Yann LeCun, arguably the father of modern machine learning, has described Generative Adversarial Networks (GANs) as the most interesting idea in deep learning in the last 10 years (and there have been a lot of interesting ideas in Machine Learning over the past 10 years).
From a trading perspective, or any game theoretic activity in which the game itself constantly evolves, GANs are big news. There is a school of thought in quantitative asset management, risk management, systematic trading etc that goes something like this:
“If I can just get enough data into my machine learning algo, then I win!”
This is wrong, because it doesn’t address the evolution of the game and the other players. It also comes up with all kinds of nonsense based on spurious correlations. Bigger data set = more false positives. Yet this approach is the basis for many startup hedge funds who get a lot of air time from the press. I’m not saying this approach doesn’t work, but it’s not the holy grail, and the results are often underwhelming, even if done correctly.
So what’s the big deal about GANs? They were invented by Ian Goodfellow in 2014. Here’s a video in which he talks about his work. It’s fascinating, and he explains his research really well. Ian Goodfellow talk about GANNS (watch from around 3:50).
I don’t want to get into too much technical detail (and if you’ve watched the video, you can skip this and the next paragraph), but here is the basic idea: A GAN is comprised of two Neural Networks: a discriminator and generator. You train the discriminator on real data to classify, say, an image as either a real photo or a non-photographic image. You then set a off generator which ‘experiments’ with trying to fool the discriminator into thinking the fake photo it generated is is from the training set. Equilibrium is reached when generator produces photos indistinguishable from real data.
But there is an other angle here. If your discriminator has any weaknesses, your generator can find and exploit them. So not only does it learn to make ‘counterfeit’ images, but it can configure data that can bamboozle the discriminator into thinking it sees something that’s not really there. In fact, you can use this approach to try and fool 3rd party models. ‘Adversarial attacks’, as these are known will increasingly become an issue in cybersecurity.
Agent based approaches
What is particularly interesting about GANs are they they have a lot in common with reinforcement learning models. These generate an agent policy from repeated interactions with the environment which is ‘scored’ by a reward function. The agent learns the best response to various environment ‘states’ over time. But to achieve this, we must impose a sate space representation of the environment onto the agent.
A GAN’s generator network interaction with the environment is mediated by the discriminator which acts as a binary classifier – so the heavy lifting around problem representation is done by training the discriminator. This means we get a game theoretic interaction framework without the need for a clunky state space representation.
‘Unlimited’ training data
The generator can create its own unlimited training set that is labelled by the discriminator. This is a powerful way of bootstrapping huge training sets from a ‘normal’ training set.
Given that the central problem of using Deep Learning models in business applications is lack of training data, this is a really big deal. So GANs should result in rapid advances in machine learning on unstructured data. Currently the focus is on video generation and enhancement, but the potential applications in finance where we have relatively limited, yet noisy data sets from a non-stationary, game theoretic domain are huge.
Deep Learning’s ‘killer app’ for finance
I would go so far as to say that any asset manager or bank that engages in strategic trading will be seriously competitively compromised within the next five years if they do not learn how to use this technology.
There are many possible uses for this technology, but the obvious ones are confusing the myriad of parasitic HFT algos trying to front-run your market orders or generating ‘photo realistic’ vol surfaces then harvest mis-pricings with your GAN generated option models.
This technology could, and probably should, form a pillar of next generation (big data and machine learning) risk management. Modelling realistic market behaviour has been the holy grail of risk managers since the widespread adoption of VaR in the 1990’s.
The problem with most other machine learning approaches to trading or asset allocation is that they implicitly assume the latent distributions in the training data are stationary. You can engineer ‘hacks’ around this (frequent re-training or calibration), but the core problem is trying to model and adapt to the chaotic non-stationarity caused by strategic agent interactions in the market.
Reinforcement learning does address this issue, but is works best on well defined state space problems (e.g. the game of Go, Video Games, etc). With a GAN you can engage in strategic interactions with a rich (unstructured) problem representation which looks more like real world problems.
Not just for trading
There is huge potential to use GANs for asset and derivative pricing, risk factor modelling and market parameter modelling. As the industry implements ‘next generation’ risk models, expect to see these popping up all over the place. We’ll still use the classic structural models of mathematical finance for generating risk reports, but practitioners that care about making money will use GANs and other empirical approaches.
Of course there are many applications in cyber-security, autonomous vehicles, and other business areas. This is new technology and I expect to see 100’s of use cases emerge over the next few years. Watch this space.