Why Normal Models Appear Often
When many similarly sized effects add and no single term dominates, sums and means can be close to normal. That helps justify Gaussian models for some measurement errors and sensor noise. Convenience is not enough, though; the generating process and diagnostics must support the assumption.
Do Not Treat the Bell as a Default Commandment
- Income and wealth often have long right tails and may fit lognormal or Pareto-family models better over relevant ranges.
- Word frequencies and city sizes can show rank-frequency or power-law-like behavior.
- Financial returns often have heavier tails and time-varying volatility, so a thin-tailed Gaussian can underestimate extreme risk.
- Mixtures of groups can be bimodal or multimodal, making one mean unrepresentative of every group.
Inspect histograms, quantile plots, residuals, and tails, and ask how the observations were generated. A probability model compresses reality; it does not dictate reality.
A risk-modeling lesson. Failures such as LTCM cannot be reduced to one normal-distribution assumption. Leverage, liquidity, changing correlations, and model risk interacted. The narrower lesson remains: relying on thin-tailed models can understate extreme losses.
Closing the Track
The normal distribution is a powerful compression tool. Mean and standard deviation, the empirical rule, z-scores, and the central limit theorem let you read data quickly—once you ask whether a bell-shaped model matches this data and its generating process.