If you are walking with no particular direction, which way do you go? Beyoncé already gave it away because research shows that you are overwhelmingly likely to veer to the left or counterclockwise.
And while many hypotheses have been put forward as to the reason for this, such as the size and layout of the walking space, cultural norms, traffic rules, early childhood education, or which hand is your dominant one, none so far explain this almost universal phenomenon. It truly seems to be what it is. And since risk measurement often is characterised by less clear probabilities, I do appreciate a likelihood approaching 100%.
But then, what is the smartest way to figure out how things will go, risk- and return-wise?
For the most part, we rely on simulations. We look at the way risk factors, most often prices, have related to each other in the past and most recently, and we use that to either calculate or, if we are more advanced, simulate as many possible scenarios as we can and then see which one, with great certainty, is the worst that can happen. We call that value-at-risk (VaR). Equally, if not more, important, we estimate what the expected shortfall (ES) would be if we end up beyond the worst-case scenario.
We know that these simulation techniques are not perfect because we make all kinds of assumptions. The most unavoidable one is that history is a good predictor of the future. And even if we accept that premise, we might, for speed or cost reasons, have to cut corners by capping how much past we include or simplifying how tomorrow relates to yesterday. Or we can splurge on compute power to run through many, many possible scenarios to capture the rarest and most extreme ones.
There is, however, an alternative that is proving to be more accurate and faster, if not necessarily cheaper. For that, we have to turn to machine learning, or as it is known nowadays, AI, and a concept called Generative Adversarial Networks (GAN).
GAN has two components, a generator and a discriminator, that are, in this case, trained on history, so we are not giving up on that assumption. The generator generates (duh!) a synthetic scenario that could potentially happen based on the real history data. The discriminator judges, also based on the real history data, whether it believes the synthetic scenario to be real or not. The iterative process continues until the synthetic scenario is indistinguishable from the real data.
An even faster version of this is Wasserstein GAN (WGAN), which, instead of the discriminator, which just deems the synthetic scenario real or fake, has a critic that measures the distance to the real data.
The smart part of GAN is that rather than reducing the problem for speed or going through thousands and thousands of scenarios for completeness, GAN can generate fewer choice scenarios that close to perfectly capture the possibility space.
That is especially true for tail-GAN, which, as opposed to “regular” GANs, focuses on getting the stress or extreme (tail) events right in the synthetic scenarios. And, as you know, VaR and ES are both tail metrics, so that is a win.
WGAN and tail-GAN can be combined and further enhanced by augmenting and leveraging what is already known about the specific historical data at hand. Think of it as boosters for high-quality synthetic scenarios.
GAN and its flavours are thus an intriguing way to get faster and more accurate risk metrics as you veer left on the distribution and approach a high degree of certainty. This should, of course, be balanced against the efforts of continually training and updating your new generator-discriminator-critic buddies.
Luckily, FRG is here to help you with that.
Regitze Ladekarl, FRM, is FRG’s Director of Company Intelligence. She has 25-plus years of experience where finance meets technology.
This article is part of the FRG Risk Report, published weekly on the FRG blog. To read other entries of the Risk Report, visit frgrisk.com/category/risk-report/.