With Gary Kazantsev, Head of Machine Learning, Gideon Mann, Head of Data Science, and Bruno Dupire, Head of Quantitative Research, Bloomberg
Machine learning cannot do everything people can do, but the technology is finding more widespread use in finance.
What is the biggest misconception about machine learning in finance?
Gary Kazantsev: That it is some sort of a magic wand that will solve hard problems in contravention of truths known from basic statistics. No amount of machine learning will help if the problem you are trying to solve is ill posed, or you don’t have a sufficient amount of data, or if you aren’t careful about issues like non-stationarity and bias.
Gideon Mann: One major misconception is that machine learning can do things that people cannot do — that it can magically accomplish things that tax human ability. Typically, the biggest impacts of machine learning come by automating simple and straight-forward human decisions, but doing it on a cost basis that makes various processing economical. This, in turn, leads to the appearance of magic.
How advanced is machine learning in finance today?
GK: It depends. The range of problems being attacked and the methods used is now vast and rapidly expanding. We are familiar with organizations which do end-to-end strategy development (from portfolio selection to execution) as an ensemble machine learning problem. There are also plenty of firms who are now only starting to investigate this field.
The level of acceptance of new technology in financial institutions varies depending on their acceptable risk profile, specific requirements for interpretability and transparency of models, and even geographical region (which influences the available pool of talent familiar withcutting edge research). This applies to machine learning even more so than many other technologies.
Bruno Dupire: Machine learning is still in its early stage, but catching up very quickly. Quantitative finance is a natural field for it as learning to establish links between input data and returns is very valuable. Data, both structured (security price time series, fundamentals) as well as unstructured (text from news/tweets/call transcripts, net searches, satellite images) are systemically exploited and the array of methods is ceaselessly expanding.
Random forests, support vector machines, knowledge graphs, recurrent nets, LSTM (long short-term memory), convolution nets, GAN (generative adversarial networks). It has changed a lot since I initially used neural nets to forecast financial time series in 1987.
How are sophisticated clients using machine learning in their workflow,and how is it impacting investment strategies?
GK: We have seen everything from counterparty risk analysis to optimal execution, and from predicting bankruptcy risk to forecasting returns, earnings or unemployment statistics. It’s also being used in portfolio construction, sentiment analysis of financial news and so on. Machine learning is becoming an integral part of the toolbox used in creation of systematic strategies.
What is driving investment and attention in machine learning in the financial industry?
GM: Machine learning has had an enormous effect on other industries and has driven significant growth. Think Google, Amazon, Facebook. There are also an increasing number of financial firms that have been able to harness machine learning to drive value. Finally, the pressure to trim costs has focused firms inward to see if they can do more with less, and enhancing employee productivity through augmentative technology has become more appealing.
What new Bloomberg machine learning application or tool are you most proud of and why?
BD: We are building a machine learning prototyping suite that enables the user to access scikit-learn, TensorFlow and our own functions, in a very user-friendly interactive environment. It offers multiple ways to visualize the data, the progress of the learning and how the algorithm operates.
GM: We have made significant investments in our neural network infrastructure, and because of our efforts have seen numerous examples of deployed neural network models. From these, the effort in understanding tables has particularly made me proud as it demonstrates the power of these new technologies on a thorny old issue.
GK: I am particularly proud of the work we have done on question answering. We have been able to make an impact on the way clients use the Bloomberg terminal despite this being a very challenging open problem.
Can you give one prediction for the future?
GM: I think the future is likely to be increasingly characterized by fairly stable periods interrupted by very rapid changes as the speed at which information and technology gets disseminated increases.
GK: Sea levels will rise, markets will fluctuate, and deep learning will not give us true human-level artificial intelligence.
GD: I think the community will soon realize that neural nets, however deep they are, cannot solve every problem. We are likely to observe a merging of neural nets and logic-based systems, especially when data is scarce. For advanced tasks, it is not enough to let data drive the learning process, one also needs to inject expert knowledge, leading to hybrid systems.