Most decisions assume a world where averages are meaningful, samples are informative, and yesterday predicts tomorrow. That world exists, and a lot of insurance is built on it. But it isn't the only world — and confusing one for the other is how organizations get blindsided. Before you reach for an average, ask which world you're in.
The familiar world
Mediocristan
Outcomes cluster around an average. The biggest observation in a large sample is rarely much bigger than the typical one. Sampling works. The past predicts the future.
Examples
- Heights, weights, IQ scores, blood pressure
- Routine auto claim sizes
- Daily call-center volumes
- Employee tenure, time-to-hire
- Standard mortality experience
Averages mean what they sound like. Sample sizes work. Past behavior predicts future behavior. Standard actuarial methods apply.
The dangerous world
Extremistan
A single observation can dwarf the entire rest of the dataset. The worst case you've seen is almost certainly not the worst case there is. Averages mislead.
Examples
- Catastrophe losses (hurricane, wildfire, earthquake)
- Cyber events, especially aggregated
- Pandemic-related losses across lines
- Social-inflation-driven verdicts and class actions
- Reinsurance recoveries in a stressed market
- M&A outcomes, startup returns, market crashes
Averages mislead. Sample sizes lie until the rare event hits. Past behavior predicts future behavior right up until it doesn't. Tails are fatter than your data shows.
Any one of these is a flag. Two or more, and you should assume tail-risk thinking applies.
- The variable can scale without natural limit. Heights cap out around eight feet; verdicts and catastrophe losses don't cap out at all.
- Outcomes depend on networks, contagion, or feedback loops. Where one event triggers others — reinsurance failures, market panics, viral litigation theories — distributions get fat in a hurry.
- The underlying process can shift. Climate, technology, regulation, social attitudes. When the generator of the data is non-stationary, historical base rates understate forward risk.
- The historical worst case keeps getting broken. Wildfire losses, cyber events, hurricane-driven losses — when "100-year events" arrive every five years, the distribution wasn't what you thought.
- Averages hide concentration. If a small fraction of policies, accounts, or events drives most of the loss, you're not really dealing with an average — you're dealing with a few outliers and a lot of noise.
- The downside is much bigger than the upside. When a good year saves you a few points and a bad year threatens solvency, the asymmetry itself is the signal.
If you're in Extremistan
What changes
- Stop reasoning from averages. Reason from worst plausible cases. The expected value is a number; the tail is the story.
- Distrust your data. Your sample almost certainly understates the tail. Assume the next worst case is bigger than the last one, not smaller.
- Protect against ruin even at the cost of expected value. Insurance, hedges, reinsurance, capital buffers, exposure limits. Paying for protection looks expensive in calm years and cheap exactly once.
- Barbell, don't balance. A mix of extreme caution on the core and bounded aggression on the periphery beats a uniform medium-risk posture. Most of the book ultra-conservative; a small slice taking real chances with bounded downside.
- Subtract before you add. Cutting catastrophic exposure beats finding clever upside. "Don't do anything stupid" is the strategy.