Fat Tails: What VaR Assumes and Should Not
Fat Tails: What VaR Assumes and Should Not
Markets are not normally distributed. Real stock returns, currency moves, and bond yields have "fat tails"—extreme events occur far more frequently than a bell curve would predict. This is one of the most important empirical facts in finance, and it's also one of VaR's deepest blind spots. Most VaR models assume returns follow a normal distribution, which has thin tails. Under a normal distribution, a 10-standard-deviation move (the kind of crash that saw the S&P 500 fall 20% in a single day) should happen once every 10 million years. In reality, such moves happen roughly once every 10 years. The difference between once-per-million-years and once-per-decade is the fat tail problem. Understanding this gap is essential for protecting yourself against the catastrophic losses that VaR confidently assures you are impossible.
Quick definition: Fat tails are the heavier-than-normal probability of extreme events (large gains or losses) in market returns. In a fat-tail distribution, 10-sigma events are far more probable than a normal curve predicts, making VaR estimates dangerously optimistic.
Key takeaways
- Real returns have fat tails: Empirical studies confirm that stock returns, foreign exchange, and commodities exhibit kurtosis (fat-tail behavior) consistently.
- VaR assumes normality: Standard VaR models use the normal distribution, which has thin tails. This assumption causes VaR to underestimate extreme loss probability.
- The kurtosis problem: Markets have "excess kurtosis" of 3–10 or higher, meaning extreme moves are 3–10 times more likely than a normal distribution predicts.
- Scaling compounds the error: When you scale 1-day VaR to 10-day VaR using the square-root rule, you assume normality. With fat tails, this scaling underestimates long-horizon risk.
- Crashes are more common than VaR expects: A loss that VaR says should happen once per 1,000 days might actually occur once per 50 days. This means crashes are 20 times more common than VaR predicts.
- Alternative distributions help: Using t-distributions or mixture models that allow for fat tails produces more realistic VaR estimates, though still not perfect.
What is a normal distribution, and why does VaR use it?
The normal distribution (bell curve) is the foundation of classical statistics. It's mathematically convenient, well-understood, and appears in many natural phenomena. In a normal distribution, about 68% of outcomes fall within one standard deviation of the mean, 95% within two, and 99.7% within three.
More importantly for VaR, the normal distribution has a clean mathematical property: if you know the mean and standard deviation, you can calculate the exact probability of any outcome. This is why the "variance-covariance" VaR method uses it—you can plug in the expected return and volatility (standard deviation) and instantly get VaR without simulation.
The calculation:
Daily return = mean + std_dev × Z-score
For 95% VaR with normality: Z-score = -1.645
95% VaR = -(mean + std_dev × -1.645)
Simplified (assuming mean ≈ 0):
95% daily VaR = 1.645 × daily_std_dev
This simplicity made the normal distribution attractive to risk managers in the 1980s and 1990s. A risk manager could calculate VaR on a spreadsheet with one formula. No simulation, no historical data crunching, just multiplication.
The problem: financial returns don't actually follow a normal distribution. They have fat tails.
The empirical evidence for fat tails
Decades of research have documented fat tails in financial markets. The evidence is overwhelming and consistent across asset classes, time periods, and countries.
Equity returns: Daily returns of the S&P 500 have been studied extensively. A normal distribution predicts that a 5-standard-deviation move (roughly a 7% daily loss for the S&P 500) should occur once per 3,000 trading days, or roughly once per 12 years. In reality, such moves occur much more frequently. Between 1925 and 2023, the S&P 500 experienced more than 50 daily moves of 5+ standard deviations. That's roughly once per 4 years, not once per 12 years. The tails are more than three times fatter than normality predicts.
Currency markets: Foreign exchange rates are equally fat-tailed. A normal model would predict that the euro-dollar exchange rate moves more than 3% in a day less than once per year. In reality, such moves occur roughly once per 20 days. Again, the tails are much fatter.
Bonds: Even government bonds, often treated as low-risk, exhibit fat tails. A 1% daily move in Treasury yields (a large but not unprecedented move) is 5–10 times more likely than the normal distribution predicts.
Cryptocurrencies: Bitcoin and other cryptocurrencies are famous for extreme volatility and crashes. Fat tails are even more pronounced in crypto than in traditional markets—crashes exceeding 30% in a day are not rare.
Kurtosis: Measuring the fatness of tails
Kurtosis is a statistical measure of how fat the tails of a distribution are. A normal distribution has a kurtosis of 3. Distributions with higher kurtosis have fatter tails and more frequent extreme events.
Kurtosis interpretation:
Normal distribution: kurtosis = 3 (baseline)
Fat-tailed distribution: kurtosis > 3
Very fat-tailed: kurtosis > 5 or 6
More useful than kurtosis is "excess kurtosis," which subtracts 3 from the total:
Excess kurtosis = kurtosis - 3
Normal distribution: excess kurtosis = 0
Fat-tailed distribution: excess kurtosis > 0
Moderate fat tails: excess kurtosis = 1 to 3
Severe fat tails: excess kurtosis > 5
Real-world measurements:
- S&P 500 daily returns: excess kurtosis ≈ 3–5 (depending on period).
- T-bond futures: excess kurtosis ≈ 2–4.
- EURUSD currency pairs: excess kurtosis ≈ 1–3.
- Bitcoin: excess kurtosis ≈ 10–20 (extremely fat-tailed).
An excess kurtosis of 3 means that extreme moves are roughly 3 times more likely than a normal model predicts. If the normal model says a 5-sigma move happens once per 3,000 days, the fat-tail reality is once per 1,000 days.
How fat tails break VaR
Assume you calculate a 95% daily VaR on an equity portfolio using the normal distribution. Let's say it's $100,000. Under normality, you expect to breach this threshold (i.e., lose more than $100,000 in a day) about once per 20 days.
But if the equity returns have excess kurtosis of 3, the true breach frequency is not once per 20 days—it's closer to once per 7 days. Your VaR is underestimating risk by a factor of nearly 3.
This compounds over time. If you hold the portfolio for a year (250 trading days), you expect about 12–13 breaches under normal assumptions (250 ÷ 20). But with fat tails, you might experience 30–40 breaches. The difference between "almost never" and "very frequently" is huge for risk management.
Real example: A bank calculates 95% daily VaR for its equity trading book at $10 million. Under normality, it expects to exceed this once per 20 days. The bank's board is comfortable with this "1-in-20 chance" and approves the position size.
But when the bank backtests its VaR, it finds that it breached $10 million almost once per week, not once per 20 days. The fat-tail property of equity returns—which the VaR model ignored—was causing the estimate to be wrong by a factor of 3. If the board had known the true breach frequency was 1-in-5 rather than 1-in-20, it would have demanded smaller positions.
The scaling problem with fat tails
Recall the square-root-of-time scaling formula:
10-day VaR ≈ 1-day VaR × √(10)
This formula is derived assuming normality and independent daily returns. With fat tails, both assumptions are violated.
Independence fails: On the day of a crash, subsequent days are more likely to also be down. Losses cluster. A normal model assumes today's loss tells you nothing about tomorrow's probability. But empirically, after a large down day, the next day is more likely to be down as well (negative autocorrelation, or mean reversion, is sometimes present, but during crises, momentum dominates). This means 10-day losses can exceed √(10) × 1-day VaR by substantial amounts.
Normality fails: The square-root scaling is only valid if returns are normal. With fat tails, the formula underestimates the true 10-day VaR. If 1-day VaR underestimates by 30% due to fat tails, then scaling by √(10) will underestimate 10-day VaR even more in relative terms because the tail gets fatter the longer you hold.
Practical consequence: A 1-day VaR of $100,000 might scale to $316,000 for 10 days under normality. But with fat tails, the true 10-day VaR might be $400,000 or $500,000. You think you're safe with $316,000 in capital reserves, but you're actually short by 25–50%.
Visualizing the fat tail difference
Why do markets have fat tails?
Several factors contribute to fat tails in financial markets, and understanding them helps you anticipate crashes:
Jump risk: Prices don't always move continuously. Sometimes, on bad news (a terrorist attack, a sudden default, a geopolitical shock), prices jump instantly. Continuous models assume you can find a buyer at any price. Jump models account for the discrete gaps. These jumps create the extreme moves that populate the tails.
Feedback loops: When prices fall, margin calls force traders to sell more, which drives prices lower, which triggers more margin calls. This cascade can create multi-standard-deviation moves in minutes. The positive feedback (price down → forced selling → price down more) creates tail events.
Correlation breakdowns: In normal times, certain assets are uncorrelated or negatively correlated. During crises, correlations converge to 1.0, amplifying portfolio losses. A portfolio that VaR models as diversified can suddenly become concentrated risk when correlation breaks change.
Liquidity shocks: In normal times, assets are liquid and traded in size. During panics, market makers vanish and bid-ask spreads explode. The "price" you see on your screen is not a price you can transact at. Large traders must accept worse prices, and this price impact widens the tails of the return distribution.
Informational cascades: When uncertainty is high, traders follow each other. If a few large traders sell, others follow (fearing they have private information), which causes more selling, which confirms the fears of others. This herd behavior creates fat tails.
Solutions: Fat-tail aware risk models
Several approaches address the fat-tail problem:
Use a t-distribution instead of normal: The student t-distribution (or Student's t) has fat tails by construction. It has a parameter called "degrees of freedom" that you can fit to your data. Lower degrees of freedom mean fatter tails. Fitting a t-distribution to historical equity returns and using that for VaR calculations produces estimates that are more realistic.
Use a mixture model: Combine a normal distribution (for calm days) with a separate distribution for tail events (for crash days). Most days come from the normal component; occasionally, a day comes from the crash component. This hybrid approach captures both normal and crisis behavior.
Historical simulation: Instead of assuming any distribution, just use historical losses directly. If the worst loss in your 20-year historical data is 8%, your 95% VaR is (roughly) the 5th-worst loss in your dataset. This method automatically captures fat tails because it's based on observed data. However, it can't predict losses worse than history—if your data doesn't include a 15% crash, historical VaR can't measure it.
Expected shortfall: Instead of VaR, use expected shortfall (conditional value at risk), which measures the average loss in the tail. This is provably more robust to fat tails and gives more information about extreme risk.
Real-world examples
The 1987 crash: On October 19, 1987 (Black Monday), the S&P 500 fell 22% in a single day. Under normal assumptions with historical volatility, this was a 20+ standard deviation event—a 1-in-100-million-year event. Yet it happened. Traders and risk managers who relied on normal-distribution VaR were shocked. A fat-tail model would still have been surprised, but not as unprepared.
The 2008 financial crisis: From September 2008 to March 2009, the S&P 500 fell roughly 55% peak-to-trough. That's a multi-standard-deviation move spread over months. Historical simulation VaR based on 1950–2007 data couldn't capture 2008 because 2008 was worse than anything in that history. Only a fat-tail model or a model using crisis periods like 1987 or the Great Depression would have flagged the risk.
The 2020 COVID crash: In March 2020, the S&P 500 fell 34% in 23 trading days—the fastest bear market ever recorded. The VIX (equity volatility index) spiked to 82.7, levels that normal models predicted should happen once per decade at most. Fat-tail models came closer to the mark, though even they underestimated the speed and intensity.
Common mistakes
Mistake 1: Trusting VaR from a model that assumes normality without checking kurtosis. Always examine the kurtosis of your asset returns. If kurtosis is significantly higher than 3, your VaR is underestimating risk. Adjust by using a fat-tail model or scaling up your VaR estimate.
Mistake 2: Assuming historical simulation VaR is immune to fat tails. Historical simulation captures observed fat tails, but it can't predict new types of crashes. If your data has never seen a 40% decline, historical VaR won't measure a 40% decline risk. Use a combination of historical simulation and stress testing.
Mistake 3: Forgetting that kurtosis changes over time. During calm periods, markets have moderate excess kurtosis (2–3). During stress, kurtosis can spike to 10 or higher. A VaR model built on calm-period data will be wildly wrong during crises. Recalibrate regularly, especially after market dislocations.
Mistake 4: Underestimating the impact of leverage. If you use leverage and markets have fat tails, your losses can exceed expected shortfall by substantial amounts. Leverage amplifies tail risk.
Mistake 5: Ignoring jump risk. Real prices jump (they gap overnight or gap at market open). Continuous models like Black-Scholes and simple VaR models don't account for jumps. If you own stocks or derivatives, recognize that a jump risk exists and could create a loss larger than VaR predicts.
FAQ
How do I know if my data has fat tails?
Calculate the excess kurtosis of your asset's historical returns. If excess kurtosis is substantially above 0 (i.e., above 0.5 or 1.0), you have fat tails. Excess kurtosis of 3–5 or higher indicates severe fat tails. You can also visually inspect a histogram of returns and look for visibly fatter tails than a normal curve would suggest.
Should I always use a fat-tail distribution instead of normal?
Yes, if you're serious about risk measurement. A t-distribution or mixture model is almost always more accurate than a normal distribution for financial returns. The only downside is slightly more computational complexity, which is trivial on modern computers.
Does using expected shortfall instead of VaR solve the fat-tail problem?
Partially. Expected shortfall (the average loss in the tail) is more robust to fat tails than VaR (the threshold of the tail). But expected shortfall still depends on your model's assumptions. If you use a normal model for expected shortfall, you're still underestimating. Use expected shortfall and a fat-tail aware model.
If markets have fat tails, why do option prices sometimes imply they don't?
Option prices sometimes imply kurtosis that's lower than historical kurtosis suggests, especially for short-dated options. This is because traders price based on implied volatility, which is partly a forecast of future volatility and partly a risk premium for jump risk and tail risk. The market's implied distribution (extracted from option prices) can differ from the realized historical distribution.
How much should I adjust my VaR if I know the excess kurtosis is 3?
A rough rule of thumb: if excess kurtosis is K, scale up your VaR estimate by roughly sqrt(1 + K/3). So for excess kurtosis = 3, scale by sqrt(1 + 1) = sqrt(2) ≈ 1.41. Your VaR estimate should be increased by 40%. This is approximate; a more precise calculation requires fitting the actual distribution.
Can stress testing help with fat tails?
Yes. Stress testing doesn't rely on any distribution assumption. Instead, you shock the market (e.g., assume the S&P 500 falls 20%, the VIX spikes to 50, etc.) and calculate the portfolio loss under that scenario. Stress testing captures tail risk directly, without assuming any distribution. Combine VaR with stress testing for a complete risk picture.
Related concepts
- What Is Value at Risk?
- Why VaR Fails During Market Crises
- How Banks Use VaR Internally
- Why VaR Failed in 2008
Summary
Real markets have fat tails: extreme price moves happen far more frequently than a normal distribution predicts. VaR models that assume normality significantly underestimate the probability and magnitude of crashes. Using t-distributions, mixture models, historical simulation, or expected shortfall can provide more realistic risk estimates. Understand your asset's kurtosis, regularly backtest your VaR, and supplement VaR with stress testing. Recognizing fat tails is the first step toward protecting yourself from the catastrophic losses that normal-distribution models claim are impossible but happen regularly.