How Fortia is using Machine learning to Handle Compliance Overload
The financial world has hugely changed in the last years. Since the last financial crisis, governments and regulatory authorities have imposed stricter rules and guidelines on investment funds.
Just in 2014, the average number of changes in banking and financial regulation reached 155 per day, while putting each of these changes into place takes at least 3 months. There is an obvious gap between the regulation change-pace and the reactivity of the industry. Our goal at Fortia is to address this issue using machine learning and artificial intelligence.
Inside an investment fund prospectus, we can find the main investment rules that the fund must comply with. The translation of this text into implemented ratios, takes a person between 5 to 20 days on average depending on the size of the prospectus. At Fortia, we propose a method for automatically detecting these rules and translating them into executable ratios.
Machine Learning And Rule Detection
Machine learning is the field, within Artificial intelligence, that allows to create, out of examples, generalizable models for decision-making. Let’s understand this through an example: let’s say we have a group of points in a plane in two colors, like in Image 1. Given a new point (in green), we would like to know what is its color. The machine learning answer is to look for the border between the colored points, and then color the point depending on the side of the border that it lays on. This method allows to classify (color) points that have never been seen before.
In the field of Machine Learning, we can find the 'Deep Learning' methods. These algorithms are characterized by being able to select the features they need to better develop a task. These are the methods we use for detecting the investment rules on a prospectus.
Image1. In red we can see the set of points 1, in blue the set of points 2. Given a new point (empty circle in green), does it belong to the red or to the blue set? See the text for more information.
As with the problem of classifying (coloring) points that we presented before, detecting an investment rule can be posed as a classification problem: is this sentence in the prospectus an investment rule or not?
That is to say, for every sentence in the prospectus, we classify it as being a rule or not. Using deep learning techniques, we can identify all rules in the prospectus in less than a second, and translate them into executable ratios in less than 30 seconds, which is a great improvement with respect to the 5 days needed by a human. This process is deployed currently at BNP Security Services worldwide and running live in other depositary banks and asset managers. We greatly believe that it will allow to highly improve the performance of their compliance management processes.