Core Path · L6.2.05

Monte Carlo VaR with fat tails

14 minMonte Carlo VaRMultivariate tt copulaGARCH pathsSimulation error

Know these first

By the end you will be able to

  1. 1Simulate a multivariate Student t with the data's covariance, fit its degrees of freedom by profile likelihood, and say what it adds to the normal model.
  2. 2Separate margins from dependence with a t copula on GARCH residuals, and explain why today's volatility moves the VaR more than the copula's ν does.
  3. 3Run GARCH-driven ten-day paths and compare them with √10 scaling.
  4. 4Put a standard error and an order-statistic interval on every simulated VaR, and decide whether two models really differ.

Why it matters

Over the last 500 days, history says the sleeve's one-day 99% VaR is £2,460,369. A normal model with exactly the same 500 days' covariance says £1,954,348: 21% less, on the same positions with the same full revaluation. The only change is the shape assumed for the tails. A Student t brings it back to £2,146,761. A model that knows markets are calm today says £1.43m. Monte Carlo can give any of these numbers. The work is in choosing the model, and in knowing how much of each number is noise.