In the previous article, we considered the client on-boarding process; in this second part, we will focus on the sales/advisory process. We will explain which machine learning (ML) models can be applied to which parts of the process. This is not an exhaustive view of the possible options and is meant as an opening to further discussion.
Sales / Advisory Process
Historically, the advisory process has been based in human interaction between the relationship manager/investment advisor and his clients because of the complexity of products and services that can be offered. Furthermore, the regulatory constraints around product suitability and tax implications remain a challenge requiring special skills and knowledge. Big data and machine learning technologies can support the process of collecting the required data and supporting the investment decision process based on defined rules and on historical data.
The following are typical processes in wealth management that could be impacted by applying machine learning techniques:
Investment Advice: advising clients about investment opportunities where the client acts as the decision-maker based on the relationship manager/investment advisor input (also called self-directed investment).
Discretionary Portfolio Management: management of the portfolio is delegated to the portfolio manager who acts on the client’s behalf to maximize the client’s portfolio performance based on an upfront, agreed-upon investment strategy.
Inheritance Planning: advising the client on the best way to define an inheritance path in order to optimize the successor’s benefits
Retirement Planning: financial planning with a defined target in terms of years and asset value
Objective-based Advice/ Financial Planning: investment planning based on quantifiable objectives; for example, save x amount of money in y years or to increase one’s wealth by a percentage
Personal Financial/Cash Management: a process more focused on day-to-day financial management and is not strictly linked to wealth management
Different types of machine learning models
Machine learning models can be used to:
• Predict future discrete values based on historical data (e.g. housing prices) or time series data (e.g. sales forecasts or number of clicks on specific product campaigns)
• Predict output classes (referred to as classification). A typical example of classification is sentiment analysis in which the models try to understand the current sentiment on specific topics (e.g. stock movements)
• Find patterns or clusters in a data set (e.g. classifying customers into segments using historical transaction data)
• Recommend products or items (also called recommender systems) based on user-user or item-item similarities. These systems are used by Amazon or Netflix to recommend the next item to purchase or movie to watch
• Recognize and interpret language using Natural Language Processing (e.g. chatbots), identify and recognize images via Deep Learning models, and translate languages
What ML technique can be applied to which Wealth Management process
Recommender systems applied to investment proposals and (self-service) sales
One of the obvious applications of machine learning techniques to the advisory process is the use of recommender systems to propose well-suited products or approaches to a client. For example, with data regarding the financial situation of a client, their current assets, their risk appetite and their goals, one could mine the data of similar clients and ascertain investments proposals that have been previously accepted by similar clients. Until this point, I have yet to see this being widely implemented by robo advisors; most likely because current robo advisors are mainly based on rebalancing algorithms and their offer is still limited to ETFs.
In complement to similarity analysis between users or products, sentiment analysis can also be used to select investments that are currently being favored on social media or recommended by investment analysts.
Moreover, financial news can be interpreted using natural language processing to detect signals that can trigger investment alerts.
Ultimately, non-financial data coming from social networks, geo-localization or e-commerce interactions would help personalize advice, but poses data privacy concerns. Personally, I would appreciate more personalized advice based on my lifestyle and spending habits, so long as I can control what data is shared with whom.
Predictive modeling applied to investment & trading strategy
Stock prices cannot be predicted, but patterns found in stock variations can be used to define or refine investment strategies; an application of machine learning that is attracting a lot of focus right now. Typical quantitative algorithms are currently being complemented by machine learning techniques that process a lot of information about companies and integrate that information into an overall algorithm. For example, we may be able to predict market reaction after a specific event (stock price drop, negative news, etc.) based on our data regarding past, similar events. Integration of sentiment analysis and news processing can also add additional refinement to the algorithms; integrating many more impacting factors and, perhaps, improving the accuracy of the predictions. Given ML’s capacity to digest an immense amount of data, it is possible ML processes may infer better predictions than humans who are using historical data to take decisions. However, if the actual future data does not hold with the predicted data, ML algorithms may not adjust in predictable ways.
Predictive modeling applied to objective based planning
As mentioned earlier, we cannot predict future stock or market performance; but using historical data, we can assess if an investment proposal was successful or not. Customers want feedback to understand what has failed and what was successful in order to augment their investment decisions going forward. Recommender systems can also be used to find similar products/strategies that could fit certain requirements. Moreover, reinforcement learning does not need to have historical data to learn, but improves its predictions over time.
Clustering applied to advice process
Applying clustering techniques to client and transaction data allows clients to be grouped together and, based on this classification, have best-fit investment products proposed to them. Also, clustering results could be made available to clients to understand what similar groups have been investing on in the past. Advanced clustering methods can be used to find unidentified patterns and go beyond the typical classifications (like geographical, AUMs and so on) and provide clients with more personalized advice than just the standard models based on static information.
Chatbots are using NLP techniques to interact with clients. As of now, chatbots are still quite primitive and cannot answer sophisticated questions; specifically in fields like wealth management. However, it is likely that chatbot creators will be able to tap into the above mentioned models and create bots that will be able to interact with clients on a more sophisticated basis.
There is currently a lot of discussion about robo advisors, but they remain limited in nature and are focusing on rebalancing portfolios using ETFs according to client risk profile. In the future, it is possible that robo advisors will integrate historical client data and contextual financial information to propose advanced investment proposals to customers not only based on best products, but also on client affinities to provide a truly personalized investment process.
Data as the key driver
The biggest challenge to implement ML/AI is the availability and quality of data; not only at the start of the process, but on an ongoing basis. The models must be continuously refined using accurate data feeds, which requires substantial effort to put in place. At the same time, data privacy regulations pose a substantial challenge as the financial advice process would need a complete overview of the client situation to be fully personalized and optimized. Therefore, a hybrid model where human interaction is still required is the best current option. Less sophisticated advice can be automated, but with limited scope.
– Text by Patrick Rotzetter
Patrick Rotzetter is an IT director who has acquired more than 20 years of experience of leading and delivering complex change projects in digital banking, asset and wealth management in Europe, US and Asia. He translates strategy into actions and ensures delivery using strategic and creative thinking.