Executive summary

Context

Superlinear developed a tool that InvestSuite can implement into its StoryTeller solution to incorporate news stories into the analysis of the performance of investment portfolios.

Goal of the project

To enable InvestSuite’s StoryTeller solution to rank news articles about the performance of specific financial assets.

Solution

A flexible, AI-driven solution that can find the best articles on a given topic during a specific period of time, thus refining even further the reporting and analysis of portfolio performance.

Case study

The challenge

InvestSuite is an international B2B WealthTech company operating across Europe. InvestSuite helps financial institutions modernize and extend their wealth management product range with a suite of white-label solutions. Its team of experts operate across machine learning, design, human insights and wealth management.

The company seeks to maximize the enormous growth opportunities created by a combination of changing customer expectations, technological evolutions and the emerging new ecosystem of financial institutions.

InvestSuite’s range of solutions are aimed at many financial institutions such as online brokers, retail banks, private banks, asset managers, pension funds or insurance companies.

StoryTeller, one of InvestSuite’s products, was developed in 2020 and is a worldwide first new way of telling the story of the performance of retail clients portfolios. Using five unique engines, StoryTeller helps to explain portfolio performance in an understandable and transparent way to retail investors. It provides an in-depth view of the returns over a chosen time period and illustrates events that have impacted the performance and to what degree.

InvestSuite wanted to further refine the reports by making them relevant in terms of real-time actuality by using current news stories to give the analysis more clarity. This will result in increased customer satisfaction, financial knowledge and engagement.

The briefing

InvestSuite asked Superlinear to develop a news article ranking engine that could be embedded in the StoryTeller solution and would rank news articles about the performance of specific stocks. The ranking tool, based on a number of criteria including relevance, content and time, needed to be extremely accurate. The top ranked article was required to give the most information and therefore the most value in terms of helping to analyze portfolio performance.

“Many retail investors receive almost no information about the reasons for the performance of their investment portfolio. When they do, it is usually either unintuitive (tables with numbers) or just a high level macro-economic explanation with an unclear link to the actual investments. We wanted to explain portfolio performance in an understandable way to retail investors through in-depth reporting. Superlinear’s commitment to excellence and our shared values made it a straightforward decision to partner for this project.”

Mathieu De Baets, Product Manager, InvestSuite

The solution

The project started with a number of known challenges. Superlinear had to find a way to rank articles in a way that mimics how humans would be reading them, which is not only hard to define, but also hard to translate into criteria for the solution.

Superlinear developed a first sprint of the solution based on data labelling. It allowed Superlinear and InvestSuite to quickly test the article rankings and to interact with the different labels.

To ensure maximum relevance of selected articles, Superlinear used a technique called ‘keywords clustering’, as opposed to an ‘embeddings’ approach. Embeddings use semantics to link keywords together by relevance.

With this technique, an article containing a lot of keywords has a high chance of getting picked up, but might not be entirely relevant. Having a related article was not good enough, as the solution needed to select only the most relevant articles for each reporting.

This is why Superlinear used the clustering method. This technique maps a number of topics through keyword clusters. Articles are then ranked based on how many specific concepts (topics) and clusters are mentioned in an article, which increases the ranking’s relevance.

During the development phase, Superlinear also suggested going from text summarization for the reports to the more user-friendly method of question answering. Instead of the AI summarizing the text of an article, it can actually answer questions in natural, flowing language. For example: ‘why did my Apple stock increase by x percent?’

If an article mentions the right keywords but isn’t closely related to the topic, the AI is asked “why did this stock go up/down?”. Based on the information, the results can be further fine-tuned.

“The collaboration with Superlinear was extremely good and we are very satisfied with the results. The Superlinear way of working is super clear and the team is extremely talented. We particularly loved it when they came up with out-of-the-box solutions to solve the problem, showing an in-depth understanding of the challenge.”

Mathieu De Baets, Product Manager, InvestSuite

“The specific challenges of this project made it clear that popular NLP solutions might not be the best fit. This allowed us to get creative. Thanks to the wealth of specialist knowledge that InvestSuite shared with us during the collaboration, in what was a very successful partnership, we designed together a solution that tackles these challenges.”

Victor Hutse, Machine Learning Engineer, Superlinear

The future

InvestSuite is currently in the process of implementing Superlinear’s news article selection engine in the StoryTeller solution and intends to improve the performance of the engine over time with trained datasets.