underestimate or completely miss extreme losses during market stress, liquidity crises, and tail events.
Key takeaways
- VaR ignores tail risk: It measures the boundary of normal loss, not what happens beyond it.
- Correlation breaks during crises: VaR assumes stable correlations; during crashes, all assets fall together, making diversification useless.
- Liquidity risk is invisible to VaR: A stock that trades normally most days can become unsellable during a panic, but VaR doesn't account for exit difficulty.
- Scaling breaks in stress: The square-root-of-time scaling that works in calm markets fails when volatility spikes.
- Model risk: VaR's accuracy depends entirely on historical data and the choice of model; different methods produce wildly different estimates.
- Backtesting reveals the problem: When a bank backtests VaR, it often finds that losses exceeded the 95% VaR threshold far more frequently than once every 20 days—sometimes as often as once every 2–3 days during volatile periods.
The tail risk problem: What VaR leaves out
Imagine you're sitting in a stadium with 1,000 people. A 95% VaR is like saying, "I'm confident the 50 people at the tail end won't suddenly rush the stage." This is useful information for planning—maybe you add more security. But it tells you nothing about what those 50 people will do if they do rush the stage. Will they knock over the barrier once, twice? Will they break equipment? VaR answers none of these questions.
In markets, the 5% worst-case scenarios (beyond the 95% VaR threshold) are precisely where the most devastating losses occur. A portfolio might have a 95% daily VaR of $500,000, meaning a loss worse than $500,000 should happen only 5 days out of 100. But when it does happen—when you hit that 5% tail—the loss might be $5 million, not $600,000. VaR doesn't measure how far into the tail you fall.
This is why many risk managers supplement VaR with expected shortfall (also called conditional value at risk), which measures the average loss conditional on exceeding the VaR threshold. If the 95% VaR is $500,000 and the 95% expected shortfall is $1.2 million, you know that on the rare days when you lose more than $500,000, you'll typically lose around $1.2 million. That's much more informative than VaR alone.
Correlation collapse during crashes
A core assumption underlying most VaR models is that correlations between assets are stable. This assumption is wrong. During normal market periods, gold and stocks might have a correlation of -0.1 or -0.2 (slightly negative, offering some diversification benefit). A portfolio with 50% stocks and 50% gold might seem well-balanced.
But during a market crash, when fear spikes and investors dump everything to raise cash, gold and stocks both sell off together. The correlation shoots toward +0.7 or +0.8. Suddenly, the diversification benefit evaporates and the portfolio losses exceed what VaR predicted based on normal-time correlations.
Real example: In March 2020, as COVID-19 fears spread, the S&P 500 fell 12% in a single week. Traditional safe havens like government bonds, gold, and commodities all sold off simultaneously. A portfolio that VaR models had estimated as "safe" because of diversification lost far more than expected. The 95% VaR estimated losses around 2%, but actual losses hit 5–6% in some balanced portfolios.
The problem is that VaR models built on several years of calm data don't capture the regime shift that happens during crises. When you're building a VaR model on data from 2015–2019, you're not seeing any crashes, so your correlation matrix is biased toward calm-market values. The moment stress hits, correlations revert to their crisis levels—which VaR doesn't know about.
Liquidity evaporates when you need it most
VaR typically assumes you can exit any position at the "market price" shown on your broker. In normal times, this is reasonable. You can sell 1,000 shares of Apple in seconds at the current bid. But during a market panic, liquidity dries up. Bid-ask spreads widen, volume evaporates, and if you're forced to sell a large block, you might have to drop the price 2–5% just to find a buyer.
For liquid mega-cap stocks, this is a minor cost. But for smaller stocks, corporate bonds, or illiquid derivatives, the cost of exiting can be enormous. A position that VaR valued as a $1 million loss might actually cost $2 million to exit if you need to sell quickly into a panicked market.
During the 2008 crisis, many mortgage-backed securities had prices listed on Bloomberg terminals, but no actual buyers existed at those prices. Financial institutions holding these bonds faced a choice: mark them at fantasy values (inflating their reported capital), or sell them at fire-sale prices (realizing huge losses). VaR models based on normal bid-ask spreads were completely useless for these assets.
Key insight: Liquidity is not constant; it's procyclical. When you need to sell, others are also selling, and liquidity vanishes. VaR doesn't model this dynamic. It assumes you can always exit at market prices, which is false during the crises when it matters most.
Volatility clustering and regime change
VaR models often assume that volatility is constant. The GARCH and other advanced methods try to capture time-varying volatility, but they typically assume that volatility evolves smoothly. In reality, volatility clusters and switches regimes suddenly.
A calm period with 10% annualized volatility can instantly shift to a 40% volatility crash. When this happens, the daily price moves that VaR considered "1-in-100" suddenly become "1-in-5." The model breaks because its core assumption—that the statistical properties of returns don't change—is violated.
During the 2008 crisis, volatility on the S&P 500 surged from 15% to 80% in a matter of weeks. A 95% VaR calculated on calm data would have been wildly inaccurate. The same with the March 2020 COVID crash: volatility jumped from 12% to 60% in days. VaR models based on 2019 data couldn't anticipate this shift.
Model risk: Different models, wildly different answers
Not all VaR models agree. A 95% daily VaR calculated using three different methods on the same portfolio might produce answers of $400,000, $600,000, and $800,000. Which is correct? The answer depends on the model's assumptions, and assumptions are often hidden or implicit.
Variance-covariance VaR assumes returns are normally distributed. This method is fast and simple but underestimates tail risk because real returns have fat tails (more extreme events than a normal distribution predicts).
Historical simulation VaR replays the worst historical losses, assuming the past is a good guide to the future. This method is robust but blind to new types of crises (e.g., a pandemic) that never happened in the historical dataset.
Monte Carlo VaR generates thousands of simulated price paths based on estimated parameters. This is flexible but highly dependent on parameter choice—change the mean or volatility estimate slightly, and VaR changes significantly.
Each method can be correct within its own assumptions, but assumptions differ. This means VaR is not an objective measure of risk; it's a model-dependent estimate. If you choose the wrong model, you get the wrong risk number, and you might not even know it.
During the 2008 crisis, banks using variance-covariance VaR (which assumes normal distributions) were especially blindsided. Their models had calculated risk as low, right up until markets collapsed. Historical simulation models performed slightly better because they at least included the extreme moves from 1987, but even those models failed because 2008 brought new types of moves never seen before.
Scaling failures under stress
Under normal conditions, the square-root-of-time scaling rule (10-day VaR ≈ 1-day VaR × √10) works reasonably well. But during crises, this breaks down.
The square-root rule assumes that daily returns are independent and that volatility is constant. During a crash, neither assumption holds. Days are not independent—each down day makes the next down day more likely as fear spreads. And volatility is not constant; it spikes.
A 1-day VaR of $100,000 might imply a 10-day VaR of $316,000 under normal conditions. But during a crisis, the actual 10-day VaR might be $800,000 or $1 million because the market can fall multiple days in a row without reversing. The scaling rule is too optimistic.
This is one reason why institutions use "stressed VaR," which calculates VaR based on a historical crisis period (e.g., September 2008) rather than on calm-market data. Stressed VaR captures the correlation breakdown, volatility spikes, and multi-day momentum that the square-root rule misses.
Historical VaR fails on new crises
Historical simulation VaR can only measure losses as bad as the worst loss in the historical data. If the worst month in your 10-year dataset saw a 15% decline, historical VaR can't predict a 25% decline—it simply won't be in the data.
This was the problem in 2008. Most banks' historical data included the 1987 crash, but 2008 was structurally different. New instruments (credit derivatives), new leverage levels, and new feedback loops (mortgage-backed securities) created losses that exceeded historical precedent. VaR models couldn't see it because they weren't trained on anything like it.
The same problem recurred in March 2020 with COVID-19. The speed of the market decline and the scale of the government response were both historically unprecedented. VaR models based on 2015–2019 had no frame of reference.
The false precision problem
VaR creates a veneer of precision that is often false. Saying "95% VaR is $497,534" sounds precise and scientific. But this false precision can lull traders and risk managers into complacency. They believe they understand risk because a number is quantified. They don't realize how much uncertainty underlies that estimate.
In reality, when you calculate a 95% VaR, you're not measuring risk with scientific precision. You're making assumptions about correlation, volatility, distributions, and liquidity—all of which are uncertain and model-dependent. The true confidence interval around a VaR estimate is much wider than the number suggests.
This false precision has led to multiple financial disasters. Before 2008, major banks prominently displayed their VaR numbers to show they were well-managed and risk-aware. Investors saw the low VaR and assumed the banks were safe. Then the crisis hit, losses exceeded VaR by factors of 5 or 10, and the banks failed or required government bailouts. The false confidence in the precision of VaR contributed to the disaster.
Visualizing why VaR misses tail risk
Real-world examples
Barings Bank 1995: Nick Leeson traded futures contracts in Singapore without proper risk limits. His bank's VaR model showed the position as low-risk because it was hedged on paper. But the hedge didn't actually exist as a binding contract—Leeson had fabricated it. When the Kobe earthquake hit and markets fell, the "hedge" disappeared. Barings lost $1.3 billion, far exceeding any 95% VaR estimate based on normal market moves. The failure wasn't mathematical; it was that Leeson's fraud broke the model's assumptions. But the lesson is that VaR is only as good as the data and assumptions fed into it.
Long-Term Capital Management 1998: LTCM built a legendary hedge fund on sophisticated quantitative models, with Nobel Prize-winning partners. The fund calculated very low VaR. Then the Russian government defaulted on debt, credit spreads blew out, and market correlations inverted. Losses were catastrophic—$4.6 billion gone in weeks—far exceeding what the models predicted. The fund would have collapsed entirely except the Federal Reserve orchestrated an emergency rescue. LTCM's models were mathematically elegant, but they underestimated tail risk.
JPMorgan 2012 London Whale: A trader at JPMorgan named Bruno Iksil built a massive derivatives position that he believed was well-hedged. The VaR on the position was reported as low to moderate. But correlations shifted, the hedge proved imperfect, and losses snowballed to $6.2 billion. The VaR estimate had missed the tail risk in the position. Even at a major bank with world-class risk systems, VaR failed to capture the actual exposure.
Common mistakes
Mistake 1: Relying on VaR as the sole measure of risk. VaR is useful but incomplete. Always supplement it with expected shortfall, stress testing, or scenario analysis. A trader who only looks at 95% VaR is flying blind to the 5% tail risk.
Mistake 2: Assuming VaR is measured with high precision. A 95% VaR of $500,000 sounds precise, but it's really more like $400,000 to $600,000 depending on the model and assumptions. Treat VaR estimates as rough guidance, not gospel.
Mistake 3: Expecting VaR to work the same way in crises as in calm markets. VaR models trained on calm data don't anticipate how correlations, volatility, and liquidity shift during crashes. If you want to predict crisis losses, use stressed VaR or crisis-period data.
Mistake 4: Forgetting that VaR assumes you can exit positions. If you're holding illiquid assets or large positions that would require time to sell, the true loss during a crisis could be much larger than VaR predicts. Adjust your risk tolerance down accordingly.
Mistake 5: Confusing VaR confidence levels. A 95% daily VaR means a 5% chance of a loss exceeding that amount. A 99% daily VaR means a 1% chance. The 99% VaR is higher (worse) because the tail is fatter. If you're reporting to regulators, make sure you're reporting the right confidence level.
FAQ
Is VaR useless then?
No. VaR is useful for daily risk monitoring and understanding normal-market exposure. The key is not to over-rely on it. Use it as one tool alongside stress testing, expected shortfall, and scenario analysis. VaR is good for answering, "What's a typical worst-case day?" It's bad for answering, "What if everything falls apart?"
Why do banks still use VaR if it fails during crises?
Regulators mandate it, and it's useful for daily trading risk management. Banks do use it—they just supplement it with other measures. The Basel Accords require banks to hold capital based on VaR. The Federal Reserve also requires stress testing and expected shortfall. VaR is part of a toolkit, not the entire toolkit.
What should I use instead of VaR?
There's no single replacement. Different situations call for different measures:
- For daily trading risk: 1-day VaR plus expected shortfall.
- For risk of ruin: Monte Carlo simulations of long-term portfolio paths.
- For regulatory compliance: VaR plus stress testing and expected shortfall.
- For crisis preparedness: Historical or hypothetical stress scenarios, reverse stress testing.
Can you improve VaR to make it work during crises?
Somewhat. Using longer historical periods (to capture past crises), switching to distributions with fat tails (to overweight extreme scenarios), and regular stress testing all help. Conditional VaR (expected shortfall) is provably better than plain VaR at measuring tail risk. But no model can perfectly predict the next crisis, because the next crisis is always, by definition, something new.
Did the 2008 crisis cause banks to abandon VaR?
No, but it did lead to regulatory changes. After 2008, the Dodd-Frank Act and Basel III imposed stress testing and expected shortfall alongside VaR. The SEC also required public reporting of VaR for large investment firms. The lesson was: use VaR, but never let it be the only measure of risk.
How often do VaR threshold breaches actually occur?
In a well-calibrated model, a 95% daily VaR should be breached (i.e., actual loss exceeds the VaR estimate) about once every 20 days, or roughly 2–3 times per month for an actively traded portfolio. In practice, many portfolios breach much more frequently, suggesting either poor model fit or that the model is underestimating risk. Regular backtesting (comparing actual losses to VaR predictions) can reveal this.
Related concepts
Summary
VaR fails during market crises because it ignores tail risk, assumes stable correlations and liquidity, and relies on historical patterns that may not repeat. When correlations collapse, liquidity evaporates, and volatility spikes—precisely during the crises you most need risk protection—VaR becomes dangerously incomplete. It measures normal-day risk well but says nothing about extreme losses. Traders and institutions must supplement VaR with expected shortfall, stress testing, and scenario analysis to truly understand downside risk.