AI in finance

If ‘data science’ is ‘statistics re-invented’, then arguably artificial intelligence must go further than that to differentiate itself. IFT’s Director of Research, Professor Sir Alan Wilson, reflects on the overlap between data science and AI, and what more it can offer. Prof Wilson reads this as the ability to offer insights from the analysis of large data sources or ‘big data’ that goes beyond statistics.

An introductory step is to sketch what AI can do and what it can’t do. Through associated methods in mathematics, statistics and computer science, combined in new ways, AI can see, hear, read, translate and write. These are remarkable achievements and can themselves transform many business processes. Many of these are being applied by banks, for example, in serving their customers. What AI can’t do is ‘think’. This led Michael Jordan to argue, in his blog post Artificial intelligence: the revolution hasn’t happened yet, that we should neglect the earlier ambitions to create a human-imitative ‘thinking’ AI, and convert AI to IA: intelligence augmentation.

If we continue to follow Jordan, we can note that much of AI is ‘machine learning’. This allows us to identify hitherto undetected patterns or structures from data, thus providing insights and augmented intelligence. This can be dramatic: DeepMind’s AlphaGo systems which recently showed how to reveal the structures of protein folding is a remarkable case in point. The simplest kind of structures are clusters, for example of population types, which provide the basis of new marketing algorithms. Another kind of product is ‘anomaly detection’ which has obvious applications in finance in relation to money laundering and fraud.

There are three kinds of machine learning: unsupervised, supervised and reinforcement learning. Unsupervised throws machine-learning algorithms at large databases and invites the delivery of possible clusters; supervised learning starts with a cluster-labelled data set and seeks to position elements of new data into these clusters; reinforcement learning combines new data with old in such a way that the cluster definitions can be improved – and in this sense, the system ‘learns’ as data is added.

In many applications, the assignment of people to clusters is the basis for decisions to be made – say, about loans or an insurance premium. The algorithm, based on a trained data set, takes a new person, assigns them to a cluster, and offers a rules-based ‘decision’.  This immediately exposes the risks of applying computer machine-learning based decisions in finance (and elsewhere): the training data may lead to biases for example. There are challenges for both companies and regulators. In the case of one insurance company, if the ‘the computer says yes’, the decision is accepted; if ‘the computer says no’, the case has to be referred to a human agent. There are similar examples related to money laundering: if a possible instance is detected, it usually has to be referred to human agency because of the number of false positives generated by the algorithm. There are clearly research challenges in this territory to make the algorithms more reliable!

There are different kinds of application in asset management and portfolio construction. This illustrates another fundamental challenge in relation to applications of AI in finance: handling uncertainty. If the output of a machine-learning algorithm using stock market data is the prediction of future prices, and, say, the identification of arbitrage opportunities, the offered results will be augmented intelligence that makes the uncertainties explicit. Very complex mathematics and statistics underpin these fintech applications.

Applying AI to finance

Even this very brief sketch indicates how AI is already widely deployed in finance and this will expand in the future. Can we envisage ‘revolutionary’ steps? To further fix ideas, it is useful to look at an example from health as the basis for articulating some possible research projects in finance. Consider the following scenario: the data from a large number of patients flow into a system on a real-time basis. The objective is to create a machine-learning-based system to offer augmented intelligence to clinicians, when presented with a new patient, for diagnosis and the creation of treatment plans. There are four phases:

1. the data – health and social data combined – is wrangled into a state where it can be entered into a machine learning algorithm;
2. the algorithm is trained on the data and output is generated for the new patient: probabilities of diagnosis and prognosis, and intelligence on the best treatment plans based on past training;
3. the clinician reviews the outputs and decides on a treatment plan;
4. over time, the impact of that treatment plan is evaluated and fed back into the data base and the algorithm is continually retrained.

This system can thus be thought of as a learning machine, working towards optimum recommendations of treatment plans for clusters of patients with particular diagnoses. Such a ‘machine’ can be the basis of personalised medicine and offers the possibilities of a health care system which is both cheaper and better. The mathematics and statistics which underpin the algorithm’s ‘model’ are immensely complicated – needing to take on board the patterns of co-morbidities in a patient’s medical history. Embryonic versions of such systems have been tested in areas such as cystic fibrosis, cardiology and dementia. Implementation on a large scale remains an R & D challenge.

Clearly such a system could be developed for, say, a bank or an insurance company: assemble data on clients, process through a machine-learning algorithm, create augmented intelligence for marketing and customer service purposes; evaluate outcomes and enter feedback. The learning machine system could be used in the first instance to put some purposeful order into ‘big’ data systems. And of course, there is a potential link between the health learning machine and a financial one: insurance companies and banks would love to be able to link the two and, for example, to have access to health prognostic data for their clients! However, this raises huge issues of privacy and ownership of data, but there must be a likelihood of this being on the table for debate, albeit still in the more distant future.