Man-Machine learning in finance

Machines are getting smarter; is it time for quants to be worried? John Hull isn't panicking yet, here's why. 

Traditional courses in statistics are concerned with means, standard deviations, probability distributions, significance tests, confidence limits, linear regression, and so on. Those topics are important, but finance professionals now need to cope with what might be termed “the new statistics of machine learning.”   This takes advantage of the fantastic improvements we have seen in computer processing speeds and data storage costs. It allows us to find patterns in large data sets in ways that were not possible before and to develop non-linear models from large data sets for forecasting. One hurdle to understanding machine learning is that there is very little overlap between the terminology used in traditional statistics and that used in machine learning. (The latter talks about features, labels, targets, activation functions, supervised/unsupervised learning, and so on.)

“Will all the interesting jobs in finance be taken over by machines?”

Machine learning is revolutionizing many aspects of finance. Fund managers use it to develop trading strategies. Banks use it for credit decisions and fraud detection. Insurance companies use it to improve their analysis of the risks they are underwriting. Both banks and insurance companies use it to understand their customers better so that they can anticipate their needs. This naturally leads one to ask: “Will all the interesting jobs in finance be taken over by machines?”

It is tempting to suggest that there will be almost no non-data-science jobs in finance in a few years. But this view is too extreme. What we can say is that in order to succeed in finance individuals need to understand enough about machine learning to use it as a tool.

This means that they must (a) understand in a general way how the algorithms work, (b) know how to interpret machine leaning output and be able to make decisions on next steps in an analysis, and (c) communicate effectively with data-science professionals.

This is what we are trying to teach business students on our programs at the Joseph L. Rotman School of Management, University of Toronto. As a step in that direction I have included a chapter on machine learning and blockchain applications in the fifth edition of my book “Risk Management and Financial institutions” which has just be published by Wiley.

There may be some lessons for those of us in finance from the world of chess. Garry Kasparov, world chess champion from 1985 to 2000, is arguably the best chess player of all time. He is also a very intelligent and thoughtful person. He has written several books including “How Chess Imitates Life” and “Deep Thinking.” For much of his chess-playing career, Kasparov was able to beat all chess-playing computer programs. (For example, in 1985 he played against 32 different computer programs simultaneously and won all 32 games.) In 1996, he won a six-game match against IBM’s Deep Blue Program (4-2), but in 1997 he lost a similar match to Deep Blue (2½-3½). Chess-playing programs have continued to improve since 1997 and there are now several computer programs in existence that can consistently beat any human chess player (even the current world champion, Magnus Carlsen).

A man-machine combination can beat a human playing on his/her own or a machine playing on its own.

Kasparov who has a great interest in artificial intelligence has introduced a new game, Advanced Chess (sometimes also called cyborg chess, centaur chess, or Ivanov chess). This is a game between two humans where each human uses a computer chess program to assist in choosing moves. This elevates the quality of the chess that is played. A man-machine combination can beat a human playing on his/her own or a machine playing on its own. Humans are better at strategic analysis. Computers are better at tactical play and brute-force analysis looking many moves ahead. When a computer identifies a number of alternative moves as being equally attractive, the human can use strategic considerations to choose between them. End games however can usually be left almost entirely to the computer.

The success of Advanced Chess suggests that it may not be optimal to completely replace humans by machines in finance. Rather we should retain at least some humans and allow them to work with machine learning algorithms in much the same way that humans work with machines in Advanced Chess. For routine decisions the machine learning algorithm will tend to be best. But when there is a regime change or an unusual situation, the human may well be better at responding.

John Hull will be speaking at QuantMinds International, in Lisbon this May. Wiley develops digital education, learning, assessment and certification to help universities, businesses and individuals bridge between education and employment and achieve their ambitions. They will be exhibiting at QuantMinds International.