Chakraborty and Joseph’s paper provides a comprehensive foundation in the main concepts of Machine Learning (ML) through the lense of an economist. The publication does require a reasonable grasp of maths, or at least to be fluent in the greek symbols and nomenclature of the field, along with the equally jargon-rich world of finance. For this reason it is assumed that the intended audience is either someone from the world of ML wishing to understand another application of this increasingly popular field or a highly numerate economist / banker, wishing to sharpen up their toolset to be more contemporaneous.
The length of the publication at 85 pages including references, requires a reasonable commitment from the reader, but does cover a lot of ground in the theory it covers as well as three very interesting and diverse case studies: banking supervision, UK CPI inflation forecasting and an analysis of unicorns in financial technology.
In their preliminary overview and consequent case studies, the authors outline all the significant steps an experienced data scientist might go through when conducting research other than one common process: dimensionality reduction. In plain speak, this is a technique to compress the inputs being used in an analysis into their most informative components with the ambition of producing more parsimonious models. Simpler is always better, and Machine Learning provides an elegant way of trading off complexity against performance. The authors do outline one method from this field in the context of visualisation, but it is surprising they don’t employ it when actually making predictions. This is particularly true in the CPI case study where the model inputs are highly co-linear (i.e. compremissible).
One surprising omission in their work, again in the case study of forecasting CPI, is that of a very popular class of models, imposingly called Gaussian Processes (GPs). These are a flexible methodology for time series forecasting, particular when the variable being forecasted has intrinsic structure. CPI is well known to revert to mean: if it goes up one quarter, it is likely to come down the next (everything else being equal), so one might think GPs would be very appropriate here. Furthermore, these models are particularly good at articulating the confidence one has in predictions when they are made - analogous to the fan charts that litter the BoE’s quarterly predictions. In many ways GPs are the embodiment of forward guidance, a concept very popular with the BoE, and it is strange they are missing in this work.
Another notable exception to their run through of commonly used algorithms / concepts at any depth is that of Reinforcement Learning. This is the research field devoted to models which take constant feedback from their environment in order to ascertain how to best exploit it and is very popular in financial services, for example in high frequency trading (always an area to be on the cutting edge of exploitative methods!).
These very minor gripes aside, the pragmatic explanations of methods in the context of the three use cases will appeal to anyone actually wishing to ascertain the utility of ML as a toolset. The authors, like any sensible practitioner, are not wedded to any particular algorithm or, even worse, optimisation method, and just explore some common and well studied methods. Other organisations, less grounded in centuries of history, might have jumped more full-heartedly onto the Deep Learning band-wagon, ignoring the impoverished ability to interpret these structures’ inner workings and their relatively reduced utility in this domain.
I believe that experts in a field should be able to have a fractal view of their subject: understanding at a very high level where it might fit into the bigger picture (e.g. central bank policy) as well as the minutiae of how one goes about getting one’s hands dirty implementing aspects of the area (for example tweaking optimization parameters in some obscure algorithm). The authors of the paper seem to operate at all scales without the irritating need to obfuscate their thinking with exclusive terminology, a method others insecure in their knowledge often seem to resort to.
In conclusion, this paper is not in any way a comprehensive survey of Machine Learning in Central Banking, or even in economics. It is more of an introduction to this tool-set in the context of the central banking domain. There don’t appear to be many surveys of ML in a supervisory role and I suspect if there were, they would be pretty brief. Many practitioners, particularly those in supervisory / regulatory roles, are still trying to get more comfortable with ML, and institutions like the BoE / FCA are not motivated to be the early adopters found in the private sector. Furthermore, banking regulation / supervision itself is a rapidly moving target, so one would expect its overlap with novel theories to remain fairly minimal as a consequence.
The important point to note here being that the poverty of publications in this area is not because of some intrinsic limitations of ML’s applicability to the domain, it’s just that it’s early days.
This Bank of England paper is a promise of more to come.