Today saw RE•WORK hosting the first day of the Deep Learning in Finance Summit in London.
This morning’s sessions were dominated by discussions on explainability and interpretability, and the afternoon turned towards the technicalities of applying DL in finance, as well as honing in on fraud detection.
Attendees came together to explore the most cutting-edge academic and industry advances in Deep Learning in Finance, and we’re taking a look at some of the key learnings of day 1 of the summit:
I’m feeling very privileged to have heard insight into the latest Deep Learning projects in Finance, they are really at the forefront of business applications and global research. — Hollie Jacques, Summit Organiser
“@teamrework I’m very much enjoying the summit, especially the second part of the morning session where speakers went into details of the models and provided a list of suggested readings.” @mervealanyali
The morning’s sessions kicked off with Jackson Hull, — CTO & COO — GoCompare Group speaking on Improving Customer Experience using AI:
“I hope I can provide you with some of the importance of data science on customers. We need AI to give us a winning differentiator and transform the customer experience. We’ve recognised the threat of ‘go AI or die’ which is why we’re using these technologies. We provide direct real-time quotes for the customer which means we can predict what they need at the moment they need it. This yields much higher sales as we’ve taken the challenges out of it for the users as we use the data to provide the best experience for them.”
Discussions also turned to ethics in the Q&A sections with an attendee posing the challenging question:
Q: Given the amount of personal data you have, how do you account for bias and ethical risks?
A: Really, the only way to do it is to widen data sets to make sure we’re not segregating a particular type of consumer.
Continuing on the theme of customer experience was Huma Lodhi from Direct Line Group who highlighted the unique challenge they’re facing:
“Insurance customers range from millennials to someone who’s 97 years old. A customer who is at either end of this spectrum will have different preferences and needs, so it’s important for insurers to understand all types of customers.”
What else did we learn today?
Manuel Proissl
Head of Predictive Analytics in Banking Products, UBS
On the Human Element in Building Trustworthy Consumer-Centric AI Products in Banking
“The key here, as it says in the title of my presentation, boils down to trust and building a robust ecosystem around observational data models to deal with the interpretability of the black boxes of deep learning. It’s important to put this in front of your products as it depends on the ways in which human interpretation was used when the data was created.”
Katia Babbar
Visiting Research Fellow, University of Oxford
Exotic Options Valuations Using Deep Learning
“In the current financial market, customers are impatient and demand to know now. Not in 3 days or 5 days. This means we can’t use the complex deep learning models that are being developed in academia. The preference has also been skewed away from deep learning towards analytics because it gives a common language in terms of pricing, but this, of course, compromises accuracy. People have started creating GPUs to try and solve this. There are simple finance derivatives, and there are more complex ones.”
Mehrdad Mamaghani
Principal Data Scientist, Swedbank
Anomaly Detection using Deep SVDD
“Fraud schemes are living organisms, there’s no such thing as a static system; the complexity is also on the growth due to the variety of disruptions. We have a lack of access to recent cases because they’re still unfolding, so we don’t know exactly what we’re looking for. Once we know something deviates we know it could be fraudulent. Our work of anomaly detection is also used in healthcare, cybersecurity, imaging, and logistics as well as finance. Anomaly detection is used with supervised, clean/semi-supervised and unsupervised learning.”
Stefan Zohren
Associate Professor (Research), University of Oxford
Deep Learning Models for High-Frequency Market Microstructure Data
“Everyone here knows why DL is interesting and has lots of applications in Finance, but it’s very different to just take DL methods and try and apply them in finance because of problems like low signal-to-noise ratio and non-stationary. Here I’m focusing on HF data which is arguably more amenable.”
Hamaad Shah
Lead Data Scientist, Deutsche Bank
Using Generative Adversarial Networks to Estimate Value-at-Risk for Market Risk Management
“Usually when we talk about GANs we think of images, so I thought it would be nice to see if we can do the same in finance. We look at the portfolio returns distribution samples from an untrained GAN, specifically here it’s a BiGAN. The trained GAN output results are remarkably similar to the real results.”
Luigi Troiano
Professor of AI, Data Science & ML, University of Sannio
Algorithmic Fairness in Finance
“There have been several talks on fairness so far, but generally we think algorithms can be ‘unfair’: let’s start a discussion about this — is it the algorithms that are unfair, or is it the humans that are unfair in developing the models?”
“We talk about algorithms as a black box, but I prefer to call them a blind box. It’s not that we don’t know what’s going on, but it’s that what we can observe is limited. We can understand, but not see.”
Rich Radley
Head of Customer Engineering, Financial Services, Google Cloud
Applying Transparent and Explainable Deep Learning in Financial Services
“We’re working on making deep learning more auditable to satisfy regulations. From a trust and transparency stance, it goes without saying that we need to reduce bias. We also need to be able to explain in human terms, the factors that went into the model’s prediction. On the interpretability side, why is this important? Increasingly, regulation is driving the need to be on top of our game in this space. Model interpretation is so important as it increases understanding and trust.”
Ammar Belatreche
Senior Lecturer & Programme Leader, Northumbria University
Leveraging Artificial Intelligence in Stock Market Manipulation Detection
“We’ve tried several techniques to craft and engineer AI to detect the manipulation, but we don’t have the labelled information to know what is manipulated. The only way left is to inject synthetic information into the data.”
PANEL:
Considerations for Utilising AI to Assist Security Efforts and Enhance Safety in Finance
Moderator: Katia Lang, CEO, The FinTech Times; Mason Edwards, Customer Engineer, Google Cloud; Adam McMurchie, Head of DevOps, Barclays; Adri Purkayastha, Head of Technology & Cyber Risk Analytics, BNP Paribas Group; Lorenzo Cavallaro, Professor & Chair in Cybersecurity, King’s College London
- Adam: “I do infrastructure and security — if you’re in Finance, your goal is really to generate revenue. This means the fraud side is the opposite — you’re most interested in not hitting the headlines. A big company would much rather maintain its customer base than gain 1%. It’s not a case of how strong is the model, because it has to be far far better than the big bank fraud models that have been very good for the past 25 years. At a bank I worked with in the past, there was an additional field added in a model and that one additional field required 6 months to add, and that’s not feasible in a rapidly changing environment.”
- Adri: “Modern governance, are we really able to explain the models used across risk? Increasingly we get vendors who say they’re using ML and AI from intrusion detection to everything else. What about adversarial AI? In some situations, we’re finding the adversaries are compromised, so that’s an area of challenging possibilities.”
Zhengyao Jiang, Zhengyao Jiang
Student, University of Liverpool
Cryptocurrency Portfolio Management with Deep Reinforcement Learning
“Reinforcement learning aims to solve Markov Decision process where there are a state and an action, and the RL learns the policy from the data. Deep RL uses a DNN to represent the policy — we’ve seen achievements here in board games and video games, so I’m now considering using DRL in portfolio management. If we could make this work, there would be no need to worry about transforming predictions to profits, but applying general purpose DRL is difficult because it has very low data efficiency — there is no simulator.”
Alexandre Combessie
Data Scientist, Dataiku
GANs for Option Pricing in Real-Time
“ I’m a data scientist so I like to put things into action. This is about how we applied a GAN for option pricing in real-time in 10 days.”
“We need to go deeper, academia is very important. Our GAN model worked, but we realised that it’s not that something works because it’s perfect — if it works, it might still not be good enough.”
Paulo Rosário
Senior Economist — M&G Prudential
Deep Learning in Analysing Large Structured Datasets for Investment Purposes
“Company directors from CEOs to CTOs, when they trade their own shares they have to announce it to the market. This is good because it lets people know if they think the market is unstable, but it doesn’t actually necessarily mean this — high level finance experts can trade for many reasons.”
Throughout the day, as well as presentations and panel discussions there were interviews and podcasts recorded, as well as attendees participating in workshops. We’ll be back for another day of Deep Learning in Finance tomorrow, and will be publishing a full summary at the end of the day, so watch this space! If you’re interested in getting access to the video presentation from the summit, let us know here.
