How Machine Learning Could Shape the Future of Finance
If you’ve ever accidentally wasted an hour of your time scrolling through your news feed on Facebook, chances are you’ve just experienced the power of machine learning (Forbes, 2017). Whether it be facial-recognition algorithms used on your new smartphone (Apple, 2017) or fraud-prevention algorithms used by major credit card companies (Lusis Payments, 2017), machine learning is proving itself to be a technology of the future. In addition to its use in these applications, it has the ability to reshape the financial world forever.
Modern machine-learning algorithms use a supervised learning model (see Figure 1). This model feeds training data, which is split into its identifying features, into the machine-learning algorithm and creates a series of triggers and outputs, called a neural network. The result is then manually paired with a label. Once the algorithm is trained with an adequate amount of data, it can be used to take feature vectors from new data and predict which label it falls under (AllProgrammingTutorials.com, 2015). Although this approach is extremely effective for picking out your face in a picture, or figuring out which song is playing in the background, it has yet to master more complex financial applications.
In the field of stock-market forecasting, the aim is for machine-learning algorithms to reach a stage that enables judgements to be made on stock values (i.e., in terms of direction and magnitude) with a success rate that significantly exceeds their running cost. This may sound like an unattainable and unachievable goal to many. However, with the prospect of a large financial reward, huge amounts of money are being poured into such research by many financial services companies, such as JP Morgan (Butcher, 2017) and Morgan Stanley (Harvard Business Review, 2017).
This doesn’t mean that you can simply create an algorithm and become rich within days. In an extensive investigation, André Anderson (Anderson, 2012) came to the conclusion that “No trading system was able to outperform the [average trader] when using transaction costs.” Further, Dr Yoshua Bengio, Head of the Montreal Institute for Machine Learning Algorithms, said, “Market inefficiencies tend to be localised in time and ‘space’ (particular markets, with a limited potential volume of profits). So it may well be that some firms have used and are using machine learning, but it’s not like [hitting the jackpot], rather like patiently pulling profits here and there, each time with a different specifically tuned model.” (Bengio, 2017) Dr Bengio, the author of the piece, goes on to say that companies are currently using a significant amount of human judgment to assess which trades to make.
Unfortunately, this is the underlying theme with current stock-prediction applications using machine learning: the algorithms just aren’t good enough to exceed the performance of an average speculative trader.
Past systems have used data from news articles to assess specific companies’ success. However, these approaches have failed to take into account the virtually random speculative investment that plays an important role in driving stock values. Speculative investment is the process of buying and selling stocks on a short-term basis with little to no evidence for an increase in value over that period (Hayes, 2017). Although counter-intuitive, this process—when performed by a significant number of people—can seriously affect a stock valuation (Roosevelt Institute, 2011). For this reason, incorporating some sort of detection of popular opinion on companies being traded is vital to accurately predicting their future values.
New, cutting-edge research performed by the Indian Institute of Technology (IIT) uses sentiment analysis (Pagolu, et al., 2016), a process of analysing language in sentences to assess opinion on specific matters. Complex sentiment analysis engines must be able to determine that, for example, “that horror movie we watched was so scary” is a positive subjective comment. This is something that is easy for humans to understand but much more challenging for computers, due to the use of contextual language (‘scary’ being positive in this case). The research uses sentiment analysis on Twitter to assess public opinion of particular companies which can later be fed into the machine-learning algorithm. Indeed, this technique is currently being employed by many research departments, including at Stanford (Mittal & Goel, 2017) and Cornell (Pang & Lee, 2017).
This, in conjunction with a computers’ ability to trawl through billions of words with ease, will enable machine-learning algorithms to detect a much larger spectrum of information—ranging from hints of speculative trading on social networks, to discussions on trading forums—in a way that humans never could. If successful, this research will mark a new era for the world of finance.