With technological enhancements increasing computing power and decreasing its cost, easing access to big data and innovating algorithms, there has been a huge surge in interest of artificial intelligence, machine learning and its subset, deep learning, in recent years. The popularity of smartphones, wearables and social media platforms has led to an explosion in the amount of data being recorded and AI is the only way to make use of it. With the surge of digital disruption in the financial services, the FinTech industry has led to hundreds of emerging startups bringing new ways for people to bank, which is causing traditional methods to undergo an innovation overhaul to integrate new technological advancements in order to compete. To celebrate London's 3rd Annual FinTech Week 2016 (15-22 July), we spoke to experts in the field to find out how and why deep learning is disrupting finance, and, most importantly, what we can expect in the future.
Jan Hendrik Witte is Data Scientist at GreyMaths, where he is is building deep learning technologies for the use in trading and investing. A mathematician by training, Jan is interested in the areas of numerical mathematical finance, systematic trading, and portfolio optimization.
Yuanyuan Liu is Senior Manager of Quantitative Analytics at AIG, where he has led multiple global projects in loss-risk analysis, client lifetime value modelling, submission prioritization, and opportunity seizing, using advanced machine learning techniques. Most recently, Yuanyuan is working on applying algorithms in insurance, using generative model, computer vision, and NLU.
Hitoshi Harada is CTO of Alpaca, where they develop deep learning software to optimize workflows and scale human expertise. Before Alpaca, Hitoshi worked in the database industry and community for ten years, and has experience in data science, machine learning and image processing for industrial applications.
Alesis Novik is Co-Founder of AimBrain, a mobile banking fraud security focused company whose core technology is fuelled by the recent advances in deep learning algorithms. Alesis aims to provide the market with a smart, secure and transparent biometric experience.
Peter Sarlin is Director of RiskLab Finland and Associate Professor at the Hanken School of Economics, and is currently working on systemic risk, machine learning and visual analytics. Peter has extensive experience in financial stability, and has worked with large institutions such as the European Central Bank and the Bank of Finland, among others.
What have been the leading factors enabling recent advancements and uptake of deep learning?
Jan: Astonishing increases in computing power and data availability in recent years have been the main benefactors of deep learning technology.
Hitoshi: Some of the easily understandable applications, such as image recognition, video captioning and beating the world champion of Go, are pushing people hard to be excited. From a technical perspective, the generality and high accuracy that deep learning has is the main motivation for using it instead of other machine learning methods. In our case, for example, our AI engine learns how traders trade from the technical chart, no matter what kind of strategy or what kind of indicators they use.
Alesis: The computational power and tools to utilize that power has definitely enabled the recent advancements in Deep Learning. Regarding the uptake of deep learning, from the business side, it has become the best tool for the job. Because deep learning is producing better empirical results than other methods in a lot of tasks (for example most problems related to computer vision), companies adapt it to improve their services. At AimBrain we have been using Deep Learning for facial and voice biometric authentication from day one, achieving state of the art results. We are now applying it to behavioral biometric authentication with the same effect.
Yuanyuan: I would vote for the exponential growth in computational power, especially the boosting usage of GPUs, as well as widely available open-source packages. These changes enable researchers to quickly try new algorithms, share knowledge and tune models. In addition, these have dramatically lowered the entry threshold for deep learning practioners.
What is essential to continuing this progress?
Hitoshi: The processing speed is one of the factors I see advancing for the foreseeable future. Deep learning inherently needs lots of computing resource as well as data, and for the next few years I think some of the underlying technology like GPU and storage/network will surely advance and deep learning results will continue to expand. Of course, the size of community also needs to grow more so that more ideas and applications should jump into this area to advance the technology.
Alesis: As long as Deep Learning will present leading results, there will be interest and research going into it. From a more practical side, especially for start-ups, there is still no good cloud solutions for training deep neural networks. At AimBrain we've been using our in-house GPU servers to run experiments and test new architectures. We believe easier access to such machines would result in wider adoption of Deep Learning.
Yuanyuan: Technology innovations (not only from NVIDIA, but all relevant others), industry focus (not only tech firms but financial institutes), mature deep productions (e.g. automated vehicles, accurate virtual assistant, financial practices etc).
What area of finance do you feel will be most disrupted by AI?
Jan: Imminent overhaul is to be expected in retail financial services (including communication, aggregation, and security and accessibility). In the long run AI, will completely change our investment industry, but (certainly on the institutional investment side) we are only at the beginning of a long and slow transition of 50+ years.
Peter: If not a full disruption, then AI is at least to impact most industries in the coming years. Within finance, I feel AI will likewise impact the full spectrum of financial services. While asset management has, and continues to be disrupted in several ways, much more traditional banking activities are also to be impacted, such as bots for handling simple tasks in personal banking.
Hitoshi: We believe the capital markets will be the most disrupted. Both retail and institutional trading will benefit from AI automation and more development of interesting strategy. This includes cover dealing in brokers and sales trading as well as personalized investment and risk control.
Alesis: We might be biased, but security will definitely be one of the most impacted areas. Most static, rule based security systems will be replaced by smart, adaptive and risk based ones. We are already seeing a shift towards biometrics (which are powered by Machine Learning) on the authentication side.
Yuanyuan: Customer experience could be significantly improved using AI by analyzing individual level attributes to make traditional service much more tailor-made. Financial advisory is another under developing area, where in future, individuals could expect a machine to suggest best investment portfolios based on their own family balance and consumption behaviors. Automated (deep) underwriting is of special interests in insurance, where we are seeking solutions to build deep Q&A system, optimize business process, deliver predictive models, combine sensor/IoT, and provide better healthcare suggestions. Claim handling (from fraud detection to accurate reserving) is another very important area, we have seen the potential using AI.
What risks and challenges do we face when applying AI to financial services?
Jan: The greatest risks are a lack of accountability and transparency of artificially ‘learned’ actions.
Peter: As Artificial Intelligence gains more traction, the technology leads to new issues related to business, security, and privacy concerns, among others. Another major uncertainty to new technology in finance is that regulation and supervision needs to fully cover all new initiatives.
Hitoshi: In my opinion, one of the biggest challenges around AI, even outside of financial services, is the reasoning and accountability. Usually the deployed model is a black box, and you get a good output from input, but you don't know "why" it generated such output. When it comes to financial services, it is a very important issue to be solved, because regulation requires the decision maker or advisor to be able to be accountable to be in charge. As in driverless cars where people are discussing whether human beings should still be gripping steers to be accountable or not, I believe financial services as a community needs more discussion about what kind of things to be equipped in such AI-backed system.
Alesis: One of the main challenges for start-ups when applying Machine Learning specifically to financial services is educating the customers on the importance of data and access to it. While financial service providers can be very protective of their data (for example for regulatory reasons), Machine Learning algorithms need access to it to learn the patterns. In many scenarios, data anonymization can be used to comply with regulations and to remove the risks of using SaaS platforms, such as AimBrain.
Yuanyuan: The key challenge for applying AI is not only restricted to financial services. People’s panic about work contents change and lack of transferrable skills are still the main concern. According to Oxford and Deloitte’s latest research, about 35% of current jobs have high potential to be replaced by AI in the next 20 years. When it comes to financial services specifically, dated regulations and education systems could be the main burden. However, the insurance industry has been aware of these recent innovations, Casualty Actuarial Society has promoted data science certificate recently as a result.
What advancements in deep learning do you feel will impact the financial sector in the next 5 years?
Jan: Retail finance and global financial information provisioning will change completely in the next 5 years.
Peter: In the financial sector, Artificial Intelligence and Machine Learning is not uncommon when applied to numerical data. While text has been processed in simple ways, I view advanced Natural Language Processing to be the key contributor to the forthcoming wave of disruption. To continue on the above examples, advances in NLP assure that bots will be able to handle simple tasks in customer service and AI systems will on the other hand provide automated information from news, press releases and other textual documents to prices.
Hitoshi: I think lots of trading will be made by AI in the next 5 years, and full-AI or semi-AI trading will compete each other without much of human intervention, as we do as human today. Also there will be more advanced risk factoring and causality analysis of portfolio management by AI, by analyzing different kinds of data such as news, corporate announcement, satellite images, along with the capital market time series data.
Yuanyuan: This is still a bit vague to me for now, as many areas in insurance sector are under innovations in parallel. These all have significant potentials to alter our traditional business modules. One thing we are a bit certain is that in the next 5 years, automated vehicles (a direct application of deep learning) would be launched. By 2020, major manufacturers are expecting to deliver half-automated vehicles; and by 2025, they would launch fully automated vehicles. This could dramatically change the car insurance market and hugely affect those insurers whose main strength is in such domain.
These experts will be speaking at the
Deep Learning in Finance Summit. Discounted passes are available until 29 July! Previous events have sold out, so book early to avoid disappointment. For more information and to register, please visit the website here.Do you have an interesting topic for discussion, or know of someone who does? Let us know! Suggest a speaker using our form here.View all upcoming events here.