Have you ever wondered how your e-mail client knows what e-mail is spam and what isn’t? Or how robots have learned to play football? How about AlphaGo, which was able to beat the world’s most talented player in a game which had been notoriously difficult for computers to compete in? How are credit card companies able to efficiently analyse vast quantities of data to accurately identify the small fraction of fraudulent transactions? As an astute reader, you can probably guess the answer: by using machine learning techniques.
Machine learning is a way of solving problems where we train a computer to find solutions based on large amounts of data, instead of programming the solutions directly. Let’s use an example: if we were to program an e-mail spam filter directly, we might command it to always mark e-mails with spoofed sender addresses and specific keywords as spam and leave other e-mails alone. If we were to apply machine learning to create an e-mail spam filter, we’d go about it differently: we would provide a suitable machine learning algorithm with thousands of example e-mails, some of which had already been classified as spam. The algorithm would then try to find patterns in the data: what kind of characteristics do the spam e-mails typically have? How do they differ from the valid e-mails? Given enough data, a suitable algorithm, suitable parameters and a bit of time, the training will yield an algorithm which is ready to start marking spam e-mails for us more accurately than a spam filter built in a more traditional way with predefined heuristics.
How can we apply machine learning techniques in the FA platform in a way which would help you, our dear clients (or clients-to-be)? That’s what we’re currently investigating and working on. Initially, we’ll focus on two specific areas: compliance and CRM. We may explore other areas in the future, but we’ll start with these two.
Machine learning is commonly used to detect credit card fraud. Although that’s not directly relevant to our clients, many of you face regulatory requirements of a rather similar nature: the obligation to conduct anti-money laundering (AML) monitoring. In practice, AML monitoring requires flagging and reporting things such as significant flows of funds, and mismatches between an individual’s financial activity and their background information. Let’s say an individual makes a sizeable deposit. Depending on the individual, flagging the deposit may or may not make sense. If the individual reported a modest salary and indicated that they are investing funds solely from this modest salary, we need to flag the transaction for closer scrutiny. If we have already established that the individual is a wealthy businessman who has easy access to such funds, we may not need to investigate any further. Distinctions like these, and spotting irregularities relative to how different kinds of clients normally behave, is an area where machine learning holds lots of promise. This is an area that we are currently investigating, with our goal being to be able to flag irregularities in financial activity more accurately than simpler heuristic methods would be able to.
We believe machine learning techniques could also yield valuable insights in the area of CRM. We could apply it to questions such as:
• Which of our clients are we at risk of losing?
• Where are the best opportunities for additional sales within our existing client base?
• We have obtained a large list of leads; which ones should we focus on?
• We just launched a new investment product; which of our clients might be interested in this?
The FA platform is unique in the fact that it includes many different types of data in a single platform. That makes it easy for us to answer questions such as the ones mentioned above from many different perspectives: perhaps we are at risk of losing clients whose portfolio has performed poorly, who have recently rated themselves as risk-averse, and who have stopped investing additional funds into their portfolio. Perhaps the client whose investments have done well lately and has indicated that they have a considerable amount of assets invested through other channels is an excellent opportunity for additional sales. Perhaps several of our leads match the profile of a wealthier investor who, if we were to win him as a client, would generate an exceptional amount of value for us in the future.
As a user of FA, you most likely already have an abundance of enormously valuable data in your system. You are also almost certainly underutilizing this data. The goal of our ongoing work in the machine learning realm is to help you unlock the enormous potential of the data that you already have. We expect our work to gradually come to fruition during Q2-Q3 of 2019 – expect another blog post around that time!