The volume of data stored and analysed by financial institutions conducting trading and investment activities has increased exponentially in recent years. Firms no longer have the human, time or financial resources to properly analyse even a fraction of this data while remaining competitive. As in the past, technology may provide a solution. Many computer scientists believe artificial intelligence(AI) will completely replace traders and investors, as has already happened to a large extent in areas such as High Frequency Trading. Presently, one developing aspect of AI, machine learning (ML), is drawing increasing attention from those providing solutions that rely on huge quantities of data. This is especially relevant within Capital Markets, given the demand for techniques that enhance the analysis required to make financial or investment decisions. They allow financial institutions to perform automated analysis on big data in less time, with smaller headcount and greater accuracy, while also saving money and increasing profitability in the long term.

What is machine learning?

Machine learning is a branch of computer science that automates both the analytical techniques of computers and the software development of humans. Machine learning enables computers to “learn from experience” by finding patterns in information and to adapt their own programming to achieve outcomes. This process is similar to data mining, since both find hidden recurring data relationships within large datasets. However, unlike ML, data mining techniques require more human interpretation and interaction with the software, allowing wider scope for human error. Machine learning applications use stored information to automatically improve process algorithms. The most common techniques used in quantitative finance applications are: (a) supervised learning and (b) unsupervised learning.

Supervised learning

The goal of supervised machine learning is to create an adaptive algorithm that makes predictions based on categories, or instructions, given by the programmer. The machine ‘learns’ from observations, exactly as a human being would do, improving continuously its performances (by absorbing examples of correct behaviour). Input data, along with known responses, are fed into a computer that identifies common patterns in information. The supervisor will then tell the algorithm which ‘successful’ patterns it should be looking for, allowing it to improve its future performance on other data sets. For example, suppose an investment bank wants to estimate whether the value of particular shares will rise or fall within the coming two months. Information on prior market behaviour is fed into the machine as training data. The system will then combine all the given data into an algorithm that estimates the future share price within the given timeframe.

The level of accuracy of ML techniques in stock market is growing widely but some limitations remain, particularly concerning the accuracy of training data – bad historical data will not help the software make good future predictions. Subsequently, stochastic models are often required. The most popular method of supervised machine learning is inspired by the human brain model and constructed similarly. Neurons exchange information through connections, and these connections become stronger based on whether or not those connections result in a successful outcome. As repeated sets of data are input the strongest connections will have a larger influence on future data sets, much in the same way as a brain strengthens the most oft-used connections.

Unsupervised learning

Unsupervised learning, on the other hand, does not use training data to make predictions. The objective is for the machine to learn how to perform a task without the user providing any initial examples or historical data. This technique is mainly used to find patterns and, as there are no desired outputs, these algorithms can only identify similarities in data. As such, unsupervised learning techniques cannot be used to make predictions, as there are no instructions as to what it is looking for, but it may be used to classify data. With reference to the Capital Markets, unsupervised techniques can be used to find groups of stocks with the same behaviour but it is unable to make predictions on their future performance. The most widely used technique is called clustering. The goal of cluster analysis is to identify the intrinsic structure of data and group them ways that objects in the same cluster have more similar structures than those in other groups.

The debate whether AI will eventually replace humans in the making of trading and investment decisions remains open. Mathematical trading models are already being created and used by humans to better enable them to respond to changes in market conditions. As advanced machine learning techniques progress, computers will be able to teach themselves how to perform this function. These “trading robots” will be able to process huge quantities of data identify complex market patterns and react to them far more efficiently and quickly than any human could. The high initial cost of implementing ML is justified by the long-term reduction in resourcing costs, efficiencies in performing processes and reduction in errors caused by human input. Machine learning development is derived from the idea that “intelligence can be so precisely described that a machine can be made to simulate it”. In the digital age, the financial services industry is increasingly replacing conventional analytic techniques with new forms of machine learning to stay competitive and even become revolutionary. No firm can afford to lag behind the curve.

http://www.onthemedia.org/story/engineering-intelligence/transcript/