The headline panel of the event had some of the best known quants in the industry, Damiano Brigo (Imperial College), John Hull (University of Toronot), Peter Carr (NYU Tandon) and Vladimir Piterbarg (NatWest Markets). The major topic was how new trends were impacting quant finance. On the subject of machine learning there was still largely skepticism. Peter Carr noted that it wasn’t new, citing papers which were 20 years old, which were using ideas like neural networks to do option pricing. The difficulties of interpretability were also flagged by Brigo.
Piterbarg noted how advances in machine learning in other areas like Google Translate and image recognition, were not matched by equally impressive advances in finance. Whilst, there might be applicability for high frequency datasets, it was not as easy for lower frequency data, such as daily. However, John Hull very much disagreed, flagging how machine learning was being used to speed up numerical solving. He advised people to learn as much as possible on machine learning, and he said it would have a big impact on finance. He has spent the past 12-18 month learning about machine learning and he is now writing a book on the subject. He noted that in finance in practice, we already use many black box models and there were new techniques being researched to look at the issue of interpretability.
Machine learning featured (excuse the pun!) in a number of different talks more broadly. Marcos Lopez de Prado’s talk was a general overview of the topic. He noted how machine learning could be used to model non-linear relationships and could handle unstructured data. However, there was a need to be careful when dealing with time series. He cited several examples, including portfolio construction. He also noted that it could be used for parts of the investment process, for example, in price prediction, to size bets (in combination with another model for the trade direction). Patrik Karlsson and Hanna Hultin (SEB) talked about how it was possible to use reinforcement learning to help simulate the execution process to develop algos.
Also on the trading front, Mark Higgins (Beacon) showed how it was possible to use a neural network which could allow better for hedging than Black-Scholes. Richard Turner (Mesirow Financial) also presented ways of doing portfolio optimisation for FX strategies looking at expected shortfall and also introducing how machine learning might be applied in this area.
We often think of machine learning as a way of developing a non-linear model to describe relationships between variables. However, there were a number of talks at QuantMinds which instead tried to use machine learning to help evaluate existing models. Katia Babbar (AI Wealth Technologies) discussed how to use deep learning to price exotic options.
The objective in using a deep neural network was to make the pricing of options more efficient, ie. using the same pricing models, but having a more efficient numerical solution. Blanka Horvath (Kings College), Mehdi Tomas (Ecole Polytechnique) and Aitor Muguruza (Imperial College) presented a way to calibrate rough volatility models in a few milliseconds for the whole implied volatility surface. The idea was to do the learning step offline, and the coefficients of the learned model could be used later to do (fast) calibration when needed.
Matthew Dixon (IIT) presented a paper on how to use Gaussian processes to do CVA computation. On the subject of interest rates, Marcos Carreira (Ecole Polytechnique) discussed various schemes for interest rate interpolation. When it came to learning from your dataset, he noted that it was important to know which regime you were in.
The area of quantum computing also cropped up in a talk by Davide Venturelli. He stressed that we are still very far away from a quantum computer which can be used to break RSA encryption. Whereas bits in computing can 0 or 1, qubits, their quantum equivalent can be both 0 and 1 at the same time. They are easy to create, he noted, but they are very difficult to control. Hence, the probability of errors are very high, and hence error correction needs to be employed to use them. One interesting point he noted was that quantum computing was not purely about speed. A quantum computer also used a lot less energy.
Machine learning needs data! Hence, it’s not surprising that in recent years the area of alternative data has become a key area in finance. Saeed Amen of Cuemacro (the author of this post) introduced the topic of alternative data in his talk. The main objective was to discuss how alternative data could be valued and to give some use cases, which included using CLS FX flow data to trade FX and the use of Bloomberg News articles to generate FX signals. Matt Napoli (1010 Data) talked also talked about alternative data, and in particular at the datasets his firm distributed, which revolved mostly around credit card transaction data. He explained the difficulties in structuring this data, with appropriate tagging and how his firm had invested a lot of time in doing this.