The Gaussian Assumption
The most widely used assumption in quantitative finance is that asset returns are normally distributed. From portfolio optimization (Markowitz) to option pricing (Black-Scholes), Gaussian returns underpin virtually every major financial model. But this assumption is dangerously wrong.
Fat Tails: The Real World
Empirically, financial returns exhibit leptokurtic distributions — they have fatter tails and sharper peaks than a normal distribution. What does this mean in practice? Events that a normal distribution says should occur once every 10,000 years happen roughly once a decade.
The 2008 financial crisis, the 2010 Flash Crash, the COVID-19 crash of March 2020 — each of these involved daily returns that were 5-10 standard deviations from the mean. Under a normal distribution, a 5σ event has a probability of approximately 1 in 3.5 million. Yet we see them repeatedly.
Kurtosis and Skewness
Two key statistics expose the failure of normality:
- Excess Kurtosis: Normal distribution has kurtosis = 3. Equity returns often show kurtosis of 5-10+, indicating extreme tail events far exceed what Gaussian models predict.
- Negative Skewness: Returns are negatively skewed — large losses are more frequent and severe than large gains, violating the symmetry assumption of the bell curve.
Alternatives: Modeling the Real World
Several distribution families better capture financial return behavior:
- Student's t-distribution: Heavier tails controlled by degrees of freedom parameter ν.
- Generalized Hyperbolic Distribution: Captures both skewness and kurtosis through additional parameters.
- Extreme Value Theory (EVT): Focuses explicitly on tail behavior using the Generalized Pareto Distribution.
Practical Implications
If you're building a risk engine assuming normality, your Value-at-Risk estimates will systematically underestimate the probability of catastrophic loss. This is precisely why the 2008 Basel risk models failed — they relied on Gaussian VaR thresholds that dramatically underpriced tail risk.
The lesson: always model the tails explicitly. Use EVT for extreme quantiles, fit heavy-tailed distributions for VaR calculations, and never trust a model that assumes returns behave like coin flips.