3.1 Introduction
Artificial intelligence (AI), and the machine learning techniques that form the core of AI, are transforming, and will revolutionise, how we approach financial risk management. Everything to do with understanding and controlling risk is up for grabs through the growth of AI-driven solutions: from deciding how much a bank should lend to a customer, to providing warning signals to financial market traders about position risk, to detecting customer and insider fraud, and improving compliance and reducing model risk. In this chapter we detail current machine learning and AI techniques being used and current applications of those techniques. We further envisage the future role for fully AI solutions as the natural next step after the widespread adoption of machine learning in helping the organisation to manage risk.
An example of ZestFinance
serves to illustrate the potential for AI
and machine learning
in risk management
. ZestFinance was founded by a former Chief Information Officer of Google and in 2016 partnered with Baidu, the dominant search engine in China, to improve Baidu’s lending decisions in the Chinese market. Baidu was particularly interested in making small loan
offers to retail customers buying products from their platform. Unlike most developed countries, the risk with lending in the Chinese market is that less than 20% of people have credit profiles or credit
ratings. Lending
to people who have either ‘thin’ credit profiles, or no credit profiles, is inherently risky as there is no history to draw on to check borrower reliability. ZestFinance
(with permission) taps into the huge volume of information on members held by Baidu such as their search or purchase histories to help Baidu decide whether to lend. They use thousands of data points per customer and are still able to make lending decisions on new applications in seconds. A reported trial in 2017 of their system led to a 150% increase in total small-item lending by Baidu with no increase in credit
losses in the space of just two months.
1
The exact nature of how ZestFinance makes these decisions is not disclosed except under the broad umbrella of machine learning and AI, but essentially what they use as a base is a core machine learning set of techniques around clustering and decision trees, and possibly deep learning. Thus a search history indicating accessing gambling websites would cluster a potential borrower into a higher risk group. While on the other hand, a history of online spending that supports the applicant’s reported income in their application to borrow, or maybe even data indicating upward career mobility, might cluster someone into a group of lower risk borrowers. The actual clusters will be significantly more refined than the simple examples given above. This approach crosses the line from machine learning to AI due to the automated nature of the lending decision-making process.
The ZestFinance example illustrates the essence of how AI and machine learning can improve risk management. A standard credit score is largely a linear calculation of a small number (about 50) positive or negative numerical characteristics about a person and thus misses out on a huge amount of additional personal information that can help to either reduce negative risk or accept positive risk. This new information is often atypical and non-numerical; the type of data which machine learning is innately suited to analysing. Our chapter thus expands on this idea of how AI and machine learning can improve risk management—particularly around the techniques being used to make decisions based on such large volumes of atypical data. The next section briefly details, in a non-technical manner, the core machine learning techniques which can be applied to improve risk management. The remainder of the chapter is devoted to the actual application of AI and machine learning to various forms of risk management, finishing with a forward-looking perspective on what is next for the role of AI in risk management and some challenges that need to be addressed.
3.2 Machine Learning and AI Techniques for Risk Management
A first step is defining what we mean by AI
and machine learning
, and this is not necessarily a straightforward distinction. In a glib sense the public relations and fundraising functions of startups tend to use the more attractive AI term when they most often mean machine learning
, but even in research there is a reasonably fluid distinction. AI is most commonly viewed as intelligence demonstrated by machines, with intelligence being defined with reference to what we view intelligence as in humans (Turing 1952, cf. Shieber
2004). As it matters to risk management
we are normally particularly interested in artificial super-intelligence; that is machines that can demonstrate a risk management-specific intelligence higher than human intelligence in this field. To compare the two terms in a more technical manner, we can say machine learning is a core technique of AI
involving learning from data, but that AI often involves additional techniques and requirements. For example, as noted by Bostrom (
2014), a full AI solution would be automated in terms of data identification, data testing, and making decisions based on the data testing. In practice, AI might involve additional techniques in addition to machine learning
, such as including hard-coded and logic rules. Machine learning
on the other hand normally involves manual data identification and testing by the data scientist, and human decisions as to how to apply the outputted information. Given the lack of technological and organisational readiness for pure AI
, and the reality that most claimed AI is in fact machine learning, in this section we outline the core machine learning
techniques applied to risk management
. In the following section and especially the last section, we move our discussion more towards AI as the logical next step to follow from the widespread usage of machine learning techniques.
Machine learning
falls into two broad categories of supervised
and unsupervised machine learning
. In supervised learning you have input data that you wish to test to determine an output. This is similar to how in traditional statistics terms you have a range of independent variables that you test to determine relationship with the dependent variable. In unsupervised
learning, you only have input data and wish to learn more about the structure of the data. Table
3.1 shows the distinction between these two categories as well as the main techniques within each category. A category of note that crosses both supervised
and unsupervised
learning is deep learning
, which is viewed as the closest towards AI
, and to which we turn at the end of this section.
Table 3.1
Categories of machine learning techniques (
Source van Liebergen
2017)
Problem type |
Supervised
|
Regression | • Principal components • Ridge • Partial least squares • LASSO
| Penalised regression: • LASSO • LARS
• Elastic nets
Neural networks and deep leaning |
Classification |
Support vector machines (SVM)
|
Decision trees: • Classification trees • Regression trees • Random forest
SVM
Deep learning
|
Unsupervised
|
Clustering
a
| Clustering methods: K- and X-means, hierarchical principal components analysis
Deep learning
|
Regression machine learning is the closest group of techniques to that usually applied in traditional determination of the causal relationship between variables. In simple terms, we might describe a traditional linear regression equation for a credit lending risk assessment as perhaps the dependent variable being the risk of loan non-repayment, which is then sought to be explained by a range of independent variables that should influence the risk of loan non-repayment. These independent variables might, for example, include financial measures such average non-repayment rates, whether the person is full-time employed, whether they have a good credit history, and whether they own property.
Regression machine learning
differs from traditional regression in that it uses regression techniques that allow for large numbers of variables to be used as independent variables and then automatically discarded if they lack explanatory power. This is a necessary feature due to the large range of data that is available to the data scientist. It also reduces the extent of theorising needed to determine suitable independent variables. Thus, LASSO
regression zero weights independent variables with low explanatory power, while Ridge regression gives lower weights to variables in a model that are highly correlated with other variables in a model. In both cases the outcome is a reduced model that allows the data scientist to move from large numbers of potential explanatory variables to a smaller subset. A LARS
regression works in the opposite direction to a LASSO
and Ridge, by initially zero-weighting all variables and only adding variables that are shown to have explanatory power. Hu et al. (
2015) provide an interesting application of this set of techniques by applying a LASSO
regression technique, among others, to the development of an automated investment trading
advice system based on stock trend analysis and showing this method has improved trading performance compared to traditional methods.
Principal Component Analysis (PCA) and partial-least squares regressions are quite similar in that they both aim to reduce the number of variables by combining variables and extracting common factors. PCA is the more popular of the two as it is widely used in traditional statistics and therefore better understood. A simple example of PCA is that for a set of potential variables to be used to determine credit repayment risk consisting of: (1) owns a house, (2) owns a car, and (3) has significant savings: a common factor might be extracted from these that could be termed ‘asset ownership’. A more applied example is provided by Zhong and Enke (
2017) who use PCA to reduce 60 correlated economic and financial measures to a smaller set of factors and then apply a set of neural network
machine learning techniques to forecast the US S&P 500 index. This latter example is a common use of PCA in machine learning
: that it is used as a first step for dimension reduction and then a more advanced machine learning technique is applied to learn from the PCA components.
The other main category of supervised learning is ‘classification’, with support vector machines (SVM) and decision trees being the most popular techniques within this group. The outcomes and visuals of decision trees are easily understood in practice and therefore amenable to explanation to non-data scientists. A famous application of the decision tree technique is to Titanic survival prospects. Starting with the initial tree trunk statistic that 62% of passengers died and 38% survived the Titanic disaster, we can build a decision tree to classify groups that had a greater or lesser chance of survival. Thus, a first classification of whether a passenger is male or female shows that only 19% of males survived while 74% of females survived. We can branch out the tree even further by, for example, looking at age groups beneath the first male–female division, and this shows that despite only 19% of males surviving, male boys under the age of 6 had a 67% survival chance.
This classification approach of creating sub-groups then helps to understand what characteristics determine outcomes. In practice decision trees
for machine learning have many more challenges than in the Titanic example, particularly around the issue of overfitting to existing data and thus determined sub-groups having poor predictive power with new data and in new situations. SVM
are more complex in formulation than decision trees
but have the same essential end goal of creating groups based on input characteristics to classify and predict outcomes. In the case of SVM
the approach is to map characteristics on a plane and classify groups based on similarity of where they are on this plane. As a financial risk example, Nazemi et al. (
2018) apply a SVM approach to predict financial loss to holders of bonds that have defaulted (normally some proportion of loss is recoverable from firm assets after bond default) and are able to demonstrate the superiority of the approach compared to more conventional methods.
Turning now to unsupervised learning, we first have clustering machine learning techniques which have some similarity with SVM in that they involve mapping characteristics on a plane. The technique differs in that it is not trying to predict outcomes but instead create similar groups. Email ‘spam’ detection, for example, is usually based on a clustering approach—if an email looks like other emails that are deemed spam then it is likely to also be spam. K-means clustering is the most popular approach, although other methods such as X-means and hierarchical clustering are growing in popularity. We’ll focus on K-means clustering to describe the general technique. In K-means a number of cluster groups the data scientist wishes to arrive at is predetermined (although in practice a range of the number of groups is tested), characteristics are mapped on a plane and a dividing line (not necessarily straight) is drawn that best distinguishes between groups. The idea of the iterative process behind the technique is to maximise the difference in means between determined groups.
Deep learning
and neural networks
2 are viewed as being at the forefront of machine learning
techniques and are often classified separately to the machine learning techniques already described. Indeed deep learning
can be both supervised
and unsupervised
forms of learning depending on the purpose to which it is being applied. The intuition behind deep learning
is to more accurately model complex relationships between variables and ultimately to better mimic human decision-making. To that extent these techniques represent the closest to actual AI
techniques, albeit still missing some of the data identification and automation features necessary for true AI. Heaton et al. (
2017) provide a detailed analysis of the recent application of deep learning to finance
, as well as highlighting the potential of the approach.
A key feature of deep learning
is the addition of ‘hidden layers’ after the input data stage that allow multiple and combined influences between input variables to be determined by the modelling (Goodfellow et al.
2016). As the input data progresses through the hidden layers variables are combined and recombined into newer factors weighted on influence from the prior layer. Thus, for example, a simple economic forecasting model fed with the input variables of GDP growth, unemployment rate, exchange rates, government budget deficit, might then recombine in hidden nodes into new factors. A random example at the first layer of one such recombination (performed by the deep learning
model, not by the researcher) might be 0.2*GDP growth + 0.4*exchange rates. As each hidden layer progresses these recombinations could become more and more abstract if that helps with the task set the model.
The addition of hidden layers between input and output is what creates the perceived problem with deep learning—that the process is a ‘black box’—in that it is not always clear how inputs have been recombined to create a predicted output. This has obvious implications for use in risk management, where the very presence of a black box at the centre of decision-making can be its own source of risk in a firm.
3.4 The Challenges and Future of Machine Learning and AI for Risk Management
There are some significant practical issues that need to be addressed before AI and machine learning techniques for risk management can claim its full potential. The most important of these is the availability of suitable data. Although machine learning packages for Python and R can easily read all types of data from Excel to SQL and can perform natural language processing and process images, the speed with which machine learning solutions have been proposed has not kept pace with firms’ abilities to suitably organise the internal data they have access to. Data is often held in separate silos across departments, perhaps on different systems, and perhaps with internal political and regulatory issues restricting the sharing of data. Important data might not even be recorded as data but rather kept as informal knowledge of the firm.
Another issue is the availability of skilled staff to implement these new techniques. A survey of the top 1000 firms in the United States on AI implementation in their firms found that their biggest concern in the implementation of AI was the readiness and ability of staff to understand and work with these new solutions (Wilson et al.
2017). Training a skilled cohort of staff is something that will take time, although Goldman Sachs, among other firms, has recently attempted to bypass this problem by developing a campus with space for 7000 workers in India where prevalence of these skills is more present.
There are also practical issues over how accurate machine learning
solutions actually are. The range of testing approaches available within machine learning is growing rapidly, and that is a good thing, but it is also driven by the evident limitations of the previous methods and the need to overcome those limitations. This suggests that firms cannot simply ‘apply’ a machine learning risk management
solution, but rather that it is a continuous process requiring a constant evaluation of whether their particular machine learning solution is currently considered best practice. When it comes to AI, where there is some or full automation of process from data gathering to decision-making, the need for human oversight will become even more pressing. The case of Knight Capital in 2012 serves to illustrate this risk, with their stock trading
automation through algorithms resulting in a loss of $440 million in the space of just 45 minutes.
3 As AI starts to automate lending and credit risk
decisions it will be imperative to ensure that such risks can be controlled before handing over control.
This last point serves to introduce the final major issue around transparency and ethics which AI-driven solutions need to further address. Transparency is especially an issue for the increasingly popular deep learning method of machine learning as the models work on hidden layers between the input data and the output decision. A black box system of this type of not conducive to effective risk oversight and can cause regulatory compliance issues especially around demonstrating model validity. A more hypothetical issue related to this is that models used by different firms might converge on similar optimums for trading causing systematic risk as well as loss to the firm. There are also broader ethical issues. For example, discrimination against race in lending decisions is widely legislated against, as are increasingly discrimination based on gender and sexuality. Normally these restrictions are incorporated as hard-coded rules in AI and machine learning techniques concerning credit risk and lending decisions. However as deep learning becomes more popular it becomes much harder to police that the model is not inadvertently making decisions based on indirect proxies for these red lines. This is especially the case as more and more atypical data is incorporated in risk management. These ethical aspects are only likely to increase in significance over time.
Leaving these important issues aside for now, it is worth considering the future role for AI and machine learning in risk management from a more positive angle. One obvious conclusion is that the time-consuming and costly nature of risk management will diminish significantly. For example, BBVA, the second largest bank in Spain, has 8000 of their 137,000 staff working in compliance. They are investing heavily in AI solutions to reduce this compliance cost base. The ability of AI and machine learning to automate repetitive tasks and to organise, retrieve, and cluster non-conventional data such as documents is naturally going to confer cost benefits on firms that move more into this area.
AI will also increasingly deliver accurate real-time information on all types of risks being taken by the firm. As data organisation becomes more orientated towards use by AI, real-time advice will become a pervasive presence. The following step from real-time knowledge of risks being taken is pre-emptive notice of risks. To some extent this is the holy grail of an AI-driven risk management system—to be able to accurately know in advance firm risks, be they market, operational, or credit risk. The techniques of machine learning offer this ability in the way that traditional statistical techniques could never hope to. Thinking even further ahead there is no technological impediment to a truly AI risk management system that will automatically intervene to prevent unwarranted risks, to immediately unwind dangerous exposures, and to dynamically adjust the risk appetite of the firm based on the system’s estimate of the broader risk environment. Although that then will present its own risks which will in turn have to be managed, keeping risk management professionals happily employed (albeit in a fast-changing environment) for the foreseeable future.
We thus conclude on a positive note (albeit with some issues to consider), about how machine learning and AI is transforming the way we carry out risk management. The issue for the established risk management functions in organisations to now consider is if they wish to avail of these opportunities, or if instead it will fall to current and new FinTech firms to seize this space.