Back to the Future - The past is not always indicative of the future
Posted on 18-05-2009 by Craig Turnbull | 0 comments
The science of predicting changes in near-term levels of risk and the persistence of those changes has developed a great deal over the last 30 years or so. This has generally focused on predicting short-term asset volatility and using these changes to make inferences about tails, rather than being directly focused on predicting the tails. This is natural as it is a more tractable mathematical problem, and it is the right place to start. But there are fundamental limitations and dangers in using the most recent 90 days’ market behaviour to make predictions about what a 1-in-1000 or 10,000 day event might look like.
In early 2008, some financial institutions were declaring in their published financial reports that they believed they were capitalized to withstand a 99.97th percentile 1-year event. They have since been bailed out by their governments. Now, maybe we did just experience a 1-in-3000 year event. But the more likely explanation (putting aside that some risk and leverage was off balance sheet and effectively ‘model-exempted’) is that there was an extrapolation from volatility to tail that was systematically sanguine and self-serving.
Many modeling approaches exist that can do a better job of measuring tail risk – but their application and calibration will necessitate judgment as much as algorithms. Whether we like it or not, financial markets are not stable scientific systems in a perpetual equilibrium that produces infinite relevant historical data. That means that risk measurement has to be about more than a formula. So the formulas need to get more complex, but even more importantly, model calibration needs to be more rigorously scrutinized rather than being assumed to be a ‘solved’ algorithmic problem.
Please read our participation policy before submitting a comment.