AI revolution: why SA asset managers should embrace early adoption to stay ahead
While the era of Artificial Intelligence may appear to be in its early stages, it is imperative for asset managers to adopt this technology and capitalise on its benefits as soon as they can.
By leveraging AI and Data Science for process automation, analysts and portfolio managers can significantly reduce time spent on routine tasks, enabling them to focus on high-value, alpha-generating research opportunities.
Secondly, gaining a deep understanding of the economics of Artificial Intelligence – and identifying the right investment opportunities – means one must get their hands dirty. At Matrix, we have embraced this approach.
By focusing on Large Language Models, we are proud to introduce our first Generative AI Application: The Ninjaturtle!
Figure 1: How the app appears on Microsoft Teams at Matrix
The Ninjaturtle is an application that Matrix staff can install in Microsoft Teams, and we can message it as we would any other colleague in the organisation.
Figure 2: The Bot uses a model trained in Natural Language Processing (NLP)
The project specification was to create a “Smart Bot/Library” that would collate all the research that the Investment Team receives daily. The Team would then message the bot with questions on specific topics across multiple sources.
The project leverages Google Cloud Services, Microsoft Azure, ngrok and OpenAI. After launching in September 2024, we initially used GPT-4o-2024-07-18, then transitioned to GPT-4o-mini, before recently switching to o3-mini-2025-01-31.
Our code processes the material by “reading” it, converting it into numerical vectors (vectorisation), stores it, and then waits for the user to “ask” a question. When a question is asked, the system retrieves relevant information and generates an answer. The AI community refers to this process as Retrieval Augmented Generation (RAG), which is similar to taking an open book exam—where the system retrieves information from a dataset before generating a response.
RAG is a great solution to Matrix’s use case because it enables the Large Language Model to be more accurate, in that we feed it data so it doesn’t have to invent things. It also has access to more recent data, as opposed to relying on its pre-trained data. RAG also ensures that the model has context – it tells us where the information is sourced which gives the user more comfort in its validity.
To learn more about this AI application, you can read NVIDIA’s blog piece titled “Generative AI for Quant Finance,” published in February 2024.
Figure 3: Detailed prompting leads to better results
Figure 3 illustrates an interaction referencing the P0160 – Residential Property Price Index Report, September 2024, which was released by Stats SA on February 13, 2025. This example demonstrates the bot responding in the requested output format while providing the required information.
While there might be teething issues with “The Bot”, as it’s referred to internally, it is an important step in AI adoption and progress. At this stage, the most pressing issue is its struggle with understanding time. For instance, it confuses projections for future events—confidently quoting index performance for March 2025 while we were still in February 2025. The AI industry terms this phenomenon “hallucination”.
The benefits of integrating with Microsoft Teams (through Microsoft Azure) is that safety and security is handled by our existing security processes together with Microsoft’s team of developers. We also gain invaluable access to Microsoft’s latest features such as the hyperlinks feature rolled out in December 2024.
Enhancing The Ninjaturtle: next steps for Matrix’s AI efforts
An important milestone would be enabling The Ninjaturtle to return the source of the context in PDF format. When we send out summaries on specific topics to the team, users often request the full context. This feature would provide a helpful self-service solution.
For the technocrats:
This project is three parts Data Science and one part Artificial Intelligence. The Data Science component requires significantly more effort, while the AI aspect is essentially a “plug-and-play” exercise once the data pipeline is established.
Currently, similar tools have only been rolled out by Bank Credit Analyst (BCA) with Intelligent AI Search and Bloomberg, both of which are still in their beta testing phase. Over time, we expect more research houses to adopt this technology.
By developing an in-house version that integrates multiple sources of truth, Matrix positions itself at the forefront of innovation. Early adopters of technology gain a competitive edge by accelerating their learning and unlocking benefits ahead of the market.
Until we become machines, we must learn to collaborate with them!
