Machine Learning And The Future of Finance
By Elliot Noma PhD, Managing Director, Garrett Asset Management
Artificial intelligence has conquered games and image recognition, but will it master investing? The short answer is yes, but how soon and how complete?
Machine learning methods have had impressive recent successes. These include defeating humans at chess, Jeopardy, poker and Go, as well as providing superior image and speech recognition. Developers strive to create tools that automate decision making and that can mimic or exceed human performance for specific tasks.
The range of tasks and the variety of methods influence current successes and the way forward. This means that there can be large differences in the short-term outlook of machine learning methods in finance, with some areas quickly embracing artificial techniques (AI) while other areas require the development of new methods.
AI and machine learning are often used interchangeably since they convey a general idea that software can make intelligent decisions, and intelligence implies the ability to learn over time. However, the ability to learn new concepts is different from the appearance of intelligent behaviour. Some methods, such as unsupervised learning, need little guidance, while others such as support vector machines require extensive training. Linear regression models have a great deal of statistical theory backing their application, but for neural nets the theory is still being developed. Moreover, models of learning vary based on the criteria used to evaluate success.
AI is being applied to finance in several ways. These include the automation of many tasks, handling basic functions and the ability to change over time as the software adapts to new market conditions.
1. Automation: When incorporated into data feeds, models extract information from inputs, classify them into useful categories and initiate actions autonomously or through human intervention. These actions are consistent and can be improved over time. Automation allows quicker evaluation of inputs for an investor trying to determine the utility of a new dataset. It enables a detailed look at large amounts of data produced by sensors in the Internet of Things. Trading models can be developed quickly as older models become obsolete.
2. Handle the mundane: Software excels at monitoring every-day occurrences. It can easily monitor the performance of a physical device to determine that the device is functioning correctly.
Tabulating the performance of even the largest pool of loans is done easily and accurately. Over time, machine learning tools can become increasingly sensitive to deviations from what is expected and are therefore highly useful tools in fraud detection. Different detectors can also be linked together to identify deviant behaviour.
3. Adapt over time: Machine learning techniques require training data so they can predict the best actions in normal situations. As data sets expand, the software can be adapted to increase the number of features that it can monitor. This makes decision making more nuanced. In addition, as new categories are added in supervised learning, the granularity of the decision process improves with more data.
However, machine learning methods need customisation for different domains. For instance, neural nets are usually trained on images of a standard size to better fit the internal network. A neural net analysing text may be better structured using a different internal configuration. In finance, these challenges may extend beyond just the geometry of connections making up the network.
1. Time to learn: Machine learning techniques require training data so they can excel at predicting the best actions in normal situations. They are weakest when classifying unusual situations where there are few training cases and exceptional drivers dominate. Also, some data may arrive at specified intervals such as the announcements of central banks or quarterly corporate financial reports. In contrast, other fields can accelerate their data collection by increasing their web traffic of recruiting evaluators. Using a variety of models to analyse data from multiple sources may help here, as techniques such as boosting maximize the contribution of each model and data set.
Furthermore, learning can take place at several levels and for different purposes. For instance, you may talk to a chatbot which appears to exhibit intelligence. It answers your questions and may have some learning as you continue to converse with it – but only so far. Another example is an algorithm which ostensibly learns to detect images of cats, but might be actually creating a rule on the brightness of the background.
These structures point to differences in our expectations of how AI learns and what it retains. This makes the individual methods impressive in many contexts, but lacking in aspects of intelligence that we take for granted in our interactions with other people. These include a memory of context and our ability to notice deviations from our expected context. Humans also assume that certain information will be learned even if we are not aware we are learning it. We also expect that when asked, we can give some justification for our decisions and perceptions.
2. Continuous stream of input data: Another challenge is modelling a continuous flow of information whose time boundaries may be unknown or indeterminate. This is unlike image processing which analyses individual images in isolation or game-playing programs which use well-defined conclusions that can be used unambiguously to define victory or defeat. In contrast, the time horizon of financial investments is usually not fixed as personal and business circumstances change over time as does the evaluation of investment success.
3. Lack of stationarity: Images of cats do not change over time as our image processing capabilities improve. However, trading strategies and markets as a whole adapt depending on external events and actions by traders and money managers. In some markets, new trading algorithms may have a very short useful life. Often these changes are due to events outside the world of finance, so they are beyond the horizon of most models.
4. Crowded trades: The use of common data sets and common machine learning methodologies can lead to crowded trades which limit profitability for some strategies. It also moves overall market risk from individual portfolios to the market as a whole, similar to how the convergence of risk management methods increased systemic risk. However, there is a wide range of methodologies for machine learning and the number of datasets is expanding as new tools and measures are created, which mitigates concentration risk.
5. Lack of transparency: Of greater concern is the acceleration of the investment process and the decreased insight into why decisions are made. Combined with the lack of understanding of how non-normal markets are handled by software this should raise concerns about how the system copes with periods of high volatility or events such as the “flash crash” of May 2010. Software and hardware are prone to bugs, and humans need to keep a tight control of the behaviour of machines as they take over more and more tasks.
Among the various applications across finance – credit, operations, trading cycle – each has different characteristics that play to the current strengths and weaknesses of machine learning. One cannot generalise within such broad categories, but certain specific applications are most compelling.
For instance, many operational and trade execution functions can take advantage of automation to recognise the normal. Routine execution and trading functions can be best conducted using algorithms in normal markets. Some brokerage and banking functions can be streamlined as investment advisers are cued to sales opportunities by software that considers customers’ past trades and investment preferences. Credit verification already uses a high degree of automation and could be made even more accurate by considering more complex patterns of the borrower and adjusting the weights to these factors based on changes seen in customer behaviour.
All these software tools are used within the world of human activities that have consequences for individuals and whose rules are set by humans. For the immediate future, humans will need to give oversight to computer activities and continue to handle the exceptional situations. This is especially true since the reasons a program makes a decision are very different from the heuristics used by humans.
Also, methods for computers to explain the logic behind their decisions to humans are still in their infancy. This lack of communication is especially important when determining the blind spots of an algorithm. Software cannot dictate an investor’s preference for risk or intuitions about specific opportunities. Much remains outside the view of models such as the activities of governments and central banks to affect the markets, especially during exceptional situations.
Just as other machines in other areas of activity have not fully replaced people, machine learning algorithms are unlikely to replace people altogether. Software will, however, change the number of people doing specific tasks and the skillsets of those remaining. The one certainty is that machines will further disrupt the financial industry.
We’d love to hear your feedback on this article. Please click here