Machine learning will transform underwriting, fund management and regulation, and could shake-up advice too
Having heard the phrase “machine learning” a number of times, it had never occurred to me that I had no idea what it meant. I had assumed it was a figure of speech; consultant speak for “a bit better at sums than your ZX Spectrum”. I spent a few hours reading up on it and was surprised by what I learned. I can see how it can have a huge impact on financial services – certainly anywhere a decision tree is required to direct outcomes or where large amounts of data need to be analysed.
Much of what I have seen under the headings of “automated” and “robo” so far has involved digitised decision trees of varying complexity, all designed by humans. The machine is engaged to crunch the numbers, present information and deliver the service but, ultimately, the organ grinder is made of flesh and bone.
Machine learning is a leap forward. The fundamental change is that the computer creates the decision tree based on its analysis of the data presented. Artificial intelligence is a term more often used but that is just the glossy veneer. Machine learning is the engine powering the technology forward.
The significance of a computer being able to analyse and interpret vast quantities of data in seconds, instead of the months and years it might take humans, is obvious in terms of cost and speed. For the healthcare sector, where huge amounts of money are being spent on machine learning, this could also save lives. A simple example I have used recently is the machine-generated analysis on the probability of survival for travellers on the Titanic. The decision tree, with survival rates, shows that women and children were clearly helped on to the rescue boats first but that, while sex was the main determinant, after that age was a greater determinant of survival for men and economic status a greater determinant for women
A simple example I have used recently is the machine generated analysis on the probability of survival for travellers on the Titanic. The decision tree, with survival rates, shows that women and children were clearly helped onto the rescue boats first but that, while sex was the main determinant, after that age was a greater determinant of survival for men and economic status a greater determinant for women.
Of course, a human could arrive at this conclusion but not at the same speed and cost. Additionally, the machine lacks the biases which might make humans miss certain conclusions or promote others.
It is not a panacea. There are multiple data models which can be used: linear regression, decision tree, random forest and neural networks. They vary in complexity. All have strengths and weaknesses. They need checking and pruning. But this work is far easier than starting from scratch.
So, how to apply it to financial services? It is hard, but not impossible, to see an application for advice. It is far easier to see how this would apply to data-driven functions such as underwriting, fund management and regulatory oversight. There is evidence of this already, albeit in its infancy.
Manchester University has worked with both a lender and a healthcare provider to apply some machine learning to their automated underwriting and decision engines. Indeed, underwriting is reckoned to be one of the most likely jobs to be replaced by a machine, according to a 2013 Oxford University study.
Fund management, as large parts of it gravitate to qualitative analysis, could similarly be replaced over time. As with Vanguard’s low-cost active solution, it could be the future for a large part of active fund management and bring costs down to compete with passives. Could niche, data-driven services like Clever Adviser become more sophisticated and commonplace with machine learning?
The FCA will inevitably be interested in machine learning for regulation. It cannot possibly interpret the volumes of data it receives at present, never mind what it will get in future, and draw conclusions with the speed needed to stop a problem in its tracks.
Machine learning, with the caveats I have already mentioned about human oversight, will undoubtedly become pivotal to regulation and government, probably globally and within five years.
Hubris will undoubtedly lead us to believe it will avoid a future market failure. We all know from history that there is always a ghost in the machine.
Phil Young is managing director of Zero Support