Extreme Value & Tails

VaR drew a line at the 99th percentile; Monte Carlo manufactured a thousand bad days from a model. Both quietly assume the bell curve, so the tail is negligibly rare — yet that tail is where 1987, 2008, and the COVID gap all live, “one-in-a-billion” moves that show up every few years. This topic is about the tail itself: the part VaR points at but refuses to size, and that decides who is solvent on the other side of a crisis.

Here’s the machinery you’ll pick up, rung by rung:

• Fat tails vs. the Gaussian — why returns aren’t normal, what kurtosis measures, and how badly the bell curve underestimates the extremes it can’t see.
• Power laws & the tail index — the one number $\alpha$ that governs tail weight, the eerie scale invariance of a power law, and the Hill estimator that reads $\alpha$ off real data.
• Extreme Value Theory — both great results: block maxima → the Generalized Extreme Value distribution, and peaks-over-threshold → the Generalized Pareto Distribution, unified by one shape parameter $\xi$ that tells you if your tail is bounded, exponential, or heavy.
• Expected Shortfall beyond VaR — the average loss given you’re in the tail, why ES is coherent (it respects diversification via subadditivity, which VaR violates), and how an EVT-based ES extrapolates past the data.
• Copulas & tail dependence — why correlation is a fair-weather friend, and how a t-copula captures the “everything goes to one in a crisis” clustering a Gaussian copula structurally cannot.
• Stress testing — scenario and historical tests, reverse stress testing (working backwards from ruin to the shock that causes it), and the honest limits of all of it.

Master this and you’ve crossed from someone who can report a risk number to someone who understands what it omits — why expected shortfall is the number regulators moved to, and why the middle of the distribution was never the problem.